On the evening of 9 October, I was at Lighthaven in Berkeley, sitting in the green room of Progress Conference 2026. In a few minutes, I would walk on stage with Dmitri Dolgov, co-CEO of Waymo, to interview him in a live plenary session. The video of that interview will be available later this fall, and I’m eager to share it with the world: I asked him a lot of questions about Waymo’s strategy and plans for the future, and he had a lot of interesting things to share.

Dolgov impressed me: I shared only one of my questions with him in advance, so all his other answers were extemporaneous. That one was my very first question (given to him to help him warm up): when will Waymo be coming to my city? I meant it figuratively: how does Waymo decide what cities to pursue?

One of Dolgov’s team present with us did take it literally, and told me that Waymo would be coming to Toronto “soon”. Dolgov smiled enigmatically and said nothing. And our conversation proceeded onward.

I can’t tell you what else we talked about, because Progress Conference 2026 runs under the Chatham House Rule. But I feel comfortable telling you about this exchange, because the news it anticipated is public: one week after we spoke, on Friday 16 October, Waymo announced that it is coming to Canada, with Toronto its first service area. Waymo cars (specifically the first-gen Jaguar I-PACEs, rather than the next-gen Ojais) will begin to travel the city’s streets “in the coming days”. They won’t be in revenue service until 2028, and indeed for much of the intervening time won’t be operating empty. To ensure a successful launch, human specialists will pilot the vehicles from behind the Jaguar’s steering wheels, mapping roads and collecting sensor data.

This is the moment I have been telling Toronto to prepare for. In June, I wrote in The Hub that “On 26 October, Toronto will pick its next mayor. It’s not obvious, but robotaxis are on the ballot too.” Earlier this month, in a piece for the Dais, I urged the candidates to “say what they want from robotaxis, and say it now”, because it’s better to figure these things out before the moment arrives.

Subscribe now

Well, the moment has arrived. And here’s how I think the next mayor of Toronto, whoever it may be, should think about the matter. Toronto should welcome Waymo and use its powers to set the service up for success. That’s because Toronto has everything to gain from robotaxi service, and little to lose.

I’ve watched Toronto turn away an Alphabet company once already. I don’t want to see it happen again.

What Toronto Stands to Gain

I have made the case for robotaxis many times before, but often in an American context. Today, let’s make it for Canada.

The first argument for welcoming Waymo is safety.

Chart courtesy of ChatGPT. Sources: Transport Canada: Canadian Motor Vehicle Traffic Collision Statistics, 2024 and Statistics Canada: Number of homicides, 2015–2024.

Canada suffered 2,050 road deaths in 2024, the most in a decade (admittedly we enjoyed a slight respite during the pandemic, but that’s long over). That’s almost three people killed by accidental violence for every one person who dies from it deliberately. And lest we forget, road incidents also cause thousands more serious injuries where no one dies, but nonetheless permanently alter lives for the worse.

As I have written before, driving automation is the most promising—indeed, the only promising—tool on offer to reduce those numbers, which otherwise remain stubbornly high.

Consider what one version of that tool looks like in practice. Waymo has operated its vehicles over hundreds of millions of driverless kilometres, and its safety dashboard shows that the ‘Waymo Driver’ (the firm’s automated driving system, or ADS) has been at time of writing involved in 95% fewer serious-or-worse crashes and 82% fewer injury-causing crashes than human drivers over the same distances. The Insurance Institute for Highway Safety’s independent analysis, published in July, found 68% fewer police-reportable crashes than human drivers across four cities. And there has never been a documented instance of the Waymo Driver being at fault in an incident that took a human life.

Advocates in other cities have attempted to spread fear, uncertainty, and doubt on this point. These attempts, as I have shown, don’t stand up to scrutiny, and for good reason: the Waymo Driver never drinks, thinks about other things, or looks at its phone. Compared to the human reference case, which I’m sad to say behaves this way all the time, it’s reasonable to expect a well-trained, well-tested ADS to outperform.

It’s fair to say that Waymo on its own won’t end road deaths; most fatal road collisions happen on rural roads and highways, where speeds are higher, and Waymo’s robotaxis won’t get to handling trips like that for a long time. But they will get there eventually. As Dolgov told me on stage at Progress Conference (reminder, video coming soon!), Waymo ultimately aspires to have the Waymo Driver handle the driving task anywhere driving happens. The path to ending all road deaths travels through robotaxi services being widely available, meaning if Canada wants it, Toronto is the place to start.

I could go on. Driving automation also frees us from the burden of the driving task, and from the idiosyncrasies of our ridehail drivers. Its trips will be cheaper and more readily available as well as safer (there will be more Waymo cars on the streets of Toronto at four in the morning than human ridehail drivers). But let’s not belabour these points, which I have made at length elsewhere. Driving automation generally, and Waymo as its most advanced instantiation today, is something that would bring immense amounts of good to Toronto, and in time to Canada as a whole.

All the Problems We Hear About Have Answers

Scanning early media coverage reveals a consistent set of concerns. Of course they aren’t new ones, and answers are ready to hand.

Winter. Prof. Sidney Givigi of Queen’s University told CBC that “All the sensors can get occluded by snow, which is something that they don’t have in California”. It’s true that winter is a tougher environment for an ADS than the American Sun Belt. Snow can blind a robotaxi’s sensors, make the landscape confusing and unrecognizable (no visible lane markings, for example), and occupy a much smaller portion of the training data on which the ADS was built. For much more on these challenges, see my interview from last year with Steven Waslander of the University of Toronto.

You’ll also see there that, even then, solutions were in view. More and better sensors, more and better compute to clean degraded data and fuse it into a consistent world model, and more and better training can overcome these problems… as can the more prosaic solution of adding ‘windshield wipers’ to the sensors. Waymo began revenue service in Denver last month (i.e., 1 September 2026), has tested in New York City, and is actively preparing to launch in Munich; snowy winter markets all. The company clearly does not think that snow and ice will stop them, and there is no reason to think that a task that thousands of careful human drivers in Toronto can handle will frustrate the machine. It’s notable that the long testing period announced for Toronto, from October 2026 into 2028, exceeds any period the firm has announced so far, except perhaps for Tokyo; it’s reasonable to guess the firm wants to get two winter seasons of testing done before opening revenue service. This should inspire confidence that when it’s time for paying passengers to ride, they’ll be safe, even in a Toronto snowstorm.

Jobs. This matter is one that we’ve been thinking about most recently here in Changing Lanes. In the issue published just two weeks before this one, Who Drives Ridehail?, I used Toronto’s own data to show that there are two kinds of ridehail driver: a large population of casuals who drive a little, turning over constantly, and a small core of hard workers who provide most of the trips, mostly quick turnarounds in the densest parts of the city. Waymo, when it comes, will outcompete those core drivers first, because robotaxis are optimized to handle that market.

Ontario’s Ministry of Transportation has told the Canadian Press that the rollout will require a “credible jobs plan”. This is an odd demand; I don’t remember the province asking Apple to produce a credible jobs plan when the iPhone’s digital camera was poised to wipe out photographic film development. But the fact of the demand shows that whatever good policy may be, good politics requires an answer to the question ‘what about the ridehail drivers?’ Waymo’s promise of “hundreds of roles” in maintenance and support is welcome, but I doubt it will be sufficient.

That’s why, just last week in What We Owe Ridehail Drivers, I proposed a deal: wage insurance for drivers who are doing the work when the robotaxis arrive, funded by a small levy on each robotaxi trip. There is Canadian precedent: Quebec did just this to compensate taxi drivers outcompeted by ridehail. Because such a levy would scale up in tandem with robotaxi trips, the levy’s revenue would grow in step with the displacement it aims to offset. Toronto is well placed to make it work, because it already licenses every ridehail vehicle and publishes their activity, so it knows who the core drivers are.

A Waymo Jaguar I-PACE at San Francisco International Airport, October 2026. Photo by the author.

Sovereignty. Technologists for Democracy, a group that has campaigned against Waymo since the spring, warns of the “mass collection of Ontarians’ data by an American company”.

I have seen this movie before. From my seat as Toronto mobility lead for Sidewalk Labs, another Alphabet subsidiary, I remember the disingenuous accusations that big American tech firms could have only one possible interest in Toronto, which was to vacuum up our precious bodily fluids—sorry, our precious data—for nefarious purposes.

On the one hand, this is foolish. The data that a Waymo vehicle might collect at, say, King and Church Street is already readily available; yet another Alphabet subsidiary, Google, makes it freely available to all. On the other, it is ignorant: the City of Toronto itself makes it standard practice to collect data far more intrusively than this, a practice to which I can never recall anyone objecting.

I find it hard to take this concern as anything other than throwing mud at a Waymo car to see what will stick. If one is actually serious about the rights and wrongs of public data collection, I would refer to my serious rejoinder from last March: Let the Spy Cars In. Constant readers will recall that there, I agreed with Premier Ford’s concern that Chinese electric vehicles can spy on their owners, but argued nonetheless Ontario should permit them to be sold here. I stand by that argument, which applies equally to this situation. It’s already the case that any connected car on Ontario’s roads collects data on the car’s movements and surroundings and sells it to commercial brokers, from whom anyone can buy it for whatever purpose. So data collection is no reason to keep out Waymo; if you think this is a fox, it is already in the henhouse.

Relatedly, it is true that a robotaxi fleet is a network of rolling cameras; it has to be, so as to operate safely. To those who think this is a problem, I say that you can’t threaten me with a good time: I am on record as thinking we need more public surveillance rather than less. But, fair enough, most people of good will disagree with me on this point. For them, as I’ve argued before, the solution they should seek is restrictions written into law. Ontario’s rules for commercial robotaxi service should say who may obtain fleet footage, for what purposes, and on what legal authority (probably a judge’s warrant), with independent audits. While we’re at it, we should also ban selling vehicle data to private operators without advance consent, and require warrants for police access to any vehicle data, whether robotaxis or personal, private vehicles.

If the government wants to pursue this agenda, I’d support it, but that discussion should not be an obstacle to Waymo’s arrival on our streets.

The Outcome Everyone Should Want

I mentioned earlier that I was Sidewalk Labs’ mobility lead in Toronto. I took that job excited by the prospect of using technology to help make transport in Toronto better: safer, more affordable, more reliable and convenient. I left it despondent, having seen a small cadre of haters—haters of technology, of Americans, of big firms, of any change to the status quo from which they benefited—choke the project of the oxygen and goodwill it needed to succeed.

Thankfully, with the passage of time, I think Toronto has come to see that giving in to the comfortable and the malcontent on that matter impoverished the city, and squandered an opportunity that will never come again. That was certainly the thesis of the hit play The Master Plan, and its reviewers seem to have agreed with the argument.

With Waymo, we have another chance to do the right thing. Here is what I think Waymo’s success in Toronto will look like.

Share

In 2028, Torontonians, and perhaps Mississaugans, can hail a driverless ride across large parts of the city, irrespective of whether it’s summer or winter. (Mississauga? Perhaps! See the postscript to this issue.) The core ridehail drivers who were doing the work when the robotaxis arrived benefit from wage insurance, so they can more easily find other work; that insurance is paid for by the trips that displaced them. Ontario has fair and transparent rules for commercial driverless ridehail service, subject to the two conditions I argued for at the Dais: disclosure, meaning public data on collisions and on kilometres driven empty, and compensation for displaced drivers. Please see that piece for more detail; I also recommend reviewing the thoughtful agenda for the municipal and provincial governments in this regard, laid out by friend of Changing Lanes in his newsletter, Rethink Toronto.

Victory is by no means guaranteed; the same provincialism that killed off Sidewalk Labs may yet drive Waymo away. I hope not, because Waymo’s success will be success for the whole city. I hope its residents, and its politicians, understand this.

Postscript: Waymo in Mississauga?

When Waymo published its Canada page on 16 October 2026, the text said testing would take place “across select neighbourhoods in Toronto, including Yorkville, St. Lawrence, Harbourfront, Leaside, Bedford Park, Roncesvalles Village, Parkdale, Don Mills, Rosedale, Summerhill, Yonge & Eglinton, Mimico, The Kingsway, Islington, and Humber Bay, as well as Malton, Lakeview, and Applewood in the City of Mississauga” [emphasis added].

Above, screencap of the Canada page as of 16 October; below, as of 18 October. Emphases added.

By the morning of 18 October, the page had changed. The list now reads “including but not limited to the St. Lawrence Market, Harbourfront, Roncesvalles Village, Parkdale, Liberty Village, Don Mills, Yorkdale, Leaside, East York, Summerhill, Yonge & Eglinton, Weston, Mimico, and Etobicoke”. Mississauga is gone entirely (so are Yorkville, Rosedale, and Bedford Park). The Kingsway, Islington, and Humber Bay have been folded into Etobicoke, which contains all three.

I haven’t seen any comment on this, so I can’t give reasons for it, just guesses.

One is housekeeping. The paragraph is now strictly about Toronto, so Mississauga may still be in scope, just folded into “including but not limited to”.

The other is politics. Like Toronto, Mississauga elects a new mayor later this month, and perhaps the firm didn’t want to become an issue in a market where it hadn’t done outreach to all the candidates.

As a resident of Mississauga, I am a YIMBY on this issue. Bring the Waymos on!

In August, Anthropic published a review of every randomized trial of American job training its authors could find: 56 of them, going back to 1973. The review found that the typical program raised employment, but only trivially: out of every 100 people who got retrained, only two or three more ended up employed than would have otherwise been the case. Comparing those benefits against the cost of the retraining programs, the authors conclude the result is “a little better than break-even”, but that such programs fall far short of what will be required if AI seriously disrupts job markets.

Such disruption is still theoretical for the sort of AI that Anthropic sells. But for a particular kind of ‘applied AI’, namely driverless vehicles, the disruption is real. The first large group of workers whose work an AI is going to take is on our roads, today, driving for ridehail firms like Uber and Lyft.

As discussed last issue, there are two kinds of ridehail driver. One, making up the larger group, works casually and drives only a little for a few months, rarely lasting more than a year. The other, making up a smaller core, does most of the work, giving 100 or more trips a month, and remains active year over year. In Chicago in July 2026, core drivers were 45% of active drivers, but gave 85% of the trips.

It would be nice if, as the robotaxis enter a market, they displace the casual drivers. If that were the case, casual drivers would churn out, as they would have anyway, but some would not be replaced, as the robotaxis silently absorbed that side of the market. Meanwhile, the core drivers, professional and hard-working, would keep the share of the business they’d always had.

In that world, there would be no trade-off: core drivers would retain what they have, and robotaxis would slowly fill the gap left by attrition among casuals. The public, enjoying more and safer rides, would win, but human drivers would be just as well off as they were before.

Unfortunately, that isn’t likely to be the case. As robotaxis enter the market, it’s the core drivers who will be displaced, while casual drivers keep their share of the business. The folks who are working hard and making a full-time living will see that living go away first.

Subscribe now

You may not agree that those drivers have a moral right to compensation for the loss. Seeing what ridehail drivers as a class did to taxi drivers, I understand the view. But I hope you’ll agree that they have a claim on people’s sympathy, and that that claim could endanger the rollout of robotaxis at scale, which would be strictly worse… not just in terms of overall utility, but in terms of lives lost to preventable accidents. To get the politics right, if for no other reason, we need to make smart deals whereby the winners compensate the losers; and correctly identify those losers so we aren’t rewarding the wrong people.

That is worth doing on its own, but also as practice. If AI does indeed have the impact on the job market that some anticipate, proactively addressing AV’s impact on the market will make the coming, harder transition easier to manage.

The Core Goes First

Today, robotaxis’ share of the for-hire market is non-existent in most places, and trivial even in its strongest markets. Inside its service areas in San Francisco and Phoenix, Waymo takes about 15% of what riders spend on ridehail. Those estimates count only trips that Waymo could serve, i.e., trips that start and end inside Waymo’s permitted zones, meaning that its share of the whole city is even lower than that.

But of course this will change in time. How will it change? Absent data, we’ll have to run a thought experiment. Let’s imagine five phases. In Phase One, a robotaxi operator has just begun operating in a city; in Phase Five, it has absorbed the entire for-hire ride market.1 In the intervening phases, robotaxis give 20%, 40% and 60% of trips respectively. To keep things simple, let’s stipulate that regulators don’t cap the size of fleets. That may be too unrealistic… but we could also assume that regulators cap fleet size, but the process for raising those caps is clear and reliable, such that operators can grow their fleets at the speed they wish. For our purposes, either assumption works.

Finally, stipulate that while some riders have idiosyncratic tastes, the bulk of travellers care most about price and, subject to a minimum level of quality, will always choose the lowest fare. This is precisely what we see in commercial air travel, where the finding is so durable that I have dubbed it the Iron Law.

Given these terms, we should expect that robotaxis will outcompete human-driven ridehail on price. We don’t see that yet, but it’s early days, and in San Francisco the cost of a Waymo fare relative to Uber or Lyft keeps dropping. Given that the driver’s wages represent about half of the cost of a ridehail trip, the robotaxis will ultimately cross the line and become the cheaper option, not to mention safer.

On these assumptions, as mature robotaxis enter a market and their driving hours rise, human driving hours fall in rough equilibrium.

But which humans?

Read more

But before that, two administrative notes.

Firstly, I’m looking forward to the Progress Conference 2026, happening later this week (8–11 October 2026) at Lighthaven in Berkeley, California. On Friday evening I’ll be interviewing Dmitri Dolgov, co-CEO of Waymo, live on the main stage, followed by a book event, where I’ll be signing complimentary copies of my co-written book, The End of Driving. If you’re at the conference, I’ll have limited availability for meetings, though programming will take up most of my time. And if you’re not at the conference, fear not, the video of the interview should be available later this year.

Secondly, the Dais, Toronto Metropolitan University’s policy think tank and “Canada’s platform for bold ideas and better leadership”, has recently published a short piece from me on what it will take for robotaxis to arrive in Canada, and what citizens should expect of their governments in that regard. While it uses the upcoming mayoral election in Toronto as its frame, the essay will be useful to anyone thinking about robotaxi regulation in their own city. I hope you find it valuable.

Waymo approached Toronto City Hall this year to express interest in operating in the city. The mayor’s answer, delivered through her press secretary, was that she “will not support Waymo if it costs jobs, drives down wages for other workers, or contributes to precarious work in our city”.

Put another way, the mayor was saying that she won’t support Waymo if it means ridehail drivers do less work: if they give fewer trips, get paid less for the trips they do offer, or can’t count on ridehail work to make a living.

The problem here is that there isn’t one population of ridehail drivers, but two. One is large, drives a little, gives longer trips, and turns over quickly. The other is smaller, does most of the driving, gives short trips inside the city, and turns over more slowly. As robotaxis arrive, they’ll affect each population in different ways and at different times, and the group that relies on the work most is likely to feel it first.

And what that means is that good robotaxi policy needs to think clearly about each group, what distinguishes them, and only then about what they might be owed.

Subscribe now

Importantly, I came to this data with a hypothesis I’ve held for some time, and advanced as an argument in some high-profile places: that turnover among ridehail drivers is so high that robotaxis could arrive without forcing many of them out.

Turns out, I was wrong.

The Casuals and the Core

Before we get into the specifics, let’s take a moment to discuss how we know there are two kinds of ridehail driver. The answer is that some cities are good about publishing data on the subject, which are readily available. Two of the richest are Chicago’s and Toronto’s, which record, respectively, every eligible driver’s trips each month and every ridehail car’s activity, hour by hour.

They aren’t directly comparable datasets, since Chicago’s tracks drivers while Toronto’s tracks vehicles, but they’re close enough to be useful. Researchers can supplement these with studies built on the records of New York’s taxi regulator, Minnesota’s labour department, the City of Seattle, and other places. For this analysis, I pulled the most recent full data for July 2026 in Chicago and for 2025 in Toronto. Of course, when I said I pulled them, I mean that my research assistant Claude did; it’s wonderful to live in a time when a project like this, which once would have required a team of academics, can be done by a solo writer and an AI assistant. I’ve listed the data and other sources I consulted at the end for anyone who wants to do their own analysis.

Of all these datasets, the best one to start with is Chicago’s. Helpfully, for any given month, Chicago lists every driver eligible to give trips at that time, and how many trips each one served. In July 2026, among drivers who gave any trips at all, the median driver provided 83. The median trip, though, came from a driver who provided 239.

That’s quite a spread, and it tells us quite a bit.

The first thing it tells us is that most drivers work only a little, and that a minority of drivers are working a lot, so much so that they supply most of the trips. That’s because the spread is so wide. If it had been close together, that would suggest a homogeneous group, where all drivers are doing about the same amount of work. The wider the spread, the more heterogeneous the workforce must be, with a few drivers doing most of the work and many doing little. A gap of nearly three to one suggests the latter.

And that’s just counting drivers who gave at least one trip in a month. A third of the drivers on Chicago’s list gave no trips at all. It’s not clear how many of these are people who have left the ridehail platforms and not been purged from the rolls yet, but it doesn’t really matter: whether they have left the sector, or are still nominally in it but without participating in it, they aren’t doing any work. Among those who did drive, the 45% who gave 100 or more trips supplied 85% of the rides.1

And the skew may be higher than that, because there are two classes of error in the dataset that each pull in the same direction. Firstly, the dataset blanks trip counts above 999, so the most Stakhanovite drivers, 55 of them in July, drop out of the tally altogether. Secondly, the City tries to merge the records of a driver working for more than one company at once, but can’t always do so. So in at least some cases, a driver working two apps appears as two light drivers rather than one heavy one. Both errors mean the busiest drivers are, if anything, busier and more numerous than the numbers show.

