I read an article in the Wall Street Journal today about my hometown of Austin, and it made me laugh before it made me think. Since Waymo’s robotaxis arrived here in 2024, they have apparently collected $9,325 in parking tickets. Tow-away zones. Metered spots they never paid. A disabled space outside an elementary school. One that idled in front of a church garage for five minutes during Sunday service while parishioners waited. A resident’s complaint in the records reads, plainly, “there needs to be some way to get them to move.”
These figures come from documents the Journal obtained through an open-records request, so I am relaying its reporting rather than confirming the numbers myself.
As a number, $9,325 is nothing. Austin collected $6.3 million in parking fines in 2025 alone, so Waymo’s two-year total is a rounding error. Spread 83 citations across more than 300 cars over two years and the per-vehicle rate is low, probably lower than what a human-driven taxi fleet of the same size would rack up in the same window.
While the dollar figure is trivial, the behavior behind it is not. It is also a near-perfect illustration of the argument I have been making.
The obvious reading, and why it misses
The easy version of this story is that the self-driving car is not ready. Look, it cannot even park. That reading is wrong, and the same article contains the reason.
An independent analysis by the Insurance Institute for Highway Safety found that over more than 50 million driverless miles, Waymo’s crash involvement rate was 68 percent lower than that of human drivers. The hard problem, the one with lives attached, Waymo has solved to a level that beats us. Parking is where it stumbles.
A system can be superhuman at its central task and fail at something a sixteen-year-old handles on the first day with a learner’s permit. This is jaggedness. Andrej Karpathy coined the term for the strange fact that a state-of-the-art model can solve a hard problem and then miss a trivial one, and a field experiment with 758 BCG consultants showed the same thing in the workplace: performance was excellent on tasks inside the model’s zone and worse on adjacent tasks that looked just as easy. The boundary is uneven, and it does not follow the difficulty ranking a human would draw. Driving safely across 50 million miles is the hard task the machine has mastered. Parking lawfully is the easy adjacent task it has not, and no amount of skill at the first predicts skill at the second. I have written about this shape once already this month. An AI office agent filled out seventeen forms in five minutes and then could not upload a file. Same jaggedness, different machine.
The dangerous part is that the failure is invisible from the outside. Watch a car drive flawlessly for an hour and you will assume it can handle a parking lot, because that inference holds for humans. It does not hold here, and the assumption is where the trouble starts.
The parking failure itself is not my thesis. Waymo will patch handicap-spot detection, and that particular fine will stop appearing. But notice what does not happen. Jaggedness does not get fixed. It moves. Patch the parking lot and the uneven edge shows up somewhere else nobody thought to check, because the unevenness comes from how the system learns, not from a single defect waiting to be found.
That is one problem, and it is real. There is a second one in the same article, and it is not a version of the first. It is a different failure entirely, and it is the one my book is actually about.
The failure that no model fixes
Read the part of the article that is not about parking.
On July 8, the National Highway Traffic Safety Administration sent autonomous-vehicle developers a letter demanding that their cars better follow instructions from first responders. The regulator’s complaint was that robotaxis often fail to recognize where they can stop without getting in the way. When a Waymo blocked an active railroad track in January 2025, an officer reported he had no choice but to have it towed. There was no other way to move it.
It does not go away with better driving. A firefighter at a scene, a police officer at a closure, a resident at a blocked garage: each of them has authority over the situation but no means to direct the machine sitting in it. The human is formally in charge and practically helpless. I keep making one distinction in my book, and this is it in the physical world. Having authority over a system is not the same as having the capacity to intervene in it. The officer had every right to move that car. He had no lever to do it, so he called a tow truck.
That gap is the coherence problem, and no amount of driving skill closes it. It is a problem of the connection between a capable system and the people who are supposed to be able to redirect it. You can make the car a better driver every quarter and leave that gap exactly where it is.
Three hundred locally rational decisions
Here is the detail in the article that matters most.
Waymo runs more than 300 robotaxis in Austin. Between trips, the article says, the cars park themselves on public streets to stay near riders and avoid adding traffic. Each of those choices is sensible. Idling near likely demand cuts empty miles and shortens the next pickup. For the fleet, it is the right call every time.
Now add up 300 right calls. You get 300 vehicles independently claiming curb space across one city, each optimizing for the fleet, none of them accountable for what they cost the curb in aggregate. The church-garage blockage was not one rude car. It was the predictