Updating My Priors on Self-Driving Cars

The tech is really good, which should cause urbanists to update their priors about direct and second-order effects.

I finally had a chance to ride in a Waymo last month when I was at a conference in Phoenix. It was an easy, delightful experience; my colleague (with whom I shared the ride from downtown to the airport) and I agreed it felt very normal very quickly. It cost about the same as a conventional Lyft or Uber.

The experience didn’t change my mind about the challenges self-driving cars pose from a policy perspective. But it did nudge me to write that post I'd been meaning to for about a year due to having updated my views.

Self-Driving Cars: The Urbanist Objection

The basic risk with self-driving cars from an urbanist perspective is deurbanization. The main “cost” to driving is the demands it places on drivers’ time and attention, and if those costs decrease, miles driven can be expected to increase. This might not necessarily increase roadway mortality (if the tech is safe) or emissions (if the vehicles are electric) but it would be very bearish for cities.

Broadly available self-driving cars would make transit less appealing, but more generally they would also tend to make the urban lifestyle package much less appealing relative to its suburban counterpart for people above some age threshold. Automated vehicles would be an important channel for this change because they make it viable to live further away and have a longer commute. If the tech gets cheap enough, parents could solve for the chief suburban parenting hassle—transporting kids—by sending them off to school, soccer, and piano in AVs.

The second-order effects for land use would be in line with those caused by connecting farmland to the city core with highways in the postwar era. Add in widely available hybrid and remote work arrangements for college-educated workers and you’ve got a recipe for a huge additional wave of urban stress and suburban flight. We know what this feedback loop looks like.

While cities are powerless to do much about AVs as such, they’re doing very little to make their lifestyle package more appealing to the types of people—young families who have significant flexibility to choose where they live—who are likeliest to jump at AVs and Tiebout-sort into the new suburban lifestyle product they unlock.

Obviously suburbs aren’t new, but I think urbanists, who generally dislike suburbs and correctly observe that the US built environment is already characterized by a high level of suburbanization, discount how much more suburban we can get. It’s a big country. Even most “built-out” metros end in farmland, not natural boundaries. Widely available AVs would expand the range of commutable land and reduce the (true) cost of commuting, thereby providing a huge tailwind to the suburban model.

Automated vehicles would be a huge positive technology shock to suburbanization the way trains and elevators were to urbanization.

Counterarguments

I’ve heard smart urbanists dismiss these risks on three grounds:

1. Uber/Lyft already accomplish a version of this but don’t appear to have been a catalyst for urban flight

2. The tech isn’t good enough

3. If the tech did get good enough, AVs would allow favorable policy changes (elimination of parking, etc.) at the local level that would help mitigate the dynamics I describe

I think the main thing to understand about all three is that the large majority of people want to own a personal car, not transact on a per-ride basis for rides as a service. You want your trunk, your carseats (kids are in these until age seven or eight typically), your level of cleanliness, etc., not a de facto public vehicle with whatever trash, stains, aromas, and other stuff prior waves of passengers left behind. This is why the second objection has been the most serious one. The third makes some optimistic assumptions in an area where I see no evidence for optimism, frankly.

On that last point: if the tech does get good enough, rather than urbanism-favorable policy adaptation I would expect primarily AV accommodation. This might include waiting areas for AVs, pedestrian gates, disinvestment in transit and pedestrian infrastructure, etc. I’m not predicting a doom loop, but this is just to say that however troubled urbanists perceive (again, correctly) US cities to be today on an urbanism dimension, those troubles can get much worse. Imagine thousands of cars circling all day, or parking themselves in some free lot. I think the likelihood of charging them per mile or whatever to curb that is just very low.¹

So, is the tech good enough? I was a stalwart skeptic, but now I’m not. It’s getting better—and now has 2020s-era chip developments and suburbanization working in its favor. These suggest an increased possibility of positive feedback loops between cost reductions and adoption. That’s even before getting into the deregulatory forces about to be unleashed in the new administration. I won’t speculate about those, but at a minimum they are unlikely to push back on any of the above.

Video of a Waymo ride from the back seat.

Self-Driving Cars 1.0 and the Threat to the Urban Lifestyle Package

For many years I was skeptical of automated vehicle technology. During this era—from 2015, when self-driving hype picked up, until about a year ago—my view ran counter to elite consensus but aligned with urbanist sentiment and a small number of dissenting engineers.

I believed engineers were primarily treating self-driving cars as an optimization problem when they hadn't figured out how to make the darn things work safely in the real world. Whereas they saw a problem that was 95% solved, I saw one that remained unsolved, largely because the remaining challenges involved precisely predicting human behaviors in inherently unpredictable urban spaces.

From an urban vitality perspective, the 95/5 division also looked different. The 95% water in the glass might have been Evian, but the remaining 5% was curdled milk. The result was undrinkable.

This seemed technically difficult to address and thus so did the economics, even given the extremely low cost of capital at the time. Here is an excerpt (which I cut for length) of what I wrote about automated vehicles in early 2019 in a draft manuscript of this law review article on transportation policy:

[I]n 2015 Tesla CEO and founder Elon Musk predicted “complete autonomy in approximately two years.” In 2019, complete autonomy does not exist on Tesla or any other car. Rather, by that time, leading companies were shuttering self-driving vehicle divisions, announcing mass layoffs of self-driving car engineers, and conceding their vehicles’ inability to make left turns in suburban areas just outside their own depots. The year before, a self-driving car killed a pedestrian in a crash that experts agree a human driver likely would have avoided or mitigated. No one knows when or even if fully self-driving cars will deploy. (footnotes removed)

It wasn’t just urbanists with their niche policy concerns, I wrote. There were technical reasons to believe progress would be slow.