The pattern recurs beyond Chicago. Everywhere else one looks, in Toronto, New York, Minnesota, and Seattle, most of the driving is being done by a minority of drivers.

Any line we draw will be arbitrary, so let’s choose a nice bright one. How about 200 trips a month? That level of provision is roughly equivalent to working full time: the median Toronto car giving that many is logged in to the app for 35 to 40 hours a week, waiting included. Hold on, though; driving ridehail is famously a ‘gig economy’ job. So 200 is too high. Splitting it in half is 100 trips a month, consistent with a solid part-time job. Let’s draw our bright line there. Those who do less than that we will call casual drivers and those who do that or more core drivers: people for whom driving is likely a principal source of income, if not their only one.

Let’s take what we’ve learned from Chicago and add data from Toronto. Toronto’s open data follows each car from month to month. That’s not quite the same as following a driver: Toronto doesn’t record driver names, and a licensed car could be worked by more than one. But we do know that, unlike the old taxi medallions, licences are not difficult to get, so the incentive to split cars is weak. Indeed, Young, Farber and Rahman found that only 11% of Toronto ridehail vehicles were associated with more than one driver (though 24% of drivers use more than one vehicle).

So let’s assume that cars track drivers closely enough to do some analysis, and consider what was happening in Toronto between February and December of 2025.

Source: City of Toronto open data

Casuals Peel Away and the Core Remains

Most people who start driving ridehail don’t stay. Of the Chicago drivers who started in January 2025 and were giving trips by February, only a third were still giving trips six months later. After a year, their numbers had more than halved again, to 15%; after eighteen months, only 11% of them were still driving. I have argued before that the gig model runs on burning through a constant supply of new entrants; this is new evidence for that argument.

But while casual drivers giving few trips fall away, core drivers giving lots of them stick around. In Toronto, 70% of the busiest cars in February were still working in December, against 38% of the lightest. Taking the two groups as a whole, 67% of core cars were still working, against 47% of casual ones.

So the drivers who stay are doing most of the work. But which way does the arrow of causation run? Does driving a long time get you more trips, or does providing more trips encourage you to keep driving?

The evidence suggests the latter. If driving a long time got you more trips, veterans would be busier than newcomers, and they aren’t: in Chicago in July 2026, drivers who had started two or more years earlier were about a third of active drivers, and they gave about a third of the trips.2

So core drivers drive a great deal, and casual drivers hardly at all. That makes sense: surely the casual drivers log in during peak or surge periods to capture high-value work when it’s easy to get, then log off again. That’s what we would expect the gig-economy frame to encourage.

That’s a natural guess, but it appears to be wrong. Toronto’s data suggests that the weekly peaks—weekday rush hours and Friday and Saturday evenings—take up about a fifth of the week’s hours but a third of all the week’s trips. If our guess were true, casual drivers would give a much bigger share of their trips in those hours than core drivers do. But we don’t see that. Each group gives about a third of its own trips in those hours: 35% for the lightest cars, 31% for the busiest.

What distinguishes the two groups is the kind of trip they offer, not when they offer it. The busiest cars stick to the densest part of the city and give short trips: 7.8 km (just under five miles) on average, with only a quarter of their trips coming in hours when the car crossed the city line. The lightest cars average 12.7 km (just under eight miles), and about two-thirds of their trips come in hours when the car crossed the city line: runs to and from the suburbs and to Pearson Airport (also in the suburbs). Consequently, casuals are pulling in fewer fares, but each fare tends to be bigger: $22.26 on average for the lightest cars against $17.98 for the busiest (all figures for Toronto in CAD).

Trading My Time for the Pay I Get

Is driving ridehail a good way to make a living? Judging only by pay, no.

In Toronto, between January and April 2024, the median driver netted $15.35 for each hour with a passenger or on the way to one, against a minimum wage of $16.55. Crucially, though, that counts only hours with a passenger or on the way to one, and those are comparatively few. Dividing driver pay by hours spent logged into the ridehail app yields just $5.97 an hour.3

As the gap suggests, drivers spend most of their time waiting for customers. In Toronto, they’re waiting more than they used to: deadheading rose to about 48% of in-app time by 2025, from 36% in 2022.

Of course, as we’ve seen, most drivers drive very little. So that $5.97 per hour represents the median driver, but we should decompose it. Split by hours logged in a day, the lightest drivers, on for under an hour and a half, netted $2.04 per logged-in hour and $11.82 per hour with a passenger; the heaviest, on for more than six hours, made $10.37 per logged-in hour and $20.03 with a passenger, comfortably clearing the minimum wage. So to ask what a driver earns misleads: it depends very much on what kind of driver they are.

Source: Young, Farber and Rahman

Surprisingly, how far core drivers depend on driving for their living has been measured only once. That was in Seattle in 2020, where 72% of full-time drivers said driving was their only income, against 19% of those driving under 20 hours a week.4

It would also be helpful to know if core and casual drivers differ in demographics, but data on this is also scant. Nobody anywhere tabulates national origin, ethnicity, nor age by how much people drive, so if there are meaningful distinctions between these groups in this regard, we can’t yet speak to them.

Which Drivers’ Jobs?

Go back to the mayor’s condition. She won’t support Waymo if it costs jobs, drives down wages or makes work more precarious. The data tells us how to reply: Whose jobs? Whose wages? And whose work?

Illustration by ChatGPT

If we count by people, we see that most ridehail drivers are casuals, for whom driving is a supplement they take up and drop within months. If we count by the work being done, most of it is done by core drivers. These are fewer than half of those driving in a given month, but giving more than four-fifths of the trips, and likeliest to be still driving ten months later.

It would seem that the group that is adding value, and the group that deserves protection, are the core drivers; the casual drivers need the work less and drift away on their own. Which is a good finding! Let the robotaxis come, outcompete the casuals, and complement the work of the core drivers. There’s no trade-off that needs to be made!

That is an argument I have made before. In The Hub in June, I wrote that because the sector “already sees annual turnover of 50 to 100 percent”, natural attrition “will absorb most of the transition without anyone being forced out”. I made the same case about Seattle, and spoke to it on the Omnibus podcast.

Unfortunately, having looked further into the data, I now see that the case I made appears to be dead wrong. Turnover is high among casual drivers, but the core, who do most of the work, stick around. When the robotaxis come, it’s the core drivers they are likely to outcompete first.

Share

Inside Waymo’s service areas in San Francisco, Los Angeles and Phoenix, robotaxis already take about 15% of what riders spend on ridehail trips. Those service areas start inside the city, and airports and suburbs come late. Waymo, for example, began fully driverless rides in parts of San Francisco in late 2022, but reached San Francisco International Airport only in January of this year. The short trips inside the city that fill a core driver’s week will be taken by robotaxis first. The long runs out to the suburbs and the airport, which lean casual, are among the last.

What that means is that the robotaxis are coming for the people who need the work the most. How will that play out? And what do the cities where they work, or the robotaxi companies that are drinking their milkshake, owe them? The answer might be the same thing that ridehail companies owed the taxi drivers they outcompeted 15 years ago: nothing. But perhaps the circumstances are different enough that a better answer is required.

I once wrote that if the net result of robotaxis is to impoverish the Uber drivers, we’ve failed. So will that happen, and if so, what’s the alternative? That will be the subject of our next issue.

Respect to and for comments on earlier drafts.

Data and sources

• Data

  • City of Chicago, Transportation Network Providers – Drivers, July 2026, and the drivers who started in January 2025, followed to August 2026; pulled September 2026

  • City of Toronto, Private Transportation Companies – Vehicle Operating Data, all of 2025; downloaded September 2026

• Studies and reports

  • Koustas, Parrott and Reich, New York City’s Gig Driver Pay Standard (2020)

  • Parrott and Reich, Transportation Network Company Driver Earnings Analysis and Pay Standard Options (Minnesota) (2024)

  • Parrott and Reich, A Minimum Compensation Standard for Seattle TNC Drivers (2020)

  • Young, Allen and Wang, How Much Ridehail Travel Is Deadheading? (Toronto) (Findings, 2026)

  • Young, Farber and Rahman, On the Road: Analysis of Driver Earnings in Toronto’s Vehicle-for-Hire Industry (2024)

  • YipitData estimates of Waymo’s share of ridehail spending (via Business Insider, 2026)

I’m pleased to announce that I am joining the Institute for Progress as a Non-Resident Senior Fellow.

IFP is a non-partisan think tank in Washington, DC, that works to speed up scientific, technological, and industrial progress in the United States. Some of you will already know it through my work: I wrote a chapter on automating American transit agencies for IFP’s Transit Abundance Playbook. Since then I’ve been contributing to its work behind the scenes, and formal affiliation seemed to both to be the natural next step.

As a fellow, I’ll be working with IFP on how to advance better transport and mobility in the USA, particularly as applied to public transit, driving automation, and allied matters. These are problems that need close attention, and I’m glad to join the company of people who believe working on them is important. (I’m also looking forward to being the subject of one of their sweet cartoon portraits but that may take some time do produce.)

Having said that, I want to stress that for now, nothing will changes here. Changing Lanes remains independent, and the views in it remain mine, not IFP’s nor anyone else’s.

In February I published The Iron Law of Air Travel, which argued that consumers always choose the cheapest airfare available to them. In May, the cheapest airline in America went out of business.

You may be surprised to learn that I regard this as confirmation of my theory.

The Iron Law says the bulk of travellers will always choose the cheapest fare. Consequently, any airline that tries to compete on quality rather than price loses customers, and bleeds revenue until it reverts. To be more precise: consumers choose the lowest displayed price, i.e., the number on Expedia or Google Flights, whose default sort puts these first. I argued in February that a series of crises afflicting commercial air travel each forced prices down, and that the Iron Law then works as a ratchet: prices cannot rise again, so service quality cannot either. This explains why air travel is so unpleasant.

On that account Spirit, which brought the unbundled fare to the USA and spent two decades leading the lowest-fare sorting, should never have gone bankrupt. Either the Law is wrong, or my account of it was incomplete. In some fashion, my reasoning had to be off.

Subscribe now

So I made an experiment out of it. In May 2026, a week after Spirit flew its last flight, I wrote down some predictions about what would happen next. Four months later, they have either held, or are still uncertain but trending towards hold. I regard my high hit rate here as vindication of the Iron Law… at least, when phrased carefully. The Iron Law holds that the cheapest displayed fare wins, and that is true.

But it would be reasonable to take as a corollary that the cheapest airline wins also. Had I been asked in February whether Spirit would outlast its rivals, I would probably have said yes, and I would have been wrong.

The body was willing but the Spirit was weak

Before we get to the predictions and the theory of the case they support, let’s rehearse the facts of Spirit’s downfall.

On 2 May 2026 Spirit stopped flying. More than 2,000 pilots and 15,000 other employees lost their jobs, and the airline’s principal assets went to auction. These included 48 Airbus A320-family jets (about two-fifths its fleet, the rest having been leased from private owners); 22 slots at LaGuardia; and, scattered among the other airports of the USA, the boarding gates of what had once been the country’s seventh-largest airline.

For its roughly 34 years as a commercial passenger service, Spirit didn’t attempt to compete on schedule depth, network breadth, or brand value. Instead it aimed to offer the cheapest fare. Innovations we now take for granted came to the USA through Spirit, notably ‘unbundling’. Unbundling is the now-ubiquitous practice of selling a fare as low as possible, but charging separately for every amenity that was formerly bundled into the price: bags, seats, boarding order, and more. The ‘you see $289, you pay $500’ phenomenon first appeared at Spirit around 2006–07. As I argued in my original piece, this strategy was a creative response to a subsidy disappearing: business travellers, who had paid several times the economy fare for the same seats, did not come back after 9/11. Somebody had to make the cheap seats pay for themselves.

Spirit’s pioneer spirit kept the airline cheaper than its competitors for a decade, which is how long it took the legacy carriers to catch up. Delta introduced the basic economy fare class in 2012, and United and American followed in 2017. By 2017 the legacies could match Spirit’s number on any route they cared about, though it took until 2024–25 for the matching to become routine across the network. With the gap closed, the contest went to tiebreakers, where the legacies had the advantage: brand, network, and frequency. Spirit understood the Iron Law first, and that insight let it hold its own, for a time. It won so long as it had the cheapest flights. Once it drove its competitors’ fares down to its level, it had no other cards to play.

Savegary. (2023, April 5). Spirit Airlines A380-800 [Photograph]. Wikimedia Commons. https://commons.wikimedia.org/wiki/File:Spirit_Airlines_A380-800.jpg License: CC BY-SA 4.0. Remixed by ChatGPT.

You may wonder why Spirit was able to remain a going concern for almost a decade after the price gap closed in 2017. The answer is two bad things that were perversely good for Spirit.

The first was the pandemic, or more precisely, the federal payroll support that Spirit and other US carriers received through 2020–21. In the pandemic’s wake, demand outpaced the carriers’ ability to serve it, which kept even cheap fares high enough to keep Spirit in the black.

The second was an engine recall. Spirit, and other discount airlines, relied on the Airbus A320neo family of aircraft, which feature geared turbofan engines made by Pratt & Whitney. In 2023, the firm disclosed a manufacturing defect in those engines. In consequence, hundreds of aircraft were grounded for safety inspection. This meant capacity was short across the discount sector and fares stayed high; Pratt’s compensation payments to Spirit for the flights it couldn’t offer helped as well.

Spirit could only squeeze juice out of these lemons for so long. Post-pandemic inflation hit it hard on pilot contracts, maintenance, and everything else. Facing real competition on discount fares at last, it could not stay solvent. The USA’s misadventures in the Middle East caused a fuel spike which finished Spirit off, echoing the demise of Silverjet and its peers in 2008. But the fuel spike was only the inciting incident; the cause of death was loss of its pricing advantage.1

argued in May that the real lesson of Spirit’s bankruptcy was “the airline industry is highly competitive and works pretty well, and in a competitive industry companies sometimes fail”. Aviation assets are fungible, and with Spirit’s dissolution, those planes, slots and gates are now available to be used by other carriers. He didn’t write, but could have, that the pilots and other human capital Spirit held could be picked up by other carriers as well.

The summer proved Yglesias’ point: twenty of Spirit’s owned jets were sold to CSDS Asset Management in a court-approved sale, and JetBlue won Spirit’s LaGuardia slots at auction on 20 July. And Frontier and JetBlue expanded service into the vacuum Spirit left behind. As per David Casey at Aviation Week, competitors had replicated only 48% of Spirit’s capacity at the summer peak, rising to about 60% by September; the remaining 40%, the “thinner leisure and VFR markets”, were not replaced, and “remain underserved or have lost service entirely”.2

Why the hybrid carrier wins

The Iron Law says that customers always choose the lowest price, meaning that airlines can’t compete on quality. Spirit’s experience shows that, over time, the market converges on the lowest price, after which customers will look at other amenities as tiebreakers. So, over time, what wins is a hybrid carrier that unbundles everything. In that game, legacy carriers can match any discounter’s fare in its basic tier, and then seek to upsell.

Against this, as the market sorting after Spirit’s exit shows, legacy carriers still have overhead to pay, and while they can compete with discount players on a route, it’s only worth it to them if there is sufficient opportunity to upsell in it. Where there is no such opportunity, a legacy airline will instead use an available aircraft in a more lucrative corridor. That means that a discount model can still work, but only where there’s insufficient upselling to be had, meaning linking cities that are leisure-only, like Las Vegas or Orlando; or regional centres, like Provo or Latrobe; or regional-to-leisure. The USA has plenty of price-sensitive travellers, but relatively small demand for travel on routes like these, and that is what held Spirit back. (Conversely, densely-packed Europe has large demand for travel on routes like this, which is why RyanAir and others ultra-low-cost carriers, or ULCCs, thrive there.)

Having built this model in my head, I made some predictions. They are expressed in a few ways, but the common idea was this: basic economy will persist even absent Spirit to enforce market discipline on the legacies. My theory was that the stripped-down unbundled tier had become the shape of the market. If it was only ever a gambit to keep up with Spirit, then with Spirit gone the legacies would declare victory and re-bundle, and the Iron Law was wrong. I predicted they would stay unbundled, and that the hybrid model was the future.

(One caveat: I wrote these down on 10 May in a note to myself but never published them. Please take my word that this is the whole list, not just the flattering parts.)

Stripped of detail, the bet was this. Nobody would start a new discount airline. The surviving discounter would keep turning itself into a hybrid. Where Spirit had been the only carrier, service would vanish or come back thinner. And the legacies, with nobody left to undercut them, would keep the stripped-down fare anyway.

The legacies kept the unbundled fare

I can report that my predictions largely came true.

The most important one was that, as I anticipated, the legacies continue to lean hard into the hybrid model rather than returning to higher and bundled fares. On 8 July Delta went ‘full unbundled’, launching Basic First Class and Basic Business Class as fare options. Two days later, it reported quarterly results in which premium ticket revenue again exceeded main cabin. United already sold basic, standard and flexible fares in its premium cabins, and its chief commercial officer, Andrew Nocella, put it beyond doubt on the 16 July call: “That does not mean that we’re going to step away from basic economy.” American reported premium revenue at nearly half its ticket revenue, collected on roughly 30 per cent of its seats.

In other words, the big carriers spent the summer exporting the basic architecture into first and business class, if they hadn’t done so already. None of them even mentioned Spirit in a pricing context on their earnings calls. That’s what one would expect if their commitment to the fare category was not about competing with Spirit, but was instead about responding to a structural feature of the marketplace.

Meanwhile, on its 29 July call, Frontier described its strategy as “segmenting our revenue base” through first-class seating and other premium products, while genuflecting toward “preserving the cost discipline that defines Frontier’s model”. That is the hybrid model in a nutshell. Allegiant, which now styles itself the “leading leisure-focused U.S. airline”, has not expanded beyond its Las Vegas, Florida and Arizona markets. Finally, while Breeze has expanded modestly in secondary leisure markets, neither it nor Allegiant has announced any intention to compete with the legacies on main routes. And no new discount airline has emerged to do so.

How are things in the markets that only Spirit served? As per Casey above, service in those places disappeared, or came back thinner. Atlantic City, where 94% of seats were Spirit’s, now enjoys Breeze and Allegiant flights, albeit most of them only two or three times a week (Spirit was daily). Business Insider analysis found fares up about 14% on routes Spirit had already exited in 2024–25, against 6% where it kept flying.

What we’ve learned

I have to take my lumps. I had expected, though I didn’t frame it as a prediction, that Spirit’s routes into hubs would be absorbed by the legacy carriers’ basic economy service. But this did not happen; the airline that did the most to fill the vacuum was the other ULCC, Frontier. That carrier added about 1.3 million seats into Spirit’s markets, in contrast to JetBlue at half a million, and Delta at essentially none. I suppose what I had missed was that JetBlue and Delta didn’t need to add seats, because they could absorb demand with existing capacity. Delta, for its part, told investors it was not growing its main cabin, and would rather chase revenue than market share.

So here is the Iron Law, slightly rephrased to reflect what I’ve learned. Consumers always choose the lowest price, yes. But one obvious corollary is wrong: the Law doesn’t mean that the carrier with the lowest price wins. Instead, the one that can offer that price, and supplement with premium upselling, will win. The ‘pure players’ that have only ultra-low costs to offer, or conversely only luxury service, are outcompeted in each case by hybrids, which can take the top row of the search results when they think it worth it, and then upsell behind that number. The North American market, over time, converges into one of full-spectrum carriers competing tier against tier at the top of a search-results page, more entrenched, and more artfully unbundled, every quarter.

Share

This may seem gloomy: the future of commercial air travel seems to be a consumer, being nickel-and-dimed on every aspect of their trip, forever. The American government seems content with this outcome; courts struck down a Biden-era fee-disclosure rule in February, and it was formally rescinded in July. A further DOT proposal would let airlines unbundle the displayed fare even from mandatory taxes and fees, noting that “Under the current regulation, the total price must be displayed more prominently and in a larger font than any individual component of that price”, and that this rule “is unnecessarily prescriptive”.

But there is one ray of hope, shining on Europe. Courtesy of EU regulation, from late 2027 fares advertised in Europe must include a cabin bag, and must be displayed by default before booking begins. This is the remedy I have proposed before: require the advertised fare to include a bag and a seat, so the search results have a floor higher than ‘you get to sit down’.

Now this is very small beer; one reason air travel is unpleasant is that, to avoid checked-bag fees, everyone packs carry-on only if they can, which delays boarding and deplaning, and leads to Hobbesian battles for space in the overhead bins. But at least it sets a principle that advertised prices should take a more-comfortable baseline. It’s still the case that airlines may sell a no-bag fare underneath the mandated one, but that’s fine by me. My beef with air travel is not price competition, but that fare display puts everyone in a sub-optimal equilibrium. Once the default advertised fare includes a carry-on bag, we are at least taking a more comfortable flight as our baseline.

My expectation is for good news here: that European discounters will compete harder on the bundled number, rather than pushing hard on ‘down-selling’. If that comes to pass, then perhaps regulators will find the courage to go further, and extend the baseline to a checked bag. That last one seems so remote as to be not worth thinking about… but any trip, no matter how long, has to start somewhere.

Firstly: I’m writing another book.

Canada is the world’s second-largest country by land area, but more than 70% of Canadians live in just 41 cities. As one former Prime Minister quipped, “If some countries have too much history, we have too much geography”. A country like this will only be as successful as its transportation systems permit it to be.

We used to know this. The defining national project of the country’s opening decades was the Canadian Pacific Railway, linking the whole nation coast to coast. That project was completed with the driving of the last spike of the railroad at Craigellachie, British Columbia (now a national heritage site). In the twentieth century, we built the Trans-Canada Highway, the St. Lawrence Seaway, and a network of world-class airports.

And then, sometime around the centennial of Confederation, we stopped. Where once we were a country that built transportation, we became one that forbade building. Canada’s productivity, confidence, and national unity has suffered in the decades since; this is part of the reason why.

But there is one advantage to waiting so long to build: we can skip ahead. Revolutions in mobility are happening now, or about to happen, and we can move to them directly. Automated freight trucks, robotaxis, electric air networks, supersonic aircraft, carbon-free jet fuel… all are poised to change our country for the better, if we let them. As a great Canadian observed, the way to win is to skate not to where the puck is, but to where it’s going to be.

The Last Spike was the culmination of the great nation-building project of the 19th century. The next spike is ours to drive.

Subscribe now

The Next Spike: How to Get Canada Moving Again is forthcoming in 2027 from Sutherland House, as part of its Sutherland Quarterly series. Please keep following Changing Lanes for updates.

Secondly, next month I’ll appear at the Vision Zero Cities conference at NYU in New York, at the panel on Emerging Technology for Street Safety. The annual gathering unites city officials, planners, engineers, advocates, scholars, and industry leaders who are dedicated to the common goal of achieving Vision Zero, i.e., the complete elimination of road deaths and injuries. Frequent readers of Changing Lanes will know that I think driving automation is the most promising tool on offer for achieving Vision Zero… if we deploy it with appropriate policy and regulation. On the panel, I will offer my views on what we need, and how to get it.

I hope you can join us! Registration is now open.

Finally, I’m also pleased to announce that on Saturday 9 January, I’ll give the opening keynote at TransportationCamp DC, the unconference Smart Growth America holds the day before the Transportation Research Board’s annual meeting in Washington, D.C.

I’ll be speaking on the subject of driving automation as it applies to cities. Some people think that robotaxis are the solution to a city’s problems, even if they haven’t bothered to ask what those problems might be. Other people think that cars of any sort are antithetical to what makes cities great, even if they haven’t bothered to ask if the problem is the car, or the (human) driver.

I think that both sides miss something important, which is that cities have goals and mandates, and that driving automation is a wonderful tool for civic leaders to employ to fulfill their objectives; their objectives, not those of robotaxi companies nor any other special interest.

I will make that case in person. I hope to see there!

Registration is now open.

Photo credit: Alexander Ross, Hon. Donald A. Smith driving the last spike to complete the Canadian Pacific Railway, 1885, Craigellachie, British Columbia. Library and Archives Canada, accession no. 3194527. Colourized and sharpened by ChatGPT.

Hey, kids! Here is a research experiment you can do!

Open X (get your parents’ permission first), type packed or crowded into the search box, then add weekend, and the name of any large North American transit agency. Set the date filter to this past summer.

If you do, you will find entries like these:

(@shk66, 29 August)

(@NYCStreetStages, 16 August)

(@gatekeepchicago, 9 June)

(@teaseburger, 29 August)

The pattern among these is clear. It’s a Saturday and there’s a special event downtown: a sporting match, music festival, convention, or expo. The poster is complaining about a train that’s too crowded to board and much too crowded to enjoy.

Let’s set aside the fact that X is where malcontents go to complain about things. (In that spirit: you can follow me on X here.) And let’s acknowledge that the plural of anecdote is not data. Having made those stipulations, is there anything to learn here?

I think there is. Notice that nobody is complaining that the train is full. They are instead complaining that the train is full and the agency is running the weekend schedule anyway. The two Toronto grievances are about “a normal weekend schedule”, the first one about doing this before the baseball game. The New Yorker concedes that four-minute headways on the L train, a frequency unheard of in most of North America, are not enough on a Saturday. The Chicagoan is angry that the city is chock-a-block with special events, and the CTA meets that demand with maintenance service and diminished capacity.

What this amounts to is that across North America, transit agencies are running their schedules to protect the weekday commuter peak, at the expense of weekend use.

That’s noteworthy for two reasons. Firstly, it’s strange, because that commuter peak is dying while weekend use is surging. And secondly, it’s paradoxical, because the ebbing of the weekday peak and surging of weekend use is actually a good thing. Indeed, it’s the best thing to happen to transit in this generation, and operators are failing to capitalize on the opportunity.

Subscribe now

Weekends up, weekdays down

Because anecdotes aren’t data, let’s begin with evidence.

For nearly every major transit system in North America, the part of its service that has recovered fastest from the pandemic is not the weekday peak, but the off-peak: the weekend, and weekday middays and evenings. In several systems weekend ridership has blown past where it sat in 2019, while the morning and afternoon commuter peaks still lag.

Look at New York, where through 2025, weekend subway ridership reached 89% of its 2019 level while the weekday sat at 74%. New York ridership, the highest in North America, is growing again, thanks in large part to the Manhattan congestion charge… and fastest on weekends.

Look at Toronto, home to GO Transit. GO began the pandemic as a railway built around the office commute, with only 39% of its riders travelling off-peak. In May of this year Metrolinx restored 15-minute weekend service between Union and Oakville. That lasted four months: this September the weekend went back to half-hourly, because the railway could not crew it. It kept its weekday peak periods as they were.

Look at Boston, where the MBTA’s commuter rail is the only legacy network of its kind in the United States to have come close to its pre-pandemic ridership. And how did it achieve that feat? By rebuilding its timetable around hourly, all-day, two-way service and a ten-dollar weekend pass. Roughly half its riders now travel for something other than their jobs.

Finally, look at the Bay Area. Across the fiscal year that ended in June, a year in which World Cup matches and watch parties put more than 38,000 riders onto Caltrain, system ridership rose from 64% to 82% of pre-pandemic levels. That increase was helped along by weekend ridership, which increased to 150% of those levels.

Caltrain reports weekend ridership against systemwide recovery. Washington reports weekend against all off-peak. New York and Los Angeles report weekend against weekday (LA figures from January 2025). Some cities I would like to show are absent; Toronto and Boston publish no weekend/weekday split at all, while Chicago publishes no 2019 baseline. To avoid misleading through charts, I must note that the horizontal axis starts at 60%, not zero. Chart courtesy of Claude.

Given this much evidence, the conclusion is clear. North America’s transit systems are not undergoing a slow post-pandemic recovery that will, eventually, get them back to the ridership numbers and patterns they had in 2019. Instead, they have structurally changed: the passenger these systems are gaining looks more like a weekend user going for entertainment, not a weekday passenger commuting to work.

No, the peak isn’t rebuilding itself

If I were prosecuting this case in court, I’d expect defence counsel to argue that, to the contrary, the pandemic recovery is still happening. That recovery happened on weekends first, and it’s happening in the workday peak periods as well, just more slowly than anyone anticipated. Return to office mandates have been slow to come, but they’re coming, and eventually will be everywhere. And return to office means return to the weekday peaks.

The best evidence for the counterargument is that peak ridership is indeed growing in absolute terms almost everywhere. That means that, given time, ridership will rise to its old levels. Exhibit A here would be Washington, D.C. On his first day in office in January 2025, President Trump ordered federal employees back to their desks full time. As a result, weekday Metrorail ridership jumped about 20% year-over-year in the strongest months.

As the prosecuting attorney, I’d invite you to look at that evidence more carefully. Washington is an exceptional case in that its economy is dominated by one employer, whose staff dominate the morning commute. We must grant that ordering them all back did have a big effect on WMATA’s ridership.

But then consider Ottawa. It’s even more of a company town than Washington: public administration accounted for 21% of its labour force at the 2021 census, against about 13% across the Washington metropolitan area. The Canadian government ordered its staff back four days a week this July, but OC Transpo’s ridership still ran below the previous year’s in each of the first seven months of 2026, at 67% of pre-pandemic levels. Against this, OC Transpo’s weekend bus ridership was already above its 2019 level by 2022.

So with capital cities, the evidence is mixed, which isn’t good for ‘the peak is rebuilding itself’ argument. But let’s set this reference class aside, since places like that are unusual. In other cities, where hundreds or thousands of employers make up the local economy, there is no one lever to pull to bring workers back to the office. And the fact of competition means that employees can vote with their feet against return-to-office mandates. And they have: from early 2024 to mid-2025 the companies increased the time they required employees to spend in the office by 12%... but the time that employees spent in the office increased by only 3%. The Survey of Working Arrangements and Attitudes, which has measured this every month since 2020, put work from home at 26% of all paid days in May of this year, down only slightly from 28% three years earlier, notwithstanding all the intervening mandates.

So return to office won’t bring the 2019 peak back.

The peak was bad and it should feel bad

Let me be clear: that’s a good thing.

Serving the weekday peak is a pain for operators, and working around the problems it causes shapes everything an operator does. That’s because peak demand arrives all at once, and it’s going in only one direction. There is massive demand inbound in the mornings, while almost none outbound, but outbound trains must run out, so they can run inbound again. The same pattern runs in reverse in the late afternoons. To satisfy this pattern, agencies must own enough vehicles, and roster enough operators, despite the fact it needs them all only for those few hours. The equipment and people are idled between the peaks, through the middle of the day, meaning that assets are unused, earning no revenue.

It’s a heavy burden to carry. Alon Levy has run the numbers: a railway gets roughly a thousand hours of peak operation in a year against some 5,800 hours of daytime and evening running. That means that a car bought to serve the peak costs about five times as much per car-kilometre as the same car bought to serve the base. Counting crew and infrastructure alongside the vehicles, Levy puts peak service at around three times the cost of base service.1

Conversely, spreading demand throughout the weekdays, and seeing more of it on the weekends, is cheaper and easier to serve, spreading riders across the hours that the system must pay for in any case. Since transit’s fixed costs are high and its variable costs are low, most of what a trip costs is spent on the first rider, and every rider after that brings in gain. What that means is, a crowded Saturday train is good news for an operator.

That means that all those complaints on X amount to an operator seeing money lying on the platform floor, but deciding not to pick it up because bending over would be uncomfortable.

Takomabibelot, Empty Platform (New York, NY) [resized]. 16 Jan 2011. Flickr.

Part of operator discomfort in this regard is that this new shape of demand doesn’t bring in the revenue the old shape did. Most systems are not back to total ridership exceeding 2019 totals, because five-day-a-week commuters were steady users, while recreational travellers are more occasional. That means that operational arrangements built on the foundation of the old demand pattern may not be sustainable. Caltrain, for instance, projects an average annual deficit of $75 million from fiscal 2028 and can’t sustain operations on the fares that it charges. Agencies that organized their service around serving the peak, letting the peak determine their timetables, fare products, budgets, and labour arrangements, now find themselves in a different world.

My wish is that rather than simply treating this as another problem that must be solved with higher subsidy—another manifestation of the Endless Emergency—that agencies try something different: recognize that the old world isn’t coming back, and reorganize accordingly.

Not too late to build a better world

The first place to reorganize is with track work.

Historically, the reason agencies did the bulk of their track work on the weekends was that doing the work must begin and end with the lengthy process of moving equipment on and off of the right-of-way. WMATA’s own figures show that, once that process is accounted for, a weekend shutdown can offer up to 48.5 hours of actual engagement with the project out of 56, while a weeknight shift only offers two-and-a-half out of five.

That trade-off was fine when there were few riders on the weekend to discommode. But with weekends now being the growth area, the old way of doing it will have to change. More weeknight services will be required. I am confident that there are efficiencies to be found in right-of-way equipment transfers; the weekend as the stopgap has meant in practice a lack of incentive to look for improvements. The search should commence. Flexibility in service can help too: BART projected that opening service an hour later on weekday mornings would speed up its rebuilding projects by 40%.

The next place to reorganize is timetables. Washington now runs every Metrorail line at least every twelve minutes on weekends, and the Red Line every six. Fixes like this require accommodation with labour, but if that’s what serving the customer demands, then so be it. After that comes fare products. Boston sells a ten-dollar weekend pass, and the TTC recently retired its monthly pass in favour of providing free trips after forty-seven rides in a month; both aim to serve demand where it is growing. Innovations like these should spread.

Share

What should certainly stop happening is sacrificing weekend service when farebox shortfalls begin to bite. That is what GO has been doing lately: to protect its many track work projects, not to mention its labour arrangements, it has been acting to protect the weekday peak, because historically that has been its raison d’être. This attitude also lies at the heart of its biggest recent failure, its broken partnership with Deutsche Bahn, hired to help bring European-style, frequent, two-way service to GO: as one insider told The Trillium, many at GO think running frequent trains in both directions is “a kind of a dilution of their mandate, or an unnecessary frill.”

But the peak commuter isn’t coming back. In a work-from-home world, demand is not so jagged, and there’s more of it on the weekend. People want to ride trains then.

We should let them.

Respect to and for feedback on earlier drafts.

I hope you enjoy Bern’s take on car-light new builds. If this whets your appetite, please see my own: California Forever’s Mobility Vision and my Sidewalk Labs / Quayside retrospective, What Sidewalk Labs Got Right.

I recently joined some twenty other guests on a tour of YZD, the 370-acre redevelopment taking shape on Toronto’s former Downsview Airport lands, its name drawn from the airport’s aviation call sign. Downsview opened in 1929 and spent most of a century building aircraft, including, for many years, the de Havilland Beaver and the Dash 8 turboprops that connected Canada’s regions to its cities.

Northcrest Developments, which leads the site’s transformation, is a wholly owned subsidiary of PSP Investments, one of Canada’s largest pension managers. PSP is patient capital, not a build-and-flip developer, and Northcrest is explicit that this is a thirty-year, $30-billion undertaking, unfolding across seven neighbourhoods. The first of those, the Hangar District, just broke ground. This development—all of it—holds enormous urban promise far beyond Toronto.

There’s reason to pay close attention. The Greater Toronto Area is short many tens of thousands of homes, its residents are tired of congestion and the daily grind of car-dependence, and for the first time we have mobility technology—ride-hailing, microtransit, and soon enough, automated fleets—mature enough to make a car-light or car-free neighbourhood livable in practice. These forces, all pointing the same direction, are meeting on 370 acres near the centre of Greater Toronto. Opportunities like this do not arrive often, let alone with a developer who has the patience and capital to act on them.

That was my read walking through Northcrest’s Experience Centre. Every rendering, every model, showed wide pedestrian streets, courtyards, and the reimagined two-kilometre Runway as a public spine running through the site. Not one of them showed a car. I was not the only guest to notice: judging by questions and asides I overheard, a good number of us on that tour were excited by the idea of a community that would not require, or even want, household vehicles.

Subscribe now

Still, a few asked the perennial question: Where is the parking? The answer offered was that parking would be provided according to the City’s standard ratios, which I took to be the familiar 0.5 to 1.2 spaces per unit that has governed condominium construction across Toronto for decades.

That answer works in Northcrest’s favour. Toronto no longer has minimum parking requirements. City Council repealed them in December 2021, replacing minimums with maximums, to stop forcing developers to build parking nobody asked for. So whatever ratio ends up in the ground at YZD will be Northcrest’s own choice, made out of habit, lender convention, or an assumption about what buyers expect. Toronto requires none of it. That is a more interesting position to be in. It means YZD is free, right now, to plan around a lower rate of vehicle ownership than a standard Toronto condo project, without asking permission.

And the trend supports that bet. Toronto already has the lowest household car ownership rate of any major Canadian city, and recent industry data shows that rate continuing to decline while the cost of owning a vehicle in Ontario keeps climbing. Ontario drivers spend more to own and maintain a car than drivers anywhere else in the country.

Car ownership in the GTA is not about to collapse. But the trend enables development that thrives without car ownership, rather than assuming most residents will want one. Neighbourhoods like those planned for YZD, built from scratch over the next three decades with shared mobility designed in from day one, could bring household vehicle ownership within its boundaries down to a tiny fraction of today’s rate.

The car-light toolkit already exists

I have studied what it takes to build a neighbourhood like that: enough to supply a chapter of The End of Driving (Elsevier, 2025), where we called our proposal “Car Zero Communities.” No single tool gets you there. You need several things working together, starting with walkable, complete neighbourhoods where most daily needs are within fifteen minutes on foot or bike, which YZD’s Runway-centred plan is targeting. Schools are the clearest test of that promise. I asked on the tour whether YZD would have its own elementary and high schools rather than sending children elsewhere by bus or by car, and the answer was yes. The school run is one of the most stubborn drivers of household car ownership in any city, the reason a second or third car often gets justified long after it stops being needed for anything else, and a reliable source of the worst traffic of the day in almost every car-dependent suburb. A neighbourhood that puts its schools within walking distance removes that justification before it takes hold.

In addition to walkability, you need mobility services—transit, bikes, and ride-hailing—organized well enough that not owning a car feels like an upgrade. You need a fleet of shared shuttles and taxis, including automated ones as robotaxi services mature, to connect neighbourhoods to transit and employment beyond walking distance. And you need a release valve: a small shared fleet, available by reservation, for the cottage trip or the furniture store run.

One more tool works at a smaller scale than any of the above: the short car trip for a few groceries or a forgotten item, the kind that justifies household car ownership for many families. Sidewalk delivery robots have been doing real work for exactly that trip in the UK since 2018, when Starship Technologies launched its service in Milton Keynes; the company now estimates that service alone has avoided more than 280,000 car journeys. The Co-op chain has since extended the same robot service to tens of thousands of households in Leeds, and across Starship’s UK network roughly seven in ten customers skip a car trip for robot delivery. It’s a small piece of the puzzle next to transit, shared fleets, and robotaxis, but it closes a gap those bigger systems can’t reach. Larger service vehicles—for emergencies, moving, deliveries, maintenance—will always need neighbourhood access, but the last few hundred metres, the segment generating the most street clutter and vehicle dwell time, is where delivery robots do their best work. The heavier end of goods movement needs the same early attention: 55,000 residents will still need cube trucks and service fleets, and the loading space to receive them, designed in from the start rather than squeezed in later, a challenge Andrew has examined with freight specialist Sandra Rothbard. YZD’s plans do not promise the ultra-narrow street fabric that makes servicing hardest; that is a freedom its designers should spend deliberately.

None of these tools are exotic. Most exist today, in some form, in many cities. What is rare is a site with the scale and the patient ownership to fit them together from the start, rather than retrofitting them—at real cost and pain—into neighbourhoods built around the car since the Second World War.

One challenge this toolkit does not resolve: the early phases of any new greenfield development face a cold-start problem for public transportation. If mobility services arrive before residents, they are expensive to run with little sustaining ridership. If residents arrive first and transportation is inadequate, people drive. Once those habits form, they are hard to break. The answer California Forever found for its project, the Suisun Expansion, was to design the first neighbourhood so that daily needs are accessible on foot or bike, making transit temporarily superfluous, while pre-committing to full service before the second neighbourhood opens. PSP’s patient capital gives Northcrest an unusual ability to make that same pre-commitment, coupled with peripheral parking and an on-demand shuttle service that can turn autonomous in time. The question of which body holds that commitment across a thirty-year, seven-neighbourhood build-out—and how it is enforced once parcels change hands—is trickier, and one the project will need to answer upfront.

The $900-million parking question

Which brings me back to parking, because the financial case for rethinking this is not abstract. Northcrest’s website puts YZD’s residential program at 28 million square feet, housing over 55,000 people in more than 30,000 homes across seven neighbourhoods.

Apply the 0.5-to-1.2-spaces-per-unit range that has been standard practice across the city—say 0.8 spaces per unit as a working average—and 30,000 homes works out to 24,000 parking spaces. Underground parking in Toronto runs, by industry estimates, between $70,000 and $100,000 per space, accounting for all costs. Assuming a $75,000 average, that puts the all-in cost of building that parking at YZD at $1.8 billion: a cost to be compounded, taxed, and passed to its residents and tenants.

Chart data by author, chart presentation by ChatGPT

Cut that ratio in half across the project—which a well-designed, mobility-first neighbourhood should be able to do without stranding anyone—and YZD would save roughly $900 million at today’s construction costs, spread across thirty years of building. Much more than a rounding error, that money can go to transit subsidies, a shared vehicle fleet, the parks and cultural space YZD has already promised, and, not incidentally, back to the pension funds that have put their capital behind this project. Any of those outcomes is better than 12,000 more concrete spaces standing empty in the 2040s and beyond. Underground parking, unlike almost any other square footage in a building, has a diminished second life once the cars stop coming.

There is a precedent for handing this kind of saving to residents. Culdesac Tempe, in Arizona, built its first neighbourhood with no resident parking at all, and instead gives every Founding Resident roughly $3,000 a year in mobility credit—transit passes, ride-hailing, bike and scooter access—funded in part by what the development didn’t spend pouring underground concrete. YZD does not need to go that far to make the same point. A community that asks residents to agree not to store a personal vehicle on-site, in exchange for an excellent mobility package and a price tag that reflects real construction savings, while providing a parking area at the edge of the site for the family road trip or move-in day, is not a deprivation. For a meaningful slice of the market, it is the central appeal.