One self-driving car company CEO and robotics professor estimates that autonomous vehicles are currently ‘0.01% as good as humans,’ and that even if their performance doubles every 16 months, parity will not arrive until 2035.

We urbanists were not just skeptical but fearful. We worried engineers would push government to radically alter rights of way, undermining pedestrian life and urban vitality—much as various forces had done at midcentury. Reckless conduct of the relevant industries past and present seemed to confirm these fears, with one official even quoted anonymously in the New York Times in 2019 calling for pedestrian gates to keep people on sidewalks.

On the whole, the self-driving 1.0 experience seemed to validate the urbanist viewpoint. The overclaims were exposed, from Elon Musk to major ridehail companies to car manufacturers and startups. Beyond the very halting success came high-profile failures: a self-driving Uber killed Elaine Herzberg in Tempe; Uber and Mercedes Benz abandoned self-driving trucking; and in late 2023 Cruise ceased operations after a crash and mounting scrutiny, despite its CEO's blusterous predictions just months earlier.²

Elites’ credulous acceptance of industry claims had real consequences. The conventional wisdom that self-driving cars were imminent gave public officials an excuse to delay or cancel transit investments. There are many examples of the imminence of autonomy as a stated reason (pretextual or not) for not enhancing transit, including in blue cities like Chicago.

Even though the 2010s hyperbole is now properly regarded as such, I’m skeptical that this experience will give way to a kind of revival of what would have happened in its absence.

Me with my dear colleague Nestor Davidson (in cap).

Self-Driving Cars 2.0 and the Suburban Lifestyle Package

As most players have ceased operations, Waymo stands as the major exception. Having maintained a low profile during the 1.0 era, they steadily logged rides and miles while expanding their service areas and city presence. They're now the only serious AV player in the United States, backed by Alphabet's substantial resources despite tighter macro conditions.

Tim Lee has covered Waymo extensively at his Substack (to which I began a paid subscription for basically that reason). Without attempting a comprehensive summary, he has shown that:

• Waymos have demonstrated impressive safety records, besting human-driven cars by a large margin on important safety metrics³

• The company is expanding geographically and partnering with Uber to offer its services in more cities on an accelerated timeline

• The product quality and price point suggest the possibility of mass commercialization

We may be at a genuine inflection point. While Alphabet doesn't break out Waymo's results separately,⁴ Tim said in November on Derek Thompson's podcast that Waymos are estimated to cost Alphabet between $100,000 and $150,000 apiece—including both the vehicle (MSRP: $72,500 base) and the technology (cameras, radar, and LIDAR). I didn’t see any stories about this statement and don’t know how accurate it is, but Tim generally has a high reputation among tech journalists.

This estimate is about one-third to one-half of previous figures and while not cheap it roughly parallels the price range of a Cadillac Escalade. It’s also about what a Cybertruck or top-end Rivian costs. That is to say, there’s a market for vehicles at the price point Waymos may be at already, even though those vehicles do not offer autonomy.

Now, Waymos are not offered for sale to consumers, but if you’re trying to do some napkin math on whether it might be viable to do so, you’d first want to know what they cost to make.

With AVs, the other side of unit costs is operations. Some have speculated that this, not hardware, is the major cost. That might suggest an even bigger advantage for Waymo, however: the cheaper the hardware, the more data Waymo can generate and deliver to Alphabet for analysis (and what company is better positioned to do that?). Whether this can be done profitably probably determines the viability of the enterprise. Alphabet isn’t saying much about this publicly but it is increasing investment when everyone else is retrenching (at a time of higher capital costs no less), which is suggestive.

Where This All Cashes Out

The multiple layers of novelty and complexity that characterize automated vehicles and the regulatory space in which they operate—which includes not just vehicle and traffic law but land use and the landscape of driving regulation writ large—means it’s still kind of silly to make precise predictions, in my view. That being said, I no longer think a general skepticism of their success is warranted. And, if they become viable to own by the end-user, I think all bets are off; we will be living in a new world. The silver lining is there’ll be an opportunity to write a new Crabgrass Frontier—and you can do it from the backseat of a car that’s driving you wherever you wish to go.

1. While I see little prospect of charging per mile driven (often termed a “VMT tax”), I’m also no longer persuaded it’s a sound policy idea. Keying transportation agency revenue to VMT will give agencies more incentive to encourage driving. I changed my views on this after interviewing Erick Guerra at Penn’s urban planning school for a forthcoming episode of the Densely Speaking podcast.

2. GM ultimately pulled the plug entirely on the Cruise business unit in December 2024, but the more relevant timestamp is probably late 2023, when they stopped operating.

3. Tim started making this claim about Waymo directionally in 2023, but has done so more convincingly this year.

4. Reporting by business segment is an interesting subject rooted in a 1976 accounting rule. It is required by the SEC of public companies, like Alphabet, once a business line becomes sufficiently large relative to the company. Alphabet is the fourth largest company in the world by market cap and has a lot of revenue, however. If Waymo got very big or were spun off into a separate enterprise, depending on the structure of the spin, we might get some better data.