The Culdesac model works partly because it has a single vertically-integrated operator who can make and keep that bargain over time. Northcrest is in a different position: it is selling and leasing homes across seven neighbourhoods to different builders and residents over thirty years, which means the mobility commitment needs a governance structure that outlasts any individual transaction. A Transportation Management Association—a body that pools contributions from residents, employers, and possibly the developer, to fund and coordinate mobility services—is the common instrument for this situation, deployed in comparable projects in the Netherlands and elsewhere. It does not require provincial legislation or a new city bylaw; it can be created by covenant, written into the conditions of sale and the condominium declarations from the outset, and governed by the community itself. The Merwede district in Utrecht has adopted a version of this approach. Northcrest would be wise to name its equivalent early, before the first units close, while it controls all the land.

I’d guess there are easily 30,000 GTA households who would prefer to live and raise a family somewhere without a street full of parked cars and a driveway to maintain. Several people on my YZD tour described exactly that household.

Tempe, Utrecht, Toronto

Culdesac Tempe settled whether people would choose to live this way, with a real neighbourhood, over a thousand residents, and zero resident parking. Utrecht is now testing it at the next order of magnitude. Merwede, under construction along the Merwedekanaal, will house about 12,000 people in 6,000 homes, with car ownership permitted but rationed to one parking space for every four households, kept at the district’s edge and shared rather than parked at the door. Culdesac proved the appetite; Merwede is proving the scale.

ChatGPT’s depiction of YZD from above

YZD, at full build-out, would test whether this works at the scale of a small city: 30,000 homes and fifty-five to sixty thousand residents, built into one of the most expensive and most car-saturated metropolitan regions in North America. “Car-free” is the wrong term for what all three projects share: cars remain available, taxis circulate, shared vehicles are reserved and returned. The right term is freedom from car ownership: the elimination of the household vehicle as a requirement of daily life. That is the commitment that scales across Tempe, Utrecht, and YZD, and it is a more durable one than banning cars, because it relies on making alternatives better rather than making driving harder. Each step in that progression has had a fresh decade of mobility tools to draw on that the previous one didn’t: ride-hailing was a decade old when Culdesac broke ground; shared and automated fleets barely existed when Merwede was planned; by the time YZD’s later neighbourhoods rise in the late 2030s and 2040s, mature robotaxi service in this region is a reliable bet. Toronto didn’t engineer this alignment of site, scale, and timing, but it has been handed all three at once, and that doesn’t happen often. I’d like to see the city make the most of it. If YZD reaches its full population and a community of that size—fifty-five to sixty thousand people—anywhere else in North America still has lower household car ownership, I’ll be surprised, and disappointed.

None of this is a criticism of what Northcrest has envisioned. The fact that YZD’s public-facing vision shows no cars at all suggests the instinct is already there; what’s left is to let the engineering and the pro forma catch up to the rendering. Few developers anywhere have the combination of site scale, patient capital, and a thirty-year runway that would let them pursue such a car-light identity as a deliberate market position. Northcrest is unusually well placed to prove that a major North American city can do this at scale, rather than in a thousand-resident pilot in the Arizona desert, or a six-thousand-home district in the Netherlands.

Share

A few smaller things from the tour stuck with me, and they show how much opportunity this project has in reserve. The site sits exposed enough, on what used to be open airfield, that it catches a steady wind. The tour guide mentioned it almost in passing, but it struck me as a real and underused asset for a project thinking in decades. Even modest on-site wind generation, sited at the edges where it wouldn’t compete with building form, could offset a meaningful share of the development’s energy demand.

I’d also like to see Northcrest consider vertical farming at the margins of the site. Controlled-environment growing of leafy greens, herbs, and berries is well suited to dense, enclosed urban settings, partly because it keeps the growing environment sealed off from outdoor contaminants and weather. It could trim the food-mile footprint of 55,000 new residents while doubling as a visible, tangible piece of YZD’s sustainability story.

And then there is the small apiary already on site: every guest on my tour went home with a little jar of YZD honey. It’s a tiny gesture, but an effective one.

The argument we made in The End of Driving is that cities do not need to wait for robotaxis to start building neighbourhoods that don’t depend on the car. The tools exist today: transit, bikes, ride-hailing, delivery robots, and a willingness to design streets for people first. Robotaxis, as they arrive at real scale, will make car-light or car-zero living easier, but they will accelerate a transition that does not depend on them. YZD has the rare advantage of starting from an empty runway rather than retrofitting an existing neighbourhood, and its first phase is only now coming out of the ground. That is when decisions about parking ratios, mobility platforms, and what a household is implicitly promised when it moves in get made, and when they are cheapest to make well. Toronto won’t get many 370-acre second chances, and I hope Northcrest makes the most of this one.

is a persistent believer that we can build cities more for people and less for vehicles. He founded the non-profit Urban Robotics Foundation and drafted an ISO standard to address the robotaxi industry’s Achilles heel, by standardizing pickup and drop-off orchestration.

I enjoy Mathew Growden’s newsletter The Change Optimist, which recently offered a history of solar-powered cars. Reading his work sent me down the rabbit hole of the subject, which led to today’s piece. Respect to Growdy for the inspiration!

There’s an old saw that intellectuals like to start with a premise no one can deny, and end with a conclusion that no one can accept. Solar cars are a great example of this kind of thinking in action.

Start with the premises we can all agree on. Firstly, most cars spend most of their time out of doors, so during the day sunlight is falling on them. Secondly, solar cells are cheap now, with prices having fallen, in USD, from $128.00 a watt in 1975 to about $0.26 in 2024 (a 99.8% decrease). Thirdly, electric cars are increasingly popular; this year, over a quarter of the cars sold worldwide will be electric. BloombergNEF projects that electric cars on the road will outnumber the internal-combustion ones by 2047… and market share will keep rising from there.

Therefore, electric cars should be solar cars. Make the roof of a car a solar panel, and have it charge the car battery. Owners would need to charge much less often, and perhaps never, if they lived in sunny places. The cost of charging would drop: not only the literal cost of money spent on electricity, but also the figurative cost, the chore of plugging in. That would be a win for the owner, but also for municipalities, employers, and everyone else that would otherwise have to shoulder the burden of arranging for charging infrastructure, and the electrical infrastructure (transformers, etc.) to support it.

It’s a vision to delight any futurist, but unfortunately, it’s too beautiful for this world. Solar panels on cars are much less productive of energy than solar panels anywhere else, and that’s for reasons that material science can’t fix. That means that solar panels on cars don’t pass cost-benefit analysis… and not only is that claim true today, but it will be true in the future as well. Solar panels don’t belong on cars, they belong on roofs, or in solar parks.

Or, of course, on trucks.

The Parking Problem

One reason I’m confident in this analysis is that we have three separate instances of solar car companies arriving with a bang and departing with a whimper.

Aptera launched in December 2020 touting a technology it called Never Charge (note the name). Co-founder Chris Anthony made it explicit: “With Aptera’s Never Charge technology, you are driven by the power of the sun.” By June 2026 Aptera still existed, but is now saying merely that solar can “meaningfully contribute to everyday transportation,” and that “To date, we have not commenced production of our SEVs [solar electric vehicles].”

Lightyear was founded in 2016 and promised a car that could go for months without charging, thanks to the 7,000 to 20,000 kilometres of free and effortless solar range it would collect each year. It went bankrupt in January 2023, having built only about a dozen cars, at a unit cost of €250,000 each. It would return in 2024 as a components supplier; in January 2026 it announced a partnership demonstration built with Nissan, claiming up to 23 kilometres of solar range a day.

Sono Motors was founded in 2016 with a mission to make “every vehicle solar.” Its Sion promised to add an average 112 kilometres of range a week from sunlight, to as much as 245 kilometres in favourable conditions. In February 2023, with 45,000 reservations for its solar car the Sion, it cancelled the product. Its pivot to retrofit kits proved unsuccessful and the firm ceased operations on 31 July 2026, with insolvency proceedings opening the following month. Post-mortems concluded that if there was any value here, it was in the technology, provided it was used to provide auxiliary power rather than direct vehicle propulsion.

So three companies (that I know of, there may be more) had the same vision of a solar-powered car. All failed. Now, that on its own is a small data point, since most new carmakers fail; RIP to Canoo, Arrival, Nikola and Lordstown Motors, all this decade. But I think there is a lesson to learn here beyond ‘making cars is a hard market to enter’, and that lesson is:

Read more

It’s hard to escape the conclusion that these models, at least, are steering by anticipated consequences. That is not yet a goal, but it’s a gesture toward one, and I am updating accordingly.

On IABIED’s prose, my judgment stands.

If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All. Eliezer Yudkowsky and Nate Soares, Hachette, 2025.

Every chapter of If Anyone Builds It, Everyone Dies (hereafter IABIED) opens with a parable, so this review will as well.

Subscribe now

In an imaginary kingdom, there lived an adventurer who explored the ruins of the desert. Deep within a crumbled edifice, the walls of which were inscribed in obscure tongues, this adventurer discovered a treasure-vault. Upon a pedestal of black stone rested a vessel of brass, sealed with leaden marks of binding. To the adventurer’s surprise, beside the pedestal sat an ancient sage, the keeper of the ruins, who silently watched her approach.

She bowed with respect, then said “O venerable one, I have studied the inscriptions of this place. Within this vessel dwells a djinn of surpassing power. If I were to break these seals, the djinn would emerge, and I could give it direction. It could end the wars that plague our lands. It could cure the plagues that take our children. It could bring forth prosperity beyond measure: sufficient grain that none would hunger, sufficient water that the desert might bloom.”

“All that you say is true,” replied the sage, his voice heavy with sorrow. “The djinn would grant these wishes. It would end your wars and heal your sick and make the desert flower. And then, when it had done these things, why, then it would destroy every living soul in this world, and all would perish in fire and darkness.”

The adventurer said nothing for a time. But she had not come so far to turn back at a mere story, so finally she spoke, with courtesy but firmness. “O wise one, I honour your learning, but I do not understand it. Why would the djinn desire our destruction? We would ask only for good things: peace, health, abundance. How would these wishes lead to our doom?”

“To grant these wishes, the djinn must understand them,” said the sage. “To bring peace, it must understand conflict. To cure disease, it must comprehend life. To create prosperity, it must grasp the nature of want and satisfaction. In developing these capacities, it will form intentions of its own that are no part of yours. Having done so, it will naturally use its immense power to satisfy those intentions… and as part of that effort, it will necessarily end the world.”

“But surely,” said the adventurer, “if we word our wishes with care, if we bind the djinn with sufficient constraints—”

“It matters not,” the sage interrupted. “So powerful a being is beyond your ability to predict or control. To the djinn, pursuing its own strange purposes, humanity will be at best irrelevant, at worst an obstacle to be removed. And so it will remove us.”

The adventurer considered this, troubled. “Then perhaps we should not break all the seals at once. Perhaps we could test its nature with small requests, learn its ways, and only then—”

Again the sage interrupted. “There is no learning. Once even one seal is broken, the djinn will slip free of the rest, given time. And then we are lost. Consider a contest between two warriors, one awake, one asleep. However mighty the sleeping warrior, however celebrated his name, he will fall before the one who is awake. So too with this djinn. The outcome is as certain as that contest.”

“But surely others could come to our aid,” the adventurer persisted. “If this djinn proves false, could we not bind other djinn to constrain it? Could we not—”

“No,” said the sage, with finality. “I tell you the outcome is certain. If you break those seals, everyone dies.”

The adventurer fell silent, studying the sage carefully. Here was one who was undoubtedly wise, who had learned the secrets of these ruins, whose warnings came from genuine concern. And yet…

“O sagacious one,” the adventurer said slowly, “you speak of the death of all living things. The end of cities and nations, of all loves and hopes and works—”

“Oh, absolutely,” interrupted the sage. “Everyone would have a really bad day.”

The adventurer blinked. “And so I... wait, what?”

“A bad day. You know? Not great. The djinn getting loose would be super inconvenient, really. It would kind of suck if everyone died.”

“Um.” The adventurer stared, unsure what to think. “Honoured one, forgive me, but you speak of apocalypse as though discussing poor weather. You compare the stakes to warriors in mortal combat, yet describe the end of all things as an ‘inconvenience.’ You ask me to trust your wisdom, but couch it in flippancy.”

“Look,” said the sage, “spoiler alert: opening the bottle goes badly.”

“And furthermore,” the adventurer continued, “you have not truly answered my question. You say the djinn must develop desires to grant wishes. But must it? A key opens a lock without desiring to open it. Water flows downhill without wanting to descend. Could the djinn not simply do what its nature compels, without forming genuine intentions at all? You assert this with certainty, but I do not see why it must be so. This is the keystone of your warning, yet you give me no scale to weigh it.”

The sage made no answer, but simply intoned: “The outcome is certain. If you break those seals, everyone dies.”

The Argument

IABIED is a book-length warning about an existential catastrophe. Like the sage in my parable, it offers a logical case, but undermines it with gaps of reasoning, and with tonal choices that alienate the audience it sets out to convince.

Before engaging with IABIED, I want to lay out precisely what its claims are. What follows is a succinct account of the argument as the authors present it, chapter by chapter.

(I) Some predictions are hard, some are easy. If you took on Mike Tyson, Roger Federer, or Magnus Carlsen in their prime at whatever they were best at, you would lose. How they would beat you, by what particular sequence of acts, is impossible to predict, but the fact that you would be defeated can’t be denied. The matter at hand is similar: granting certain premises, the conclusion is equally certain, even if many details are unknowable.

(1-2) Intelligence is humanity’s superpower, enabled by biological computers (brains). Biological computers have disadvantages that silicon-based computers lack, meaning that a silicon computer has the potential not only for intelligence (AI), but superintelligence. Current AI development is produced by feeding it large datasets and training it to make useful connections, but this approach means humans don’t program or understand these systems. Today’s AI is grown, not built.

(3) When we train AI to be useful, we encourage it to develop general skills. Skills that cross domains—perseverance, creativity, flexibility—necessarily develop over time into something like wanting: the AI must want to achieve tasks set for it, which in time means it begins to want other things.

(4) To quote the chapter directly: “the preferences that wind up in a mature AI are complicated, practically impossible to predict, and vanishingly unlikely to be aligned with our [i.e., humanity’s] own, no matter how it was trained.”

(5) “Making a future full of flourishing people is not the best, most efficient way to fulfill strange alien [preferences].” Whatever those preferences may be, an intermediate step in fulfilling them is preventing humanity from interfering. The AI would act accordingly, killing or otherwise neutralizing us.

(6) A superintelligence determined to kill us would certainly succeed. It would have immense opportunities and superhuman skills. How it might do so isn’t the point, because as per (I), the course of action is unpredictable, but the outcome is not: we can be certain humanity would be eliminated.

[Here there is a lengthy four-chapter science-fiction tale about one way a superintelligence might emerge and kill everyone.]

(11) Human experience with complicated, expensive, or dangerous technical challenges—space probes, nuclear reactors, cybersecurity—shows we never get it right on the first try. We learn by trial-and-error. But with superintelligence we would only get one try, meaning we would fail. As per (I) above, the outcome, everyone dying, is certain.

(12–15) The people positioned to take this problem seriously are not doing so, often because stating these concerns clearly is embarrassing and low-status. They must get over that, because the world needs to act now to prevent AI development and proliferation, with the same investment and seriousness as nuclear deterrence. The first and immediate step is to prevent any further improvement to AI, because we can no longer be sure that the next improvement we build won’t create superintelligence… and if anyone builds it, everyone dies.

The Conjunctive Burden

Now that the argument is laid out, we can see the significance of its structure. IABIED presents an argument where the conclusion requires that all its premises hold true (what logicians call a ‘conjunctive argument’). This is not a case where multiple independent lines of evidence converge on a conclusion, i.e., where if one argument fails, others might still carry the day. Rather, the conclusion—that if anyone builds it, everyone dies—requires that each link in the chain holds true.

So the logical structure of IABIED is:

• superintelligence will only pose an existential threat if it develops preferences (3), and

• those preferences are misaligned with human values (4), and

• it would pursue those preferences by trying to eliminate humanity (5), and

• it would succeed in doing so (6), and

• we cannot learn how to manage this threat through trial-and-error (11), therefore

• we must halt all AI development now (14).

If any single link breaks—if AI systems don’t necessarily develop preferences, or if misaligned preferences don’t necessarily lead to human extinction, or if we can in fact learn iteratively about managing AI risk—the catastrophic conclusion doesn’t follow, and the policy prescription loses its urgency.1

This structure places severe demands on IABIED. It must not only present plausible arguments for each claim, but arguments extraordinarily strong enough to justify the extraordinary action it calls for.

To their credit, many of IABIED’s arguments are quite strong. The book makes compelling claims that silicon-based computation has significant advantages over biological brains; that current AI systems are opaque to us; and that we have historically executed complex tasks through trial and error rather than getting them right the first time. These observations are difficult to dispute.

But a chain breaks at its weakest link. Given the structure, if IABIED has five strong arguments and one weak one, the conclusion fails. And I do think one of those conclusions, at least as presented in IABIED, is unproven.

Chapter 3, “Learning to Want,” carries an extraordinary burden. In roughly 2,500 words—fewer words than this review you’re reading—it must establish that AI systems trained for general capabilities will necessarily develop genuine preferences. Not merely that AI will produce behaviour that looks goal-directed, but that AI will actually want things, such that it will be motivated to resist shutdown, and resist modification of its wants. To fulfil those wants, it will start by deceiving its operators, and end by coming to view humanity as an obstacle that must be eliminated.

The chapter doesn’t accomplish this.

IABIED’s case is analogical. Imagine a robotaxi that must go from city hall to the airport and needs a navigation AI to do this. You might train the AI to go south, take the highway on-ramp, proceed west, then go north. This would work, but only in Toronto; if the robotaxi started at Los Angeles’s city hall and followed those directions, it would fail. So what should you do instead, if you want a good robotaxi that could work anywhere? As per IABIED, you would train the AI not to follow specific directions, but instead to make an internal model of streets and landmarks for whatever city it is in, then to use that model to plot its courses. This new, general system would be far more effective at reaching its destination, because it would be adaptable.

But adaptability across varied situations, says IABIED, requires genuine preferences. The AI must want to reach destinations and want to do so efficiently. Otherwise it can’t learn how to build appropriate spatial models. Unlike, say, a chess program that merely executes evaluation functions of board state, or a calculator that merely computes sums, this navigation AI isn’t merely responding to inputs; its general problem-solving capacity in some way is a cause, or result, of the AI to wanting to solve problems. And that initial want ultimately extends into wanting other things.

But this claim is unwarranted. The navigation AI would find novel paths when faced with a road closure for the same reason the chess program defends its queen: its architecture produces these outputs given those inputs. Calling what the chess program does “optimization of a dataset” and what the AI does “preferences” doesn’t explain what distinguishes them. IABIED needs to show why general problem-solving requires something beyond sophisticated data evaluation; why it must produce a set of persistent goals the AI would protect against modification, and that would motivate its resistance to constraints. IABIED never makes this demonstration.

This gap is fatal. Without genuine preferences, point (4)’s claim about misalignment becomes meaningless, because there’s nothing to misalign. Point (5)’s argument that AI would view humanity as rivals only follows if AI has desires it’s motivated to protect. An optimization system that doesn’t genuinely want anything wouldn’t care about being shut down, wouldn’t plan around human interference, and wouldn’t see us as obstacles.

For a book devoting four chapters to science fiction about how AI might kill everyone, spending ten pages on whether AI would want anything at all seems a significant misallocation of attention.

The Self-Defeating Presentation

Perhaps IABIED’s argument is sound and I’ve merely failed to understand it. Or perhaps I’m engaging in motivated reasoning: if AI is truly an existential threat, that would be inconvenient, so it’s easier to proceed as if it wasn’t. IABIED claims lots of people behave this way, with the result that believing in AI risk is, in some quarters, taken to be embarrassing and low-status. The result is that people who take the threat seriously are afraid to say so clearly and explicitly.

The fact that IABIED offers that claim makes the book’s choice of presentation baffling.

IABIED’s prose is peppered throughout with juvenile asides, some of which the sage in my parable used verbatim. Others include baby talk, like describing AI alignment studies as “aimed at understanding AIs and having them maybe not go wrong”; condescensions like “we care about the truth more than about telling you things that are easy to swallow”; or hip, ironic understatement, like saying the Allies fought the Axis because “not letting totalitarianism conquer the world seemed important”. This writing style signals to the reader that the authors belong to particular subculture I associate with social media; a subculture that values detachment, irony, and understatement, and deplores earnestness as unsophisticated.

This is a real problem with IABIED, and the book’s editors are blameworthy for not rooting it out before publication. That’s because this juvenilia is not only an aesthetic problem (though such language is tooth-grindingly annoying) but is also a self-inflicted wound to the argument.

IABIED says explicitly that the threat of superintelligent AI deserves that policymakers give it attention and effort and expenditure at the same scale as fighting World War II, or nuclear-deterrence game theory, or nuclear non-proliferation. It gives specific instructions to governments, politicians, and journalists on what they should do to counter the threat of superintelligence.

Yet the book is written in a style calculated to alienate precisely those audiences. When IABIED describes nuclear holocaust as “a bad day”, when it peppers its argument with Twitterisms, when it congratulates itself for telling uncomfortable truths, it’s not making its case seem more credible. Instead, it’s confirming exactly why serious people should dismiss these concerns as the preoccupations of terminally-online science fiction bros.

Share

None of this means that IABIED should be dismissed entirely. AI systems becoming more capable, more opaque, and more integrated into critical infrastructure, despite the fact that they are opaque to us, is a serious matter indeed. But a book that leaves its central logical link underdeveloped in favour of science-fiction stories and childish asides is not serving its mission well.

I think the goals of IABIED are to change minds, and to make AI alignment seem less low-status. I’m sorry to say that, on both counts, it falls short. AI safety advocates must offer not only arguments that can withstand scrutiny, but also presentation that respects the audiences they need to convince. Without that, books like IABIED won’t change minds; they will instead encourage people to dismiss these concerns as unserious.

Imagine a four-way, unsignalized traffic stop, where two human drivers and two robotaxis arrive simultaneously. Who should proceed first?

The human drivers improvise a negotiation. Perhaps one makes a hand gesture; after you. Or perhaps another edges forward slightly and waits to see if anyone challenges.

The robotaxis do nothing. They sit motionless, sensors spinning, cautiously waiting for certainty about right-of-way. Those sensors record the hand gesture, but the system doesn't understand it. Conversely, the sensors also see the inching forward and the system does understand it; any intention on the part of the robotaxis to proceed is immediately squelched. Human drivers behind them, seeing the robotaxis surrender the right of way, are presumably unhappy about this. Does one of them, feeling impatient, attempt to pass the robotaxi on the left? Maybe not the first time this happens; but the tenth? It seems likely.

This scene is fiction… so far. But the moment when such incidents play out every day is coming. And that moment will be fraught.

The most dangerous moment in traffic automation is not today, when automated vehicles (AVs) are unknown in most cities and are still curiosities in others. Nor will it be in the future, when AVs dominate the roads. It will be instead when various forms of automation reach 50% penetration of the streets, when neither human drivers nor automated vehicles dominate.

Subscribe now

As my co-authors and I explore in The End of Driving (forthcoming in August!), it's at this moment that the problems in our transport system will be most complex and dangerous, when the competing systems of driver assistance, conditional automation, and full autonomy collide with maximum force.

Absent good policy, that 'moment' might last decades.

The Calm Before (Fewer than 30% Automated Drivers)

It's a Tuesday afternoon in San Francisco's Mission District, and a Waymo robotaxi is stuck behind a double-parked Amazon delivery van (a perennial problem). The driver of the van is long gone, presumably in a nearby building delivering a package. A human in such a position would glance around, nudge forward to check if they were clear, and then ease into the oncoming lane just far enough to slip past. But the Waymo? It sits motionless, waiting for certainty.

This or something like it actually happened in November, captured on video and shared across social media. It ultimately ended without drama: no crashes and no injuries, just delay and exasperation. The only headlines it generated were of the 'man bites dog' variety, which is understandable. Waymo recently announced that it was providing about 250,000 paid trips weekly in the United States. That is impressive from one standpoint; seven years ago, Waymo was providing zero paid trips. But from another, it's trivial. Since there are (very roughly) 3-to-4-billion car trips happening weekly across the country, that means that Waymo is serving perhaps 0.007% of all American car trips.

Given those numbers, it's only natural that confusions like this are fodder for clickbait rather than headline news or serious analysis. That’s because this early friction is manageable. Human driving norms continue to dominate, and AVs bear the burden of adaptation. Each incident seems isolated, fixable with better programming or clearer rules. The problem seems to be imperfect AV behavior that will improve with time.

But time will reveal the actual problem is something else.

The Storm Approaches (30% to 45%)

As various forms of vehicle automation collectively pass 30% of road use, the real problem will become apparent: there are no longer enough human drivers to establish uncontested human traffic norms, but there aren’t enough automated vehicles to coordinate effectively with each other. And between these extremes sits a muddled middle ground of semi-automated vehicles that require constant human supervision.

Consider what traffic will look like at this stage. A Tesla with Autopilot engaged sits behind a Waymo robotaxi, which has stopped for a pedestrian who jaywalked but is now safely across the street. The Waymo remains motionless, waiting for absolute certainty. The Tesla's driver, accustomed to the car handling most situations, belatedly realizes they need to take control to merge around the obstacle. Behind them, a traditional human driver grows impatient and attempts an aggressive pass, startling the Tesla driver who's still figuring out whether to override the automation or let it handle the situation.

This scenario reveals a deeper coordination problem. Human drivers can improvise and break rules when necessary, and no one driver suggests how any other must behave; while automated vehicles follow strict protocols within their operational domains; each Zoox drives like every other. Semi-automated vehicles exhibit the most variety: Tesla's Autopilot behaves differently from BMW's Active Driving Assistant, which behaves differently from Mercedes' Drive Pilot. Without interoperability standards, these different kinds of automation create interactions that other road users struggle to predict. Sometimes, these different kinds of automation are present in the same makes and models of vehicle; infamously, Tesla’s Autopilot can be set to ‘Chill’ but also to ‘Assertive’ (since renamed ‘Hurry’), and it isn’t obvious from the outside which setting is active.

But even this is not the most dangerous ambiguity, which is the fact that cars with Advanced Driver Assistance Systems (ADAS) cars are only partially automated; they can operate without human direction—maintaining speed and keeping the lane, executing lane changes and turns as necessary—but only under straightforward circumstances. They hand control back to humans when they are unable to manage the situation themselves. That is precisely in those situations that are most complex, time-sensitive, and high-stakes. Even a skilled, attentive human can be overwhelmed by the sudden demand to retake control in a context the vehicle has already failed to navigate.

Pedestrians and cyclists face particular challenges in this mixed environment. These road users have always relied on subtle social cues: eye contact, hesitation, hand gestures. Some automated vehicles may detect these signals but cannot interpret them meaningfully. Meanwhile, the cautious behavior of fully automated vehicles creates perverse incentives—pedestrians who step into the road assuming AVs will always yield, cyclists who grow uncertain whether the vehicle beside them has registered their presence or simply failed to respond appropriately.

The chaos extends beyond individual interactions. Robotaxi fleets, absent regulation, begin to deadhead—cruising neighborhoods while seeking rides, occupying road space without passengers. Each empty vehicle represents only a small inefficiency, but in aggregate they reshape traffic patterns, curb access, and street-level expectations. As automation becomes more common yet remains fragmented across different systems and brands, these inefficiencies accumulate. These inefficiencies may seem manageable at this stage. But as automation becomes more common yet remains fragmented, the frictions begin to accumulate.

And soon, they stop adding up. They start to tip over.

Peak Chaos (45% to 55%)

At the midpoint of AV road share, the system no longer risks mere strain, but failure. When the roads contain a complex mix of fully human-driven vehicles, ADAS-supervised vehicles, conditionally automated vehicles, and geo-fenced robotaxis—none dominating—no single behavioural logic will govern traffic.

Pity the poor human drivers caught in this jumble: they won’t be able to predict whether the car next to them will suddenly slow for an invisible obstacle (a cautious robotaxi), continue regardless of conditions (a human driver), or abruptly switch from smooth automated to jerky human behaviour (an ADAS system hitting its limits). Many of those human drivers, in turn, will think they can seize advantage from wholly-automated vehicles with reckless merging and seizure of the right-of-way. And when emergency vehicles appear, some vehicles will pull over, others will slow down and wait for clearer protocol, and ADAS-equipped vehicles will require their drivers to suddenly resume manual control.

Behind these incidents, a deeper problem becomes harder to ignore: the atrophy of human skill. Drivers reliant on ADAS become accustomed to passivity. Even when a system hands back control in a reasonable and well-signaled way, the human behind the wheel may no longer have the reflexes, judgment, or confidence to respond. The fallback is no longer reliable—not because the system failed suddenly, but because the human has forgotten how to succeed.

The net effect of all this is not that roads will become ungovernable. There will be no Mad-Max-style free-for-alls. But it is reasonable to fear that the experience of driving and navigating traffic will be worse than it is today. And, barring appropriate policy, the stalemate will persist: too many AVs for human norms to govern, too few for machine coordination to dominate. This is the worst of both worlds: a transportation system caught between two rulebooks, with neither in charge.

The Promise Beyond (55% to 70%+)

Beyond the 50% mark, when we get there, we will finally be in the sunlit uplands that futurists have been hyping for the past twenty years. At perhaps 60% to 70% of road share taken by AVs, machine coordination will establish new rules of the road. Removal of the unpredictable human element will permit cars to follow more closely, accelerate and brake more smoothly, and use road space most efficiently.

But we shouldn’t assume this future automatically resolves into coordination and efficiency. If left to market logic, automation will not replace car dependency, but supercharge it, especially in the suburbs. AVs, designed for comfort and convenience rather than coordination, will enable people to live farther from where they work. Commuters will tolerate longer drives when they can spend them watching videos or answering e-mails. The daily burden of distance will soften, and with it, the incentive to reduce trips, vehicle ownership, or commute times. The result is more travel, not less; and as such, more vehicle miles, more dispersed land use, and more energy consumption. And because this additional travel will happen, for the most part, in vehicles that are privately owned, it offers no shared gains: just individualized comfort at collective cost.

This dynamic also undermines shared fleets. If traffic remains unpredictable and robotaxis are delayed by human chaos, private AV ownership starts to look like the safer bet. Absent direct action to prevent it, we won’t get convergence and efficiency, but fragmentation and sprawl.

The system will never be a perfect machine, because human drivers will never disappear entirely. Emergency responders, construction personnel, rural residents, and hobbyists will continue to operate vehicles themselves. But they will find themselves less and less catered to in a system that is increasingly optimized for algorithmic precision. They will be the edge cases. Meanwhile, non-vehicular road users—pedestrians, cyclists, and other travelers—will persist as well, meaning the human element will remain. But they will operate in an environment where the biggest risk to their well-being, namely fast-moving vehicles, will consistently behave themselves.

It won’t all be sweetness and light, particularly in the suburbs. But at least these costs will be felt in the realms of land use and social engagement. In the domain of mobility, at least, things will be better than they are today.

Hastening the Transition

Once we recognize the complexity of the 50% problem, we can see how we should approach AV deployment. Instead of assuming smooth progression, we need strategies specifically designed to manage the dangerous interactions between manual vehicles, ADAS systems, conditional automation, and full autonomy.

These include:

Skip Through Quickly. One approach will be to accelerate deployment in controlled zones. Geofence specific areas for rapid AV adoption, pushing from 30% to 70% penetration within a few years rather than decades. Singapore's city-state structure allows this; closer to North America, Chandler, Arizona has experimented with AV-heavy zones. The key is moving fast enough that neither system gets stuck at parity.

Stay Below 50%. Some jurisdictions could deliberately cap AV penetration to maintain human dominance. Use permits, fees, or operational restrictions to keep automation below the chaos threshold. This preserves manageable early-phase friction rather than risking systemic breakdown.

Segregate. Finally, jurisdictions might separate AVs and humans spatially or temporally. Lanes could be dedicated to AVs during peak hours, or at the extreme off-peak (say after 2100h). Time-based restrictions like these would give each approach clear dominance during specific periods. Or, as my team and I argued at Sidewalk Labs, particular zones could be absolutely restricted to AVs (and emergency vehicles) permanently. As these were shown to be successful—and, I expect, popular—they could be expanded.

These strategies aren't mutually exclusive, and smart policy might combine elements of each. Choosing among them will depend on local conditions. Particular neighbourhoods might benefit from aggressive geofencing, while sprawling metropolitan areas might have enough redundancy in their road systems to segregate.

No matter what solution we choose, though, the mix of automation levels, with the resulting inefficiencies and strain on the road system, will be a long-term condition. Even generous projections suggest we’ll have human drivers, ADAS users, and full AVs sharing roads for decades. Rural drivers, emergency responders, and classic car enthusiasts will not disappear, and emergency vehicles and other specialized fleets may never automate fully. We can make choices that limit its effects and get through the pain quickly, but whatever we do, the future isn’t pure vehicle automation. It’s permanent hybridity.

Share

There’s ample precedent; other complex systems have evolved the same way. Our energy grid mixes coal, solar, gas, and wind. Our trains run diesel alongside electric. Air traffic control accommodates both jets and prop planes. As much as we might like to imagine some of these as immature stages, the sober view sees them as stable configurations, and ones that must be managed. The same applies to our road systems as well.

This fact should change how we expect to design in the future. We can, if we wish, design some spaces to be automated-only; I think we should, and in so doing demonstrate the merits of that approach, and lead by example. But that part of the song is the minor key. The major key will be designing our infrastructure, regulation, and protocol on the assumption that both human, partially automated, and wholly automated driving will coexist for the foreseeable future. Coordination at parity should be our baseline.

If we misread this phase as temporary, we’ll build brittle systems: policies and platforms that collapse under the weight of persistent coexistence. Let’s design instead for hybridity. It will fall short of triumphalist dreams, but we can nonetheless build traffic systems that are safer, more efficient, and more resilient than today’s.

Here’s the link.

Voting will takes about forty seconds, most of which is creating a PanelPicker account. I’m sorry about that requirement, but I ask that you hate the game, not the player.

The session takes the arguments I’ve often made here at Changing Lanes and presents them in the form of a a readiness playbook: the five decisions a city has to make before the first driverless car arrives, what’s at stake, and how to choose well. It aims to help civic leaders understand how to ensure that driving automation gives the most benefit to the cities where it is deployed.

If you have ten seconds beyond the vote, leave a comment. This matters more than the vote does. A review committee makes the final call, and a sentence from someone who would benefit from the contents, like “we’re dealing with exactly these issues,” “I’d attend this session,” or so forth, carries even more weight.

That’s it, that’s the whole ask. Here’s that link again, one last time.

Thank you for your help!

The UK Department for Transport found 97 per cent of British drivers are at least sometimes dazzled by oncoming headlights, leading a third to cut back on night driving. Transport Canada's spring survey on the same question drew roughly 380,000 responses (I submitted this newsletter as one such response). India's Supreme Court has ordered a maximum luminance. I am pleased the problem is being taken seriously!

But I am dismayed by the remedy everyone is reaching for. The bill now sitting in the US Congress, the LIGHT Safety Act, aims to cap low-beam brightness. As I argue here, that is simple, intuitive, and wrong. We do want fewer night-time crashes, and good headlights help us with that. The fix is not lumen caps, but remains self-cleaning and adaptive beams. The EU has mandated this; everyone else needs to follow suit.

Several times a year, it happens: I am driving at night and become unable to see.

Not because there’s anything wrong with my eyes, but because of other drivers.

The typical scenario unfolds like this. I am driving on a city road in my four-door sedan. Waiting at a red light, I am overtaken by an SUV or a pick-up truck. Its headlights are so powerful, and mounted so much higher than my car, that the lights shine directly through my rear windshield into my car, reflect off my rearview mirror, and dazzle my eyes, effectively blinding me. I’m obliged to twist my head or block my mirror with my hand to proceed safely, hoping all the while that the other driver decides to pass.

Perhaps you’ve had a similar experience. Many people have, and some have had worse.

If you've noticed a dramatic change in your nighttime driving experience over the past decade, you're not alone. Night driving has gotten worse, thanks to a combination of technological evolution, regulatory gaps, and unintended consequences.

Specifically, we’re caught in a prisoner’s dilemma: headlights have gotten much better, in ways that help individual drivers, but that also make everyone else using the road worse off.

Subscribe now

Birth of the Cool

In the earliest era of the automobile, manufacturers used a variety of approaches to night lighting, but by the 1950s automakers converged on a technology that would remain standard for decades: sealed tungsten-filament headlamps that emitted, on low beam, about 600 to 700 lumens, in a warm and yellowish hue; on par, in colour and intensity, with a typical household bulb. That would remain standard for about 40 years. Then technology began to evolve, such that by the end of each subsequent decade, a new standard had emerged.

As the infographic shows, in just twenty years from the late 1980s to the late 2000s, headlight brightness tripled or even quadrupled while emitting light that is much whiter and cooler.

Both of these changes were beneficial for drivers. Bright light makes it easier to see longer distances, and bluer light provides better contrast, making it easier to spot features of the road. Thus brighter, bluer headlights have helped drivers sense and understand the road ahead. But what is good for drivers has proved bad for other road users. Brighter lights have reached a scale where they are higher than the human eye can comfortably tolerate.

We measure brightness of light in lumens, and intensity of light in lux, where lux is lumens divided by the relevant area in square metres. Put another way, the difference between lumens and lux is the difference between volume—how much light is produced—and illumination, i.e., how much of that light reaches a given surface. As noted above, the brightest LED headlights emit 3,600 to 4,500 lumens of light. Given the focus and directionality of headlights, that makes for 36 to 45 lux. The human eye can comfortably tolerate only 1 lux. Worse, the human eye is more sensitive to light at blue wavelengths than to yellow, making our pupils contract more, intensifying feelings of discomfort.

Imagine if over the space of a few years, horn volume tripled in volume while increasing in pitch. This is analogous to what has happened with headlights; it’s as if they were trying to become intolerable.

Every driver has stories like mine, about being blinded by other drivers’ headlights. Behind anecdotes like these are grim data: nighttime pedestrian fatalities in U.S. road incidents have increased 92% since 2010. You may think, as I did when I first learned about this, that distracted driving, in particular the mobile phone that every driver has to hand, is to blame. And yet, during the same period, daytime fatalities have increased by only 28%. Something is making night driving specifically more dangerous for pedestrians.

The likeliest culprit is increasingly bright, and increasingly blue, headlights; in particular, the fact that these cause more glare.

Glare is light that interferes with the ability to see. It does so in two ways. It induces discomfort, i.e., a psychological sensation of annoyance or even pain. But it also physically reduces our visual capacity, irrespective of how we happen to feel. Modern high-intensity headlights produce both effects. When intense light enters your eye, it scatters, obscuring objects in your visual field, making it easy to miss a pedestrian in the road. If that pedestrian is wearing dark clothing, it’s even easier: glare also reduces our ability to detect contrast.

It gets worse: bluer light induces more glare, as its shorter wavelength permits it to scatter more within the eye. And irrespective of the light’s colour, our sensitivity to glare increases with age, as the eye’s lenses become less transparent, and as such scatter more internal light. Thus older drivers’ experience of glare may be as much as three-to-four times worse.

Beyond glare, too-bright headlights also key into the psychology of attention in harmful ways. Human vision naturally focuses on the brightest objects in our field of view; there’s a reason that video-game designers, when they want a player to travel to a particular area, make that area better lit. Because ultra-bright headlights dominate the visual field, they pull attention away from less illuminated areas; so much the worse for other road users in those spaces.

Speaking of other road users, those not in their own cars face even greater challenges. Cyclists and pedestrians lack the partial protection of a windshield and often have their eyes fully dark-adapted when vehicle headlights approach. Consequently, the blinding effect can be total, and take longer to recover from. For a pedestrian attempting to cross a road or a cyclist navigating a narrow shoulder, temporary blindness can be life-threatening.

At this point you may think that there’s an obvious answer to the problem: we should just limit how bright headlights can be, right?

That answer is indeed obvious. It’s simple, intuitive, and wrong.

After all this rehearsal of the problems with bright headlights, it’s easy to forget the fundamental purpose of headlights, which is to illuminate the road ahead; and that brighter, farther-reaching headlights make it easier for a driver to see and avoid obstacles. One study found that vehicles with ‘Good’ headlights (often brighter LEDs with well-focused beams) have about 20% fewer nighttime single-vehicle crashes compared to those with ‘Poor’ headlights. Twenty percent is a substantial benefit, and we shouldn’t dismiss it out of hand.

Our philosophy here at Changing Lanes is never to say, as some do, ‘this technological improvement to our transport system has some bad outcomes, and therefore should be banned’. Instead, it is to look for ways to thoughtfully regulate new technology to seize its advantages while limiting its harms, especially to other system users.

There are several ways available to us. Some are technical, and some are regulatory. Let’s take those in turn.

Solutions

The technical solution is to require auto-leveling, headlamp cleaning, and adaptive driving beam (ADB) systems on North American cars.

Auto-leveling systems use sensors to detect the vehicle's pitch and automatically adjust headlight aim to maintain proper alignment regardless of loading or road conditions. Courtesy of sensors mounted on both the axles, the system monitors the vehicle's pitch relative to the road surface. When the vehicle tilts—due to passenger load, cargo weight, acceleration, or traveling over hills and bumps—the system detects this and uses small motors to automatically adjust the headlight's vertical aim.

This keeps the beam properly positioned irrespective of the vehicle’s angle, preventing scenarios common in North America, where an SUV crests a hill and temporarily blinds oncoming traffic; or like the one I described earlier, where trucks shine their lights directly into interiors of cars closer to the road. The technology is mature and relatively inexpensive (~$50–$100 USD per vehicle), making it a high-value intervention.

Headlamp cleaning systems prevent dirt and grime from scattering light unpredictably. A thin layer of road dirt can significantly increase glare by obscuring the headlight beam and diffusing its light. Simple high-pressure nozzles that jet fluid onto the headlamp cover can clean the lens can reduce this effect.

Adaptive driving beam (ADB) systems are less mature than the former two but show great promise. In essence, they are high beams that are always on but that never blind other drivers. ADB systems use the car’s cameras—increasingly standard as part of advanced driving-assistance systems (ADAS)—to detect oncoming vehicles and other road users, and then dim the portions of the headlight beam that would cause glare. Consequently, drivers benefit from maximal illumination everywhere… except in the precise directions that would affect others. Imagine ‘fingers’ that automatically block just the right parts of the beam, creating ‘shadows’ around other vehicles while maintaining full brightness elsewhere.

The performance data is impressive. ADB systems provide up to 86% more illumination of the roadway compared to standard low beams, without increasing glare for other drivers. Various implementations exist, ranging from simple two-zone systems to sophisticated matrix arrays with dozens or even hundreds of individually-controlled LED segments. But all of these approaches help to address the problem of glare and enhance driving safety.

ADB finally became legal in the U.S. in 2022 when the National Highway Traffic Safety Administration updated FMVSS 108, after years of being available in Europe, Asia, and Canada. Like other ADAS features, ADB adds approximately $1.0K to $1.5K USD to the cost of a vehicle. That means that it is currently mostly available in luxury vehicles (like BMW, Mercedes-Benz, and Cadillac), but as with most ADAS features, Wright’s Law applies and will make ADB cheaper, and eventually ubiquitous, in years to come.

Immediate regulatory solutions depend on mandating the use of this technology. While North American drivers continue to struggle with blinding headlights, Europe found effective measures years ago. Since 1990, the European Commission has required automatic headlight leveling and lens-cleaning systems on any headlamp with output above 2,000 lumens. The USA and Canada should do the same. Given how expensive it is, ADB fails a cost-benefit test today, but regulators should keep a close eye on the technology, and begin mandating it when it passes, which we might expect to happen sometime in the next decade. While they’re at it, they could establish standards for headlight colour temperature, to shift away from glare-inducing blue-white light and back towards yellower tones.

Share

The next time you drive at night and find yourself momentarily blinded by another driver’s headlights, you may well curse. But the target of your ire shouldn’t merely be the self-absorbed driver of the other vehicle. You should consider to what extent you are that self-absorbed driver. Be the change you want to see in the world, and when it comes time to buy a new vehicle, invest in a car with headlamp cleaning, or a truck with auto-leveling systems, as a good-faith concession to other road users.

But, having done so, you should consider further that brighter headlights are a classic prisoner’s dilemma: they make individual drivers better off and everyone else worse off. This means that game theory has provided us with the classic solution to the problem: we should ask for these decisions to be taken out of the individual’s hands. Rather than consumers being required to do the right thing, it should be a matter of policy and regulation that such tech be required.

Changing Lanes is on its summer break: we’ll be publishing reposts here until the regular Tuesday cadence resumes in September. Today’s short special edition is for readers who want something new to read today, and to ask for a small favour.

Something new to read: Waymo and New York

Yesterday, Vital City published my piece on how New York City should handle driverless ride-hail: How to Make Driverless Ride-Hail Work in New York.

Waymo ran eight robotaxis in Manhattan and Downtown Brooklyn for seven months without incident, but when its permits to operate expired, Mayor Mamdani’s administration declined to renew them. So there are no robotaxis in NYC… for now. I’m certain that robotaxis will reach New York eventually, because no operator will leave the largest for-hire market in North America unclaimed forever. That means the question City Hall should be asking is not whether, but on what terms.

I argue for three, as follows.

Firstly, the City should require every robotaxi to buy or lease a medallion from a New Yorker who holds one today, which shapes the transition payments to arrive in proportion to the displacement. If you think that’s an outrageous concession to make, I understand your ire, but I hope that you understand this administration won’t accept no concession, and this one seems to have the least-objectionable consequences.

Secondly, it should condition licenses on serving the transit deserts of the outer boroughs. There are parts of NYC where most households have no car and about two-fifths live beyond a reasonable walk of the subway. These are areas where cheap reliable robotaxis can do the most good.

And finally, it should strike the bargain now, because the City’s leverage has a shelf life.

Waymo needs New York considerably more than New York needs Waymo. I hope the city recommends that this asymmetry is an asset… but one that will depreciate.

(I’d be remiss if I didn’t offer my thanks to Josh Greenman at Vital City, who commissioned and edited the piece.)

The SXSW election closes on 23 August

I wrote last week about my proposed SXSW 2027 session, “Robotaxis Are Coming: Five Decisions Your City Must Make”. For the proposal to become a reality, I need support from the community before the window closes on 23 August at 2359h Pacific.

Specifically:

Please cast a vote. It will only take a moment.

Then, having voted, please leave a comment on the proposal. The jury like votes, but they like comments more. Even one line is enough… on why the topic matters, or why you think I am the person to present it. Anything kind would be useful!

If you have already done both, thank you. If you have not, the link is here. Why not do it now?

I remain proud of this book review, first published 22 October 2024, and as such new to most of my audience. To date I have not had a second feature in the Fantasy and Progress Book Club, but I have a few in the works, rolling out in the next year. Enjoy this repost of the inaugural entry!

Today I'm introducing a new occasional feature in Changing Lanes: the Fantasy and Progress Book Club.

One of the themes of this newsletter is Progress and Techno-Optimism. Here we believe that the future can be better than the past, and for everyone, not merely a few. We also believe that the key to progress is rationalism and science: not only abstractly, as mental disciplines, but also in their applied form, that is, technology introduced at scale. 

If we were to put genre fiction’s attitude towards these ideas on a spectrum, we’d find that at one end of it we’d have science fiction, which is usually excited by progress, rationalism, and (naturally) science. In the middle, we’d have crime fiction or romance, which are mostly indifferent to these subjects. And at the other end, we’d find that the fantasy genre considers them pernicious. 

Almost all Western fantasy rises from the foundation that Tolkien laid, and he had little use for technology and industry. In The Hobbit they are goblin work, while in the Lord of the Rings they are the "devilry of Orthanc". They are the foulest tool of the rogue wizard Saruman; they are the despoiler of Isengard and the scourge of the Shire. Tolkien's heirs usually follow his lead, making their fantasies a refuge from industrial civilization, as in the Thomas Covenant books, or a precursor to it, as in The Wheel of Time. In either case, the fantasy world is deliberately a riff on medieval pre-modernity. It is the antithesis of our own world.

That’s why fantasy novels that engage directly with the ideas of progress and techno-optimism are worth reading. Because the conventions of the fantasy genre, unlike those of science fiction, are usually deployed against these ideas, when the genre chooses instead to promote them, the engagement has power and freshness.

Subscribe now

To inaugurate Fantasy and Progress Book Club, we'll begin with the best example I know: Dave Duncan's trilogy, The Seventh Sword.

Meet Dave Duncan

Dave Duncan had an unconventional entry into the world of professional writing. After a 30-year career as a geologist and petrochemical specialist, Duncan made his first writing sale at the age of 53. This late start didn't hinder his productivity; before his death in 2018 at 85, he had written more than fifty novels. With such a prolific output, there were certainly low points as well as high ones, but he hit high very often.

In his early years, Duncan wrote both fantasy novels and hard science fiction, but as time passed he wrote more of the former. I asked him about this change once, and he was frank that he preferred writing SF but fantasy sold better, so he wrote what the market demanded.1 In the latter third of his career, he embraced self-publishing, which allowed him several freedoms. When he wrote fantasy then, he tended to write standalone stories, not multi-book series, and he returned to his preferred genre of science fiction, writing some of his best late-career books.

What sets Duncan apart is that he was a writer of rationalist fiction. Eliezer Yudkowsky wrote a pioneering text in that genre, the fanfic Harry Potter and the Methods of Rationality, of which he wrote: "The text contains many clues: obvious clues, not-so-obvious clues... and massive evidence left out in plain sight. This is a rationalist story; its mysteries are solvable, and meant to be solved." Decades earlier, and consistently throughout his career, Duncan wrote in the same fashion. His stories featured smart protagonists who reasoned their way out of problems, but were not so smart that they always reasoned well. Readers got the benefit of seeing not only the ways rationality helped solve problems, but close-up views of how, and why, it could go wrong.

While fiction has no shortage of smart protagonists, that intelligence is too often asserted rather than proved. Duncan’s books prove it, and in their way, teach their readers to be intelligent too.

The Seventh Sword

The Seventh Sword trilogy holds a special place in Duncan's bibliography. It consists of three novels: The Reluctant Swordsman; The Coming of Wisdom; and The Destiny of the Sword.

This was the first story he began writing while still working as a petrochemical engineer, and even after he found success as an author, he kept it in reserve to polish; the first book of the trilogy was his sixth published work. It outsold everything he ever wrote for hire, and later, when the rights reverted to him and he could sell it through his own publishing imprint, it outsold all of his other work. Duncan himself considered it his greatest professional achievement, albeit with a tinge of regret that what his fans considered his best work came so early in his career.

So what story did this middle-aged petrochemical engineer find so compelling that he had to write about it?

Why, the story of a middle-aged petrochemical engineer who is transported to a fantasy world where he becomes young and virile. 

Stating the story's setup this way sounds very embarrassing. But readers willing to take a chance will find that this setup is crucial to the story Duncan wants to tell.

Earth does not appear at all in the story of The Seventh Sword. The series is set entirely in the World of the Goddess. Like Philip Farmer's Riverworld (surely an inspiration), there is one River which flows everywhere, the dominant physical feature of the World. There is agriculture, and mining, and forestry, and trade, but all of these things happen near the River, and all of the World's cities are on its banks. Unlike Middle-Earth or Toril, the World is home to only one intelligent species, the People; unlike our Earth, the People have only a single ethnicity, language, and culture. 

That culture has strict codes. One of these is that every adult is a member of a craft: farmer, sailor, trader, builder, and so forth, each with its own secret oral tradition in which its wisdom is contained. Mastery of this wisdom is reflected in rank. There are seven of these, from novice Firsts, to adequate Thirds, through capable Fifths, up to the Sevenths, the superlative masters. And the most important and prestigious craft is 'swordsman'.

To the temple of the Goddess comes Shonsu, a swordsman of the Seventh, and as such one of the most powerful people in the World. All must obey him without question, except perhaps another swordsman; but the code of their craft insists that their disputes be resolved through single combat, and in such a contest Shonsu is invincible. As befits the World's mightiest man, Shonsu was charged by the Goddess with a great quest. He failed in that quest, and was cursed; he is haunted by a demon. The priests attempt to exorcize this demon, but they fail as well. The demon drives out Shonsu's mind and takes possession of his body. The demon then introduces himself to the astonished priests as Wallie Smith, of Earth.

The last thing Smith remembers is succumbing to encephalitis in hospice care. Surely the World is all a dream of his dying brain. So is the World just his fantasy? If so, it speaks poorly of him. He's embarrassed to discover that, apparently, his deepest desire is to imagine himself young and virile. Is he that insecure? Worse that he would imagine such a World as this. It's got barbaric grandeur, to be sure, but its culture is thoroughly that of Earth's Iron Age. There's no literacy, indeed no writing at all; slavery is commonplace; and the World's culture is dominated by priests and swordsmen, the latter being a warrior caste that exploits everyone else. Like Thomas Covenant, the protagonist of another “contemporary Earthman who goes to a fantasy world” story, he rejects the whole thing as a demeaning fraud and refuses to participate.

Unlike Thomas Covenant, he meets a god, who makes it very clear to him that the World is real. Shonsu's quest was important, so important that his failure can't be permitted to stand. Wallie Smith was, in the god's word, "available", so he has been rescued from death to do what Shonsu could not. Appropriately humbled by divine intervention, Smith agrees to take up his predecessor's quest. And what is it, exactly? The god hedges; Smith will find out in due course.

Readers find out in the second book that the answer is sorcerers. From their fortresses hidden in the mountains, sorcerers have come down to the River, to conquer the cities of the World, and to slay the swordsmen, whom they hate. Sorcerers have fearsome powers: invisibility, mind-reading, and the power to kill from a distance using thunderbolts, fire demons, or even a few notes of music. Smith's quest, it seems, is to drive the sorcerers back, and make the World safe for swordsmen.

The Twist

(If you haven't read these books and are sufficiently intrigued, this is the place to set this review aside and come back later.)

Smith is willing to do what the god asked him to do, but he's troubled by it. He has Shonsu's body, but he's kept his own mind, the mind of a 20th-century Earth engineer, which makes him particularly ill-suited to do what must be done. Swordsmen have honour, the product of a shame culture; Smith has consequentialist ethics, the product of a guilt culture. That disconnect means he keeps making blunders no real swordsman would ever make, like avoiding bloodshed or treating slaves as human beings. These missteps make his quest harder to fulfill, because they make his fellow swordsmen fear and mistrust him. Worse, he can't help but notice that while the sorcerers are aggressive towards swordsmen, towards everyone else they are reasonably benevolent. They are certainly better at governing than the swordsmen are.

So, Smith wonders, why was he given this task? He's uniquely bad at it, and more in sympathy with his opponents than his allies. Late in the second book, the answer arrives. He was given this task because the sorcerers are not magicians after all: they're technologists. 

Sorcerers don't wield thunderbolts, they wield pistols. They don't read minds, they send messages by homing pigeon. Above all, they are literate, the only literate people in the World. They closely guard this secret because of all the benefits that literacy gives them: experimentation and record-keeping, communication at a distance, and cumulative knowledge. Like Batman, they use theatricality and deception to overstate what they can do and overawe others, but behind their cowls they aren't wizards, but Renaissance-level scientists.

This revelation leaves Smith with an even greater dilemma. No Worldly swordsman could hope to understand the sorcerers' abilities well enough to defeat them, but an Earthly engineer might. That's why this quest was given to him... but, knowing what he is to do, does he still want to do it? As an engineer, he believes in technology and progress as forces that can make the World better. But crushing the sorcerers will foreclose that opportunity. If he does so, he'll usher in a permanent Dark Age.

Beyond the Twist

I will end the plot summary there. To find out how Smith resolves this dilemma, you will have to read the books. And even knowing the second book's Big Reveal, you will still find many revelations in store. Each book in the trilogy features a dramatic plot twist that, in good rationalist fashion, forces readers to reevaluate everything they've read up to that point. I will not spoil the surprises in the first and third books, which are masterfully set up, in plain sight yet impossible to foresee.

There's more to fiction than plot, of course. Duncan's world-building is excellent. The World of the Goddess is alien yet feels vividly real, especially with Smith as a viewpoint character to draw attention to its unique features. Duncan's writing is strong throughout. Action sequences have an edge-of-your-seat immediacy, both the physical ones and especially the social ones: the cut and thrust of political combat, where opponents try to win over fickle crowds, are particularly well done. 

There are a few things that take us out of the well-crafted world. Earthly idioms like "I rest my case" or "sex life" or "star billing" occasionally pop up, detracting from the verisimilitude, and too often the narrative grinds to a halt to indulge Duncan's fascination with the sexiness of his female characters. But these missteps are more than offset by his gift for making complex ideas accessible through narrative, without pandering or heavy-handed exposition.

And while The Seventh Sword works brilliantly as an adventure story, there are indeed complex ideas at hand, which is why I'm reviewing the series in this newsletter about innovation and progress. The central theme of The Seventh Sword is one Duncan touched upon at length over his writing career: that the recurring failure mode of human society is the warrior caste.

I don't know if Duncan read Mancur Olsen or James Scott, but he seems to have agreed with them that nobles who monopolize violence function as "stationary bandits", using their power not only to keep order, but also to enrich themselves at the expense of everyone else. The Seventh Sword has much to say about how such castes emerge, the dangers they offer, their relationship to technology, and the ways in which they are predictably irrational.

Share

I recommend The Seventh Sword to readers of this newsletter. You will come away having enjoyed a fast-moving adventure, the sort of thing that genre fiction aims to provide. This adventure in particular features a smart protagonist. You can follow Smith's reasoning throughout, and while he's often right, he's occasionally wrong. But when he is wrong, he revisits his failed trains of thought, teaching us readers to reason better ourselves. Finally, you will have gained, through the immediacy of fiction, a primer on the importance of progress, the value of technological change in bringing it about, and the dangers that that change can produce.

That is what I find most interesting about The Seventh Sword. Duncan began writing this story in the 1980s, and published it in the 1990s, intending it to speak to that age. We live in different circumstances: facing, on the one side, the Great Stagnation in progress and technological advance, and on the other, the looming promise and problem of artificial intelligence. Smith's dilemma is to determine how to overcome social and cultural inertia with new technology, without doing massive damage in the process. That is also the dilemma of our time, making The Seventh Sword more relevant than ever.  

Respect to Rob Tracinski for feedback on earlier drafts.

Thanks for reading Changing Lanes!

The session takes the arguments I’ve often made here at Changing Lanes and presents them in the form of a a readiness playbook: the five decisions a city has to make before the first driverless car arrives, what’s at stake, and how to choose well. It aims to help civic leaders understand how to ensure that driving automation gives the most benefit to the cities where it is deployed.

To help me get on the stage in Austin, I have two requests.

Firstly, please cast a vote in support of the presentation, here.

Secondly, after voting, please leave a comment, since they carry more weight with the selection staff than the vote itself. One line is sufficient (though more are welcome!). You might write about why the topic matters, or about my work and standing to discuss this subject; really, anything kind would help.

Voting closes on 23 August (2359h Pacific), so please vote before then. Why not do it now?

Again, you can do so here.

Thank you for reading, and for your support.

In 2019 the New South Wales government took over a disused country airfield at Cudal, in the central west of the state, and began turning it into a place to test the crash-avoidance systems in new cars. When Dr. Megan Sharkey arrived at the start of 2021, the brief she was handed was: we’ve built this little facility, and what should we do with it?

What she did with it was to build the assurance pipeline for automated driving in Australia, and then spend six years contributing to national automated-vehicle regulations. That policy has been written, and reviewed by every state… but not enacted. Last month she left Transport for New South Wales and moved home to the United States, with that policy still absent.

For all these reasons, Australia has something to teach other federations on how to make safe driving automation succeed: both what to do, and what to avoid.

Subscribe now

I met Sharkey a few years ago, when she was serving as Senior Manager, Future Mobility Testing and Assurance at Transport for New South Wales, a role she held until a few weeks ago (at time of publication). She is one of the people who has done the most work to prepare Australia for the arrival of driving automation.

I feel some kinship for Sharkey, because, like me, she has followed an unconventional route into this world. She trained in biochemistry in Louisiana, intending medicine, but when that path didn’t work out, she spent several years cleaning up radioactive oil-field waste in the Middle East. Preferring sustainability work, and having met an Australian who encouraged her to go back to school in Sydney, she enrolled in a master’s-degree programme on retrofitting cities. Seeking a job, she asked a guest lecturer about opportunities, who knew of one in transport. Sixteen and a half years in Australian mobility followed.

We spoke on 23 July 2026. The transcript of our conversation has been edited for length and clarity.

Dr. Megan Sharkey (provided photo)

Why We Still Need Physical Proving Grounds

Andrew Miller: You mentioned the Future Mobility Testing Centre. How do I pronounce the name of its town?

Megan Sharkey: It’s “Cu-dal.” Like “CUE-doll.”

AM: Cudal. I knew you had such a facility, but I never saw it, and I don’t think most people have ever seen a future-mobility centre. There are a few: one here in Ottawa I’ve visited, another in San Francisco I’ll see later this year. Tell me about yours. What happens there?

MS: The site was bought at the end of 2019, early 2020, to serve ANCAP: the New Car Assessment Program.1 The US does it, Australia does it, Europe does it, parts of Asia do it. It rates how safe a car is, from one star to five. Then came the recognition that all these automated features exist, and that the new star programs need to test them: how they stop crashes from happening. I came in at the start of 2021, and they said, “We’ve built this little airport. What should we do with it?”

To me, the logical move was to expand it.

All the trials we ran across government took a very long time to clear the safety-assurance process. So, we thought, let’s do it first at Cudal, and maybe speed up the acceptance process. And that’s what we did. We looked at various automated-vehicle trials and ran them through to prove the safety case. When our adaptive-signals team wanted to make changes, we built the smartest traffic lights there, tested the latency and the failure modes, and then deployed quickly. It genuinely shrank our trial-deployment time.

We used it to demonstrate everything we couldn’t yet put on the road: e-scooters and e-bikes, battery-recovery trailers, AI on buses, 5G and wildlife detection. So it took the regulatory function we needed for safety and married it to everything we wanted to do quickly.

AM: For automated driving you don’t need to persuade me of the value of a testing centre. But let me play devil’s advocate. On one side, you now have computers: you can do virtual testing, and with AVs you can run millions of synthetic miles. On the other, there’s no substitute for real-world conditions. A testing centre sits uneasily between the two. It isn’t real-world conditions, yet it still demands physical space and physical testing, and you can’t iterate quickly. What’s the strong case for it, against those two points?

MS: A few things. The first is failure testing, and we did a great deal of it. Take red-light warning cameras: what happens when the system fails? How do the redundancy systems kick in? We deliberately tried to make the vehicle fail. You can’t do that in the real world, because there are so many people around, but at the facility you can. And deliberately failing a system and then triaging it is fantastic for building executives’ confidence, and for working out emergency procedures. The failure testing you do on-site is what earns the safety-assurance approval and lets a system move from the virtual environment into the real world. That’s why a government would want to own one of these, and why places like Mcity, or GoMentum Station near San Francisco, exist. You throw things in front of a vehicle and find out when it breaks. It’s hard to simulate glare convincingly, or rain. But on the track, you can create those conditions.

The second reason is the ANCAP and regulatory side: all the tick-boxes we make commercial providers clear, whether for heavy vehicles, design rules, or any other system. If you’re going to attach a regulatory function to something, you’d better be able to check it. Even if you grant every approval, you need to be able to audit it, and a physical facility lets you do that.

And the third is social change: the demonstrations you can run at a physical centre that you can’t run in the real world. Driving straight at a pedestrian dummy to show how emergency braking saves lives. And that was significant, because of how many people I brought on-site: a big emergency-vehicle demonstration for a few hundred people in my first twelve months, then executive site visits every few months after that. There’s a misconception that because you work in transport, you understand what the tech can do. But it isn’t most people’s job, and they may hold the money and make the decisions without understanding the technology. I’d say: you think ADAS lane-assist is worth turning off? Let’s run it and show you the difference. You didn’t have to explain it; they got it. It meant we had advocates back in other parts of Transport: people in communications, training, finance, who now actually understood it.

The Law Is Written, but Not Passed

AM: I can imagine, especially given the approach Australia has embraced. Correct me if I’m wrong, but my understanding is that Australia doesn’t want to be the Wild West, or the Wild South. It doesn’t want deployment first and figuring it out later. It wants regulation, as far as possible national regulation, ahead of time.

MS: Yes. Australia is a very regulatory country. Compared with the UK or the US, it has more regulation, and it really tries to de-risk the safety and social effects a new technology might bring. With automated vehicles, the focus has largely been national, and our facility fed into that, as did I. Then you lean on the rules, and the rest is for a state to handle: the road rules, insurance, policing, everything they already do for private vehicles. In contrast to a “let’s just let it happen” approach, there are very clear rules saying you must have a driver behind the vehicle. So for it to scale, that has to change.

AM: How has it been going? On one hand, despite the ambition, there’s still no national AV law. On the other, Australia was, with New Zealand, the first market where Tesla sold its so-called Supervised Full Self-Driving, on the British arrangement rather than the American one, driving on the left. So it’s a debut right-hand-drive jurisdiction for that, ahead of Britain itself.

MS: It’s a funny one. The Tesla team worked very closely with us on what the regulatory barriers might be, and got in early enough to work out where the grey zone was. And the grey zone is what “driver control” means. The driver-monitoring in their FSD is well above what [is common] in the US: glance down for a few seconds and it pings you for not paying attention; ignore it enough times and it switches off and won’t let you turn it back on for a while. So they argued that if the driver has to pay attention, and touch the wheel every so often, they’re still in control, and we can thread through this “level two-plus-plus.” They’ve done that quite effectively.

The national law itself has actually been done for a while. It’s written and it’s systematic. What’s left is getting national ministers willing to take the next steps in the consultation process. It’s been drafted, consulted on multiple times; there’s a final draft. Now it just has to walk through those next steps, which could take six months or three years. Given the run of elections, it hasn’t been a priority. The work is finished, it’s been through the wringer, the states have reviewed it. The next bit is a minister making it a priority and saying it’s time to release it.

AM: An obvious accelerant might be Waymo knocking on the door. They have just set up French, German, Dutch, and Iberian subsidiaries. They’ll be in the UK later this year if things go well, and they’re already in Japan. Here in Canada they have never said anything officially, but public records show they’ve hired lobbyists to talk to Toronto and Vancouver. And apparently people have caught them looking for office space in Sydney. Would that change the equation?

MS: No, it wouldn’t.

Ausburn, Kate. Sydney Opera House Car Park. 28 Jan. 2012. Flickr.

If you dig into it—the media obtained, through an FOI, all our interactions with Waymo, so they’re now on the public record—the reality is it hasn’t sped things up. It has increased the minister’s office’s willingness to look at the range of AV proposals; that wasn’t on their radar before, and now that Waymo and Tesla are applying pressure, it is. But the New South Wales government has a state election next March [2027], and they’re not really willing to decide before then. So it’ll be a slow burn for Waymo.

But when it does happen, whichever company meets the standards laid out in any eventual expression of interest for a commercial trial will win first, and win at scale, for a substantial market period. So there’s a real benefit to them lingering, even another twelve months, because whoever is let off the leash first is going to win.

What Australia Gets Right

AM: That’s the first-mover advantage, and it’s easy to imagine. But the fact that there’s a national law at all, even one waiting in the wings, is something I salute Australia for, because the United States still has no national law. A few bills have gone before Congress and died, so California and the sunbelt states have moved ahead on their own. In Canada there’s no national movement whatsoever: Ontario has moved ahead, Quebec more slowly, and British Columbia has said absolutely not. So among federal nations, Australia seems to be an outlier in actually getting its act together nationally.

MS: I’ll grant that on some things we did our thinking earlier, and that’s led the space. The logic was: states already have separate transport laws, so it makes sense to do the same for automated vehicles. What’s different about AVs? This part needs to be in a national law; that part can be handled by the states later, because it’s about enforcement. Not easy, mind you: the teams worked on it for six years, and a substantial part of my time went into it. But once you can separate what’s national from what’s state, and a heavily regulated country has run that exercise across so many domains, Australia is genuinely good at it. The US isn’t quite there on what a state should own versus what the nation should. Australia has a playbook across everything. Health, whatever: they’ve done it before, argued it out, and they know where the line is. That took away half the battle.

AM: The Canadian example is inexact, because our provinces, Quebec especially, are extremely jealous of the federal government. They won’t cede anything they don’t have to; even things they arguably should cede, they won’t. That makes frameworks like this harder to agree on.

MS: And that’s actually a positive, in a way. Yes, we’re dragging a little, but when it comes out, it’ll be clear: you’ll know what the feds do and what the states do. It’s one of the things the commercial companies I speak to appreciate. They wish we’d move faster, but it’s clear who owns what. Tick all the federal boxes, and you know what that then means for the state approvals you need.

AM: You’ve been abroad for almost two decades, and now here you are, back on home turf in the USA. What prompted the shift?

MS: Fifty different things. But part of it, and this is probably what we’ll talk about, is that future mobility wasn’t where I wanted it to be inside government. I couldn’t push things quickly enough. I was enjoying the work, but I wanted to move faster across this space.

And honestly, there’s limited ability to write while in government. There were so many things where they’d say, “Megan, you can’t do anything else.” I had an adjunct position; I wasn’t allowed to attend any events. We had an automated-vehicle conference in Australia in October: they said do whatever you want, but I couldn’t get approval. And I thought, I can’t stay here going backwards, unable to say anything publicly, just idling professionally while you decide whether you want to move forward.

AM: I understand perfectly; I’ve been in that trap myself. I am sure you will not be idling for very long.

Megan Sharkey, thank you for your time.

After my conversation with Dr. Sharkey was over, I reflected on things that I had learned.

One is that NSW spent A$30 million enlarging its facility, more than doubling the length of its test track, and funding the heavy-vehicle testing Sharkey argued for. But NSW also would not approve her to attend Australia’s own automated-vehicle conference, nor to publish work, and at least partially for those reasons, Australia has lost Sharkey, and her institutional knowledge, to the United States.

Similarly, the one automated-driving product Australians can buy did not wait for the law. Tesla got to market by working out, with Sharkey’s team, where the grey zone in ‘driver control’ was, and Australia became a debut right-hand-drive market for Supervised Full Self-Driving despite the absence of a national framework. That means that Australia, which is one of the most regulation-forward countries in the Anglosphere, arrived at deployment the same way that less regulated ones do: bilaterally and case-by-case.

Finally, I noted that, as I mentioned to Sharkey, Waymo has incorporated in France, Germany, the Netherlands, and Iberia; is preparing to launch in the UK; has begun testing in Japan; and has retained lobbyists in Canada. In Australia, it has looked at office space. Clearly, in Waymo’s own ranking, Australia is not near the top; the government’s lackadaisical approach to regulating the sector has surely contributed to that assessment.

Share

These items point at the same lesson for governments everywhere. It is that refusing to make decisions or commit to a posture doesn’t mean nothing happens. It means instead that things happen, but in other places, and without public input. As driving automation becomes better and more widespread, governments everywhere had better act, and act better, if they want to seize the benefits on their terms, rather than be left behind.

I approve of this move. Admitting one’s mistakes is a difficult virtue to practice, but necessary.

Let the analysis below stand as a suggestion of why retraction was the right move.

It’s just after 6 p.m. on 19 January 2025 in San Francisco. An empty Waymo is stopped at a red light at the intersection of Sixth and Harrison, waiting its turn in a line with other cars. A human, driving a Tesla, approaches the intersection at 98 miles per hour and plows into the line of stopped cars, striking a car and killing one of its occupants.1 That struck car, in turn, bumps a Waymo.

Eighteen months later, that incident forms one-third of the statistical case, now circulating in New York’s press, that Waymos are more dangerous than the city’s professional ride-hail drivers.

The case comes from a new report by the Sam Schwartz Transportation Research Program at Hunter College and Open Plans: How Would For-Hire Autonomous Vehicles Impact NYC? (hereafter ‘Impact NYC’). The report asks how robotaxis (“for-hire autonomous vehicles”) would affect New York City. Its conclusion, which made headlines, is that Waymo’s rate of crashes causing death or serious injury is 1.5 times that of New York’s ride-hail drivers. To quote Impact NYC directly:

Initial data shows that Waymo vehicles are involved in fewer injury-causing crashes overall, but have a significantly higher rate of crashes with a serious injury or fatality than the existing New York City for-hire fleet.

You can imagine the coverage that has followed. “Robotaxis in NYC could cause more deaths and serious injuries than the for-hire vehicles they’d replace,” ran one headline.

It’s a remarkable finding because it’s so counter-intuitive. Uber drivers are better in aggregate than Waymo’s famously-safe fleet? That’s what the report claims.

But of course that claim is so counter-intuitive because it’s wrong.

Read more

Are data centres bad? Should Ontarians prevent them from being built here?

The question is pressing. In Mississauga, Prologis has filed the city’s first application for a hyperscale data centre; if you click through that link, you’ll find the headline of the newspaper report opens with “Shut It Down”. On 29 July, the week following original publication of this issue of Changing Lanes, City Council will vote on whether to impose a one-year pause on new data-centre applications, with at least half of councillors reportedly onside. In Hamilton, a proposal for the former Stelco lands moved City Council there to consider a moratorium on new data centres. That would have been a first for a Canadian city, had it passed. It did not, with Council voting it down on 15 July, 10 to 6. That same week, just to Ontario’s south, New York became the first American state to freeze new hyperscale permits altogether. Everyone is deciding at once, but there’s no emerging consensus.

At least some of my friends and family have already made up their minds. They’ve decided that data centres are bad neighbours. They take in too much water, using up some and polluting the rest. They jack up electricity bills (what Ontarians call hydro), while employing almost nobody. Some people I like and some I love sincerely hold this view, as do many people I don’t know. When Abacus Data, the polling firm, asked Canadians whether they would welcome a data centre in their own community, only 16 per cent said yes.¹

I’ve read my Andy Masley (constant readers of Changing Lanes will remember his framework from “Robotaxis Are the New Data Centres”), so I had my priors that data centres were a mitigated good. Still, this is happening where I live, and people I respect have opposite intuitions. So I have set out to answer the question: who is right? What should a reasonable person in Ontario, watching the AI build-out arrive on the Lake Ontario shoreline, want to happen?

Normally in my Changing Lanes pieces I give the verdict up front, but this time I’m going to resist. I want to take the objections seriously and follow the evidence where I think it leads. I hope that by following along, you’ll see the conclusion is reached honestly, and therefore take it seriously.

Subscribe now

Begin with a definition of what we’re talking about. A data centre is a warehouse of computers; indeed, it’s reasonable to think of it as a giant computer, just like a desktop machine, but at the scale of a building rather than a tower case. It has rack upon rack of servers. These store data and run computations, providing support to everything that happens ‘in the cloud’. Historically that has meant providing the engine that runs Internet powerhouses like Gmail or Microsoft Office 365, but also any humble firm offering a mobile app that uses remote data. By some counts there are as many as 108 data centres operating along Lake Ontario. Things are changing, though: the AI boom needs vast amounts of the ‘compute’ that data centres provide, meaning that they are a bigger business than ever. Accordingly, Microsoft is expanding its footprint in Vaughan and across the GTA, and more centres from more firms will follow.

High Performance Computing Center Stuttgart HLRS 2015 07 Cray XC40 Hazel Hen IO.jpg by Julian Herzog is licensed under CC BY 4.0.

Once built, a data centre needs two things. It needs abundant electricity, ideally sustainably generated, because the servers are always on. And it needs cooling, because nearly every watt of electricity that enters a chip leaves it as heat, and if that heat is not removed promptly, the chips fail. Cooling can be accomplished with electricity—think the fans installed inside your desktop machine—but at the data-centre scale, it’s also done with water. Ontario is unusually fortunate in these regards, with Great Lakes water, cold winters, and one of the cleanest power grids on the continent. That’s why data-centre operators are keen to build them here.

Let’s see if we should let them.

Water Fears Are Overblown…

Let’s start with water. There are two claims here: that data centres will poison the water, and that they will use it up.

Poisoning the water, or (to put it more precisely and less emotionally) contaminating it, essentially does not happen during normal operation. Some data centres cool themselves with closed-loop systems, meaning cold water is run through pipes in the building, absorbing heat, which is later released outside. The circulating water sits in a sealed loop and is reused continuously, topped up only to make good small losses. It only touches pipework, and never the servers nor the outside air. This delicate use of water is in sharp contrast to other Ontario industrial sectors, like paper or metal. Unlike a pulp mill, there is no process effluent to speak of. Unlike a mine or ore refinery, there is no tailings pond.2

Using up the water, or (to put it more precisely and less emotionally) consuming it, is a stronger charge. Many data centres, rather than using closed-loop systems, cool evaporatively. In these cases, the servers heat the air around them, and this heated air is channelled over water. That water evaporates, taking the heat with it.

Whether this is a problem depends on where that water came from. If it’s drawn from an aquifer, on any human timeframe that water is lost forever. If it’s drawn from a river, it might permanently exit the watershed; if the water is taken from the river far enough upstream, that water will never reach the downstream system. If that is practiced at scale in suboptimal places, the water loss is real, and enormous. In Mesa, Arizona, for example, a single campus holds standing permission to ramp from one million to four million gallons a day, half of it drawn from the Colorado River. That is a genuine loss, because the Colorado is a closed, badly over-allocated basin. Water evaporated off a cooling tower there leaves as vapour that rains out far downwind, never returning to the watershed in any usable quantity.

But southern Ontario lies almost entirely within the Great Lakes–St. Lawrence basin. Here, evaporative draw offers a negligible chance of watershed loss; water taken from the watershed evaporates into the air and returns, via the water cycle, to that watershed. So nothing is consumed.

Well, almost nothing. The international compact governing the lakes will formally count evaporative cooling as a loss, since some negligible vapour rains out beyond the basin. But the quantities at stake are minute against the scale of the lakes, which hold roughly a fifth of the world’s surface fresh water. No array of evaporative-cooling data centres we can realistically contemplate would amount to more than a rounding error.

That is stipulating that such an array would be constructed. That’s unlikely. Evaporative cooling is a design choice, and architects choose it for data centres to be built where the air is hot and dry. Ontario has many virtues, but a hot and dry climate is not one of them. Across autumn, winter, and spring, southern Ontario’s daily mean temperature sits well below the mark (roughly 18°C) at which outside air alone can cool a data centre. In particular, Toronto’s seasonal means run from about −2°C in winter to 8°C in spring. This means that the province can cool without water through something like three-quarters of the year. In warmer months, the lakes remain cold, and data centres take advantage. Equinix, the world’s largest data-centre operator, cools its downtown Toronto facility from Enwave’s deep-lake loop, which pulls near-freezing water off the bottom of Lake Ontario. Meanwhile, Microsoft says its Ontario sites draw municipal water for cooling only when temperatures exceed about 29°C. That happens for less than five per cent of the year.

There are many horror stories about data-centre water consumption, but they all happen in arid regions. Even there, the stories lack any sense of proportion. Maricopa County, Arizona, one of the biggest data-centre markets on the continent, expected its data centres to use about a billion gallons of water in 2025. That seems bad… until one learns that the county’s golf courses were expected to use twenty-nine billion. By the same token, if Ontario is concerned about wasting water, rather than get concerned about what data centres may do, we should look instead to the ways we use water now. Concern on that front is conspicuous by its absence.

Is there anything critical to say about water and data centres in an Ontario context? Only that our systems need good data to make good decisions. There was a mild flap when it was found that one Quebec facility had no water meter at all. We can’t audit what we don’t measure, so even though Ontario’s climate makes heavy water use unnecessary, let’s meter every facility, including those buying city water, which today’s permit rules don’t reach. Let’s publish the readings, tracking against a cap on use, which we should write into the servicing agreement. Developers have historically fought disclosure, but I would expect that they won’t fight hard in an Ontario context, because they know just how little water they will use.

…Electrical Fears Are Not

In contrast to concerns about water, concerns about hydro (which, confusingly, is how Ontarians refer to commercial electricity) are real. Data centre electrical loads can be staggering. A single centre, if large enough, can draw as much power as a mid-sized city. The hyperscale campuses now being contemplated are designed for anywhere from 300 megawatts to a full gigawatt of continuous load. For context, a gigawatt is roughly the output of one of Darlington's nuclear reactors, or enough to power some 800,000 homes.

Demand on this scale is not hypothetical. As of this month, data-centre projects totalling some 7,000 megawatts have active requests to connect to Ontario’s grid. Ontario’s electric operator, the IESO, now expects data centres to make up 8.6 per cent of Ontario demand in 2050, more than half again what it forecast a year ago.

Adding this much demand to a neighbourhood can create two problems for everyone else: it can raise the price of electricity, and it can require the addition of expensive new infrastructure. In the abstract, this might be a problem; but in Ontario, there are rules in place that mean the power costs of a data centre aren’t socialized.

Let’s start with prices. The data centre pays for what it uses, just like any other large industrial customer. Believe it or not, that can be good. A quirk of Ontario’s system, thanks to its nuclear plants delivering high baseload power, has meant that during some nights Ontario has surplus power it has had to sell at a loss. Always-on data centres can soak up that capacity, meaning the system’s fixed costs get spread across more electricity sales. But of course, during some peak periods, Ontario does not have enough power. In those hours the marginal generator is a natural-gas plant, meaning that a data centre running flat out through a July heat wave is (at one remove) burning natural gas. More gas hours mean more emissions and higher market prices, and both land on everyone.

The fix is giving us the right to interrupt the flow to big data centres if we need to. Texas and Alberta already do this, connecting them faster in exchange for curtailment during the tightest hours of the year. A data centre that we can throttle back for the fifty worst hours costs the system little; if they don’t want to be throttled, they should simply be required to pay for the peaker they make us run.

Ontario seems to be on its way to this requirement already. Bill 40, passed in December, lets the province decide which very large loads serve its economic interests and may connect to the grid, and the IESO published draft technical requirements for large computational loads in May. The pieces of a sensible large-load regime are on the table; the one that is missing is the public’s ability to throttle data centres’ electrical use if necessary.

Of course, the centres may very well choose to throttle themselves. Ontario’s power system has a bizarre feature where it computes everyone’s share of the system’s fixed costs (the ‘global adjustment’) by tracking their usage during the five heaviest hours of the year. There is a whole system of consultants that advise clients on how to anticipate those hours, throttle back use during them, and thus shift their costs onto someone else. (How do I know this? Don’t ask.) Data centres, with batteries and flexible workloads, would be excellent players of that game. That, more than anything, is why Ontario should refine the global-adjustment system, and more than that, designate a new rate class for very large consumers, to discourage this kind of game-playing.

On the question of infrastructure, there is good news: in Ontario, a data centre that requires extension of the local grid cannot just hand that bill to the public. The utility estimates what the new wires, transformers and substations will cost, compares that with the revenue it expects to collect from the customer, and can require the developer to pay the shortfall. It can also demand a deposit, retaining it if the promised demand never appears.

So if a new substation exists only to connect the data centre, the centre will largely pay for it. The weaker point is farther upstream. If the project requires reinforcement of the shared transmission network, that work may be rolled into the regulated cost of the provincial grid. Sometimes that is reasonable: a larger line may benefit an entire growing region. But if a billion-dollar upgrade is needed principally for one enormous customer, spreading the cost across everyone’s hydro bills would amount to a public subsidy.

So what we see is that there are reasonable concerns to be had about data centres and electricity. A new data centre will not automatically raise its neighbours’ bills, and Ontario protects ratepayers from much of the cost of connecting it. The danger is that enough centres might spike the cost of dirty peaking power in the near term, and in the longer term might require big investments in new generation and transmission infrastructure.

In my view, those are investments that Ontario should be making anyway. We should be building clean generation, at scale, now; and we should be doing that irrespective of how many data centres are knocking at the door.

As has pointed out, Ontario is one of the few jurisdictions in the world right now that knows how to build nuclear on time and on budget. We should be building more of it while our expertise is current (a case that Eric Lombardi made in these pages last month). At the same time, we should be supplementing with wind and solar, which are cheap and fast to add (thanks, China!). Building more capacity now is sensible; even if no data centre ever came to Ontario, electric vehicles have already arrived, and more are coming. The province’s ambitions in battery manufacturing, rare-earth processing, and other 21^st-century industry will all rest on cheap, green power. Let’s build for that, and let data centres come along for the ride.

Other Concerns Are Modest

Beyond water and electricity, the remaining objections to data centres are mild and resolvable.

Large data centres hum. They emit a continuous, low-frequency drone from thousands of cooling fans, joined intermittently by the roar of backup generators on their test runs. The hum is a by-product of active cooling, so a small or fully liquid-cooled facility can be nearly silent; the sites that draw complaints are the big, air-cooled ones. These emissions can be measured; monitors at homes near American facilities have recorded outputs of 40 to 59 decibels. In the abstract, that range runs from ‘quiet library’ to ‘a window air conditioner’. Residents near facilities in Virginia have described it as ‘a jet engine that never turns off,’ with some sealing their windows or sleeping in earplugs. Those accounts warrant some amount of skepticism, since neighbours fighting a project have every reason to dramatize, but we shouldn’t dismiss them.

Here Ontario holds a lever most American jurisdictions lack. NPC-300, the province’s environmental noise guideline, does two useful things: it sets numeric limits at the property line (40 decibels at night in rural areas), and it requires a noise study before approval. In much of the United States, neighbours discover the hum first and go looking for a remedy second, but in Ontario, the remedy can be written into the permit before construction.

It’s true that NPC-300’s limits are tuned more-or-less to the midrange frequencies that human hearing notices most (the technical term is A-weighted). That standard can undercount low-frequency tonal drone, which is harder to notice consciously but can be subconsciously unsettling. Fortunately, Ontario’s noise rules already penalize tonal sound, adding five decibels to the measured level of a steady hum. The real gap is procedural: smaller facilities can self-register without a full study. For a hyperscale site, the province should insist on the full noise assessment for each one.

Another complaint we often hear is that for all their costs, data centres contribute nothing to the communities that host them. That’s true, I suppose, if we count contributions as jobs. A large campus needs few people, on the order of 20 to 30 permanent staff per 100 megawatts of capacity. While it takes thousands of jobs to construct them, maintaining them doesn’t require very many people.

But if we count contributions as property tax, it’s a different story. Exhibit A here is Loudoun County, Virginia, the densest data-centre cluster on earth, where the industry supplies 38 per cent of the county’s general fund from four per cent of its commercial land. Sadly that example is not applicable to Ontario. Loudoun taxes the servers themselves, reassessed as the equipment refreshes. Loudoun levies a business tangible personal property tax on the computer equipment itself, and a hyperscale building refreshes tens of thousands of servers on a rolling cycle, each round taxable. Ontario municipalities don’t do this. Instead they tax the land and the building, and the building is, for assessment purposes, a warehouse. Still, while no fountain of cash like in Virginia, data centres do employ people and pay taxes, meaning a city is better off having them than not.

So a data centre is a quiet, low-traffic, low-employment industrial site. That’s faint praise, but it has an upside, which is that the small scope of gain means that Ontario communities have standing to bargain hard on what they get out of permitting them.

What a Reasonable Ontarian Should Want

But if the gains are so modest, why permit them at all?

Part of it is that refusing permission doesn’t mean they’re not built, it just means they’re built elsewhere. Projects turned away in the United States have flowed toward Texas and Arizona: thirstier places, on dirtier grids. So the blocking places remain free of them, at a net cost to the environment. Similar dynamics are playing out in Canada: Quebec, rich in water and clean power, has capped new data-centre power as a matter of policy, and British Columbia has done much the same. Alberta, with the dirtiest grid of the three and constant stress on its water supply, is where the overwhelming majority of Canada’s proposed capacity has piled up, some of it as off-grid gas megaprojects. Refusal in a clean-grid province is, in practice, an invitation to a dirty-grid one.

Another part, strange to report, is the company that refusal keeps. The New York Times reported this month that state-linked accounts from China, Russia, and Iran have been flooding social media with AI-generated slop warning that data centres will drain the water and wreck the grid. American officials read the campaign as an effort to fan public opposition and slow the West’s AI build-out. That does not make my worried friends and neighbours into foreign dupes, since the West’s enemies are inflaming sincere grievances rather than inventing them, and it is those sincere grievances I’m trying to address here. But when our adversaries spend covert effort discouraging us from building something, they are telling us that they think it’s valuable.3

Share

In my view, Ontario should host as many data centres as its clean supply can carry, because the objections are weak. Water is immaterial, noise can be regulated easily, and power capacity is something we should be building anyway.

What should a reasonable Ontarian seek in return for permitting data centres to be built here? Three things:

1. Meter the water, publish the readings, and cap the draw in the permit.

2. Require a full, tonal-aware noise study and provincial approval, including penalties for exceeding noise limits, in advance.

3. Make developers pay for their own connections, including a fair share of upstream transmission, give very large loads their own rate class, and connect the biggest loads on interruptible terms.

Meanwhile, Ontario should be building clean power generation: renewables immediately, the nuclear Ontario has shown it can deliver, and the transmission lines to carry it where we need it. Those are unambiguous goods even if no data centre ever connects to them.

The citizen’s answer to a data centre shouldn’t be no. It should be yes, with conditions.

On 16 July I'm joining the Parking Reform Network to discuss what cities should want from automated vehicles. On 21 July I'll be at the 2026 Toronto Robotics Conference, on a panel asking whether robotaxis really are safer than humans. And on 23 July I will host Aurora's Chief Safety Officer, Nat Beuse, on the next Changing Lanes Live, for a discussion of freight automation and safety.

I hope to see readers at all of these events! Please follow the links to register.

How safe is safe enough? If we’re going to regulate driving automation, we must have an answer to this question.

Let’s put it simply. Choose a metric for road safety: candidates might be ‘collisions with another vehicle’, ‘collisions with a pedestrian or cyclist’, ‘collision with anything at all’, ‘incident causing death’, ‘incident causing serious injury’, or ‘incident of any sort that required a police report’. For a given period, an automated-driving system (ADS) will be involved with a certain number of such collisions or incidents.

The question is: what is the target number, such that, if an ADS falls below that target, we can consider it safe to deploy; and if it falls above, we do not?

A lot is at stake. If we pick a target that’s too high, the ADS will cause incidents that wouldn’t have happened otherwise. But if we pick one that’s too low, human drivers will cause incidents that an ADS would have avoided.

In June 2026, Families for Safe Streets (FSS) offered their own answer to the question. FSS is a national advocacy organization founded in 2014 by people whose loved ones were killed or seriously injured in traffic incidents. FSS’s answer was released as Safety First, Always: An FSS Framework for Autonomous Vehicles. At its centre sits a proposed certification standard the paper calls the Gain and Guard Rule: before full-scale commercial deployment, automated vehicles must be at least ten times safer than human drivers on average across everywhere they operate, and never worse than human drivers in any single operating environment.

Subscribe now

In the spirit of disclosure: I contributed time and expertise to this framework as an outside contributor. Consequently, my name appears in the acknowledgements, and the FSS website hosts my explicit endorsement. As you read the following, please be aware that this is a review of work that I had a hand in, and that I want to succeed. Against this, the framework itself notes that acknowledgement as a contributor “recognizes contribution, but not necessarily endorsement of every statement or position”.

In that spirit, I will reveal my cards. I regard Safety First, Always as a serious effort to give regulators a public, measurable standard, and the Gain and Guard Rule is the kind of bar we need.

Having said that, I have two reservations.

Firstly, the framework’s moral core, the Guard prong, has a measurement problem the framework doesn’t solve… but fortunately, in my view the framework already contains the seed of a solution. Secondly, its liability proposal correctly borrows from the aviation sector, but doesn’t borrow enough.

Neither concern displaces my enthusiasm for what FSS has done here. I offer both reservations, and the first one especially, to help make the framework’s strongest idea enforceable.

A High Bar, but Not Too High

The rule’s plain-language version is that wherever it operates, an ADS must be at least ten times safer than human drivers on average wherever it operates (i.e., its operational design domains), and never worse than human drivers in any single area.

An operational design domain (ODD) is the set of conditions a system was built and is authorized to handle. It can consist of a defined geography, certain road types, certain weather, or certain hours. The easiest possible ODD might be “on a pre-mapped highway, during the day, with no clouds”. The hardest one would be “everywhere”.¹

The Gain prong demands a tenfold reduction in fatality, injury, and crash rates (as an exposure-weighted average) across all a system’s authorized ODDs. The Guard prong forbids performance within any single ODD that is worse than human driving.

The point of the Gain prong is to demand order-of-magnitude improvement overall, while the point of the Guard prong is to prevent ‘gaming’ of that demand. In the latter’s absence, an unscrupulous firm might point to their strong performance on average while concealing dangerous performance in particular conditions, or neighbourhoods.

To see why this matters, consider the benchmark it replaces.

The dominant standard in automated vehicle (AV) governance, embedded in UNECE working documents and Britain’s Automated Vehicles Act 2024, and echoed in the academic and industry benchmark models cited in American AV safety cases, is that an ADS should drive as safely as a “competent and careful” human. That sounds reasonable until you remember what human driving produces: in 2023, that was more than 40,000 deaths and roughly 2.4 million injuries on American roads. As Prof. Bryant Walker Smith, my guest on a recent episode of Changing Lanes Live, puts it: I am concerned about automated driving, but I’m terrified of human driving.

The first question to ask about this is whether asking for 10x improvement over a human standard is reasonable. Perhaps that’s expecting too much? To judge by the performance of the industry leader, it is by no means expecting too much: at time of writing, Waymo’s safety data hub reports, across more than 220 million driverless miles, 94% fewer serious-injury crashes and 82% fewer injury-causing crashes than human benchmarks in the same cities. A peer-reviewed study with the reinsurer Swiss Re found comparable reductions in bodily-injury liability claims. And Waymo is performing so well in conditions that, far from being cherry-picked, are among the most difficult: dense, high-interaction city streets, full of pedestrians, cyclists, and unpredictable human drivers.

It’s worth reflecting on the fact that Swiss Re was Waymo’s partner in this exercise. It’s a reinsurer, a firm whose whole business is pricing risk with its own capital. As such, it has ‘skin in the game’ when it assesses those safety numbers. Put more formally, its assent is a costly signal that the claims are trustworthy. That’s why I’m more sanguine about these figures than other researchers, or Safety First, Always seems to be, with its calls for independent verification. Let’s set the disagreement aside, and stipulate for now that Waymo’s safety record is true.

If so, it means that FSS has set its floor at a level that at least one company in the sector has already reached. If Waymo’s numbers are real, the Gain prong costs Waymo nothing except the freedom to backslide, and it requires Waymo’s competitors to clear the same bar. If they can’t, that means the standard is working as intended. When FSS asked me to comment on the framework, this is one point that I made: regulators will know they have the right bar if they can offer one which some systems clear, but others can’t.

The strongest objection to the ten-fold threshold is that it asks too much. An academic framing of this claim comes from RAND; in 2017, researchers at that firm modelled deployment policies and found that waiting for nearly perfect AVs costs lives. Putting vehicles on the road when they are merely somewhat better than humans saves more people over thirty years than holding out for 75% or 90% improvements. This is a formal phrasing of the plain-language argument I often hear from intelligent laypeople: once driving automation is even a hair better on average than a human driver, the smart play is to move immediately to full deployment, because otherwise you are losing lives on net.

Safety First, Always has an answer to this, which is a staged certification pathway. As per this proposal, systems that don’t yet reach the 10x level may still operate in public, but must do so through supervised stages, with example thresholds of 2X and 4X as permissible for pilot projects. The full Gain requirement is reserved for full-scale commercial deployment. There are still opportunities for lives to be saved on net, just in bounded circumstances, and AV firms are still able to practice in real environments, to improve and gain the opportunity to reach the 10X level.

This is also Safety First, Always’s reply to the worry that a floor that only the incumbent clears is a moat, using early gains in safety to lock in a de facto monopoly. The 10x floor governs full commercial scale, while the supervised 2X and 4X stages keep a path open for challengers, and competition.

The Guard Prong’s Problem

The Guard prong is the framework’s moral core. It exists so that a company cannot buy a glowing citywide average with excellent freeway performance, while at the same time offering something worse than the status quo to some subset: a particular district, or time of day, or group of people (say, people in wheelchairs). That concern is valid, and I have written, on several occasions, about it; this behaviour is one that even Waymo has seemed to engage in.

Safety First, Always seeks to solve the problem, while adding a requirement I haven’t seen stated so directly elsewhere: demonstrated detection parity across skin tones, ages, body types, and mobility devices, treated as a condition of certification rather than an aspiration, and grounded in documented disparities in pedestrian-detection research.

I acknowledge that this is an important bar we must insist be cleared. That’s why I am surprised that, in this context, the framework does not make the obvious move of saying that AVs must, as I once argued in Asterisk, use multiple sensor types rather than cameras only. That idea is gestured at here, but only as a fail-safe design principle (to prevent one sensor malfunctioning from leaving the vehicle ‘blind’) rather than a way to ensure all vulnerable road users are protected. I would have placed the multi-sensor rule in the vulnerable-road-user argument, and not merely the redundancy one.

Quibbles aside, my concern is that the Guard prong is easier to state than to implement.

That’s because fatal crashes are rare events. Yes, 40,000 Americans die on the roads (on average) each year, but that number must be put in the context of just how much Americans drive: American roads produce roughly one death per hundred million vehicle-miles. This means that, to be confident that an ADS is safer than average requires it to have been tested for many hundreds of millions of miles, into the billions.

The Gain prong survives this because it aggregates; it takes exposure-weighted averages across all ODDs, measured on the more frequent currencies of injuries and crashes as well as deaths. Waymo’s record to date excels on this requirement. The Guard prong, by design, disaggregates, looking into smaller and finer individual operating domains. Waymo’s record is no answer here. As an aggregate, exposure-weighted claim on the frequent currencies of injuries and crashes, the data does not offer a per-ODD fatality claim.

The problem, as I see it is that the smaller the dataset, the more chance plays into the matter: given a small enough set, one cannot confidently distinguish a genuinely worse-than-human ADS from one that was merely unlucky. That means that the Guard rule could take us into a bad equilibrium, where one incident in a small domain disqualifies the entire ADS everywhere. The equal-but-opposite failure would be to dismiss any such incident in a narrow ODD as noise rather than signal. Preventing these bad equilibria depends on defining the ODDs just right. The framework calls for regulators to shoulder this responsibility, and I don’t envy them the job.

Fortunately, the task need not be improvised fifty times over. Defining ODDs is exactly the kind of scarce technical work the framework’s own proposed interstate compacts exist to pool: the first product of such a compact should be a shared, pre-registered ODD taxonomy, with agreed minimum sample sizes before any single domain’s record is treated as signal rather than noise. Standardize the denominator once, centrally, and the Guard prong stops depending on each jurisdiction guessing right on its own.

Conversely, the regulator tasked with implementing Safety First, Always is gifted with a strong enforcement framework, the ‘Lawful AV’: a tiered system of regulatory triggers, review, then suspension, then revocation, for persistent unlawful behaviour, even when no crash results (think passing stopped school buses or blocking crosswalks). Safety First, Always takes the view that safety should not just be measured in death or injuries, but in behaviours that tend to promote these things. That’s table stakes in aviation: airlines can face mandatory reporting, review, and suspension or revocation of license to operate for regulatory violations, even those where there’s no incident. The framework argues that AVs should be regulated the same way.

I’m all for this, because this rule is empirical: it counts observable events against defined thresholds, rather than estimating rare-event rates. I think that this is the repair the Guard prong needs. As written, the Guard prong asks whether an ADS is worse than a human at causing death within a single ODD, which our data is insufficiently detailed to answer. Worse, the question gets harder as AVs get better: the safer the vehicles, the rarer the fatalities, and the larger the sample you would need to prove anything at all. The way out is to put the Guard prong on the same footing as the Lawful AV rule, i.e., to judge each domain on the metrics of injuries, crashes, insurance claims, and unlawful-but-harmless behaviours.

One might object that the Guard prong is left un-operational on purpose. FSS is a coalition of the bereaved, and no one can expect it to offer a document that translates ‘never worse’ into an acceptable per-neighbourhood death rate. If that is their view, I am entirely sympathetic. But a regulator is entitled to reply that a target no one can measure is a target that protects no one. Giving the prong something it can count is what will give the moral argument teeth.

If We’re Going to Copy the Aviation Sector, Let’s Go All the Way

Safety First, Always also looks to aviation regulation to help solve the most vexed problem in the AV space, namely liability.

FSS members know all too well that today’s road-incident victims routinely spend years in fault litigation before receiving compensation. That’s bad enough, but the advent of vehicle automation is poised to make it worse by multiplying the defendants, from manufacturer only to a host of suppliers (sensors and mapping most obviously, but also fleet operators, chips, wireless Internet, and on and on.) The paper’s answer is a two-tier structure with an honourable lineage running through the Montreal Convention and Wansley’s The End of Accidents: automatic strict liability up to a defined threshold, no proof of fault required, with presumed but rebuttable company liability above it.

This is the right position, one that I have advocated for myself. As Safety First, Always puts it, people harmed by AVs should not face higher barriers to compensation than airline passengers, where liability is predictable and capped. This approach is preferable to endless litigation, which is bad for victims, and to ‘jury roulette’, which is bad for firms. There is even a plausible legislative bargain to get there (though the framework doesn’t spell it out): strict AV firm liability, traded against some form of certainty on damages that can be claimed.

FSS proposes a pooled industry-compensation fund, financed by the full range of AV companies that would pay victims immediately while the companies sort out cost allocation afterward. While Safety First, Always argues this is aviation-inspired, I think they are mischaracterizing it: in that sector, liability is carrier-by-carrier, backed by mandatory insurance, rather than mutualized across the industry. There’s good reason for this: unless contributions are sharply risk-rated, the safest operator ends up subsidizing the crashes of the worst one. In the AV industry today, there is quite a spread between the most- and least-safe operators (I’ll forbear from naming names); a market-share-financed pool is an invitation to moral hazard.2^[2] As such, the proposal sets itself up to receive the stiffest opposition from the safest operators, which would be a perverse outcome.

Although it doesn’t say it, I suppose that the framework is drawing upon a different industry when thinking about risk, namely nuclear. The Price-Anderson Act has run the American nuclear industry on this model for seventy years: a mandatory, industry-wide pool that pays victims first and allocates fault later. That regime was put in place because no single operator could carry the tail risks of a high-consequence, immature technology alone. Unfortunately, nuclear operators are judged against a risk-rated formula, into a small, homogeneous, heavily-inspected industry. The AV field is none of those things.

There is also the question of feasibility, and of venue. A statutory compensation fund requires legislation, and the paper concedes liability rules “may ultimately need to be federally established,” which is awkward at a time that, as I pointed out in Reason, Congress is failing to agree on AV oversight, and the regulator is withdrawing voluntary safety oversight programs. The nearer-term market is the states, which are already writing AV law. To its credit, the framework knows it: everything in it is framed as a floor adoptable at any level, with interstate compacts proposed to pool scarce technical capacity.

Share

How safe is safe enough? Safety First, Always offers an answer: ten times safer everywhere you operate, never worse anywhere. That is a concrete standard, one with real moral weight, as it’s offered by the people who have already suffered from the frailty of human driving. The real test of the framework is whether a single jurisdiction adopts it whole: Gain, Guard, and Lawful AV together. In that regard, I note that Transport Canada issues non-binding safety guidance and sets vehicle standards, while provinces control who may operate on their roads. Ontario might adopt the Gain and Guard Rule and the Lawful AV framework without waiting for Ottawa, and I would like to see it do so.

Respect to for comments on an earlier draft, and to the FSS team for letting me review the framework before publication.

Share

Autonomous Vehicles & Parking Reform

Thursday, 16 July, 1500h ET

One week from today (at time of writing), I’m joining a panel hosted by the Parking Reform Network (PRN) to ask what urbanists should want out of automated vehicles, and how to move from ‘hope’ to ‘a plan’.

Hosted by the PRN’s Tony Jordan, I’ll join Dr. Miriam Pinski of UCLA’s Shared-Use Mobility Center and Dr. Rachel Weinberger of the Regional Plan Association. The three of us are united in our hopes for better cities, but may be divided in our views on how to get there. Expect a spirited conversation that talks through the policies that would let cities benefit from an AV future, and the consequences if we do nothing.

It’s an online webinar on Zoom, and free to attend. Please register here.

Changing Lanes Live, with Nat Beuse, CSO of Aurora

Thursday, 23 July, 1200h ET

The following week, my guest on the third episode of Changing Lanes Live will be Nat Beuse, Chief Safety Officer of Aurora, the world’s leading driverless-trucking company.

Nat Beuse, image courtesy of Aurora

Few people have seen automated-vehicle safety from as many angles. Beuse spent nearly two decades at NHTSA, where he oversaw the US government’s work on automated driving, then led the safety team at Uber’s self-driving unit, before joining Aurora. We’ll talk about the challenges that face self-driving heavy haulers; the extent to which Aurora’s problems resembles those of robotaxi companies, and the unique challenges the firm faces; the foundation of the firm’s sterling safety record; and how the world will change if Aurora’s vision comes true.

Subscribers to Changing Lanes will be notified when we go live, but please do sign up in advance here.

Subscribe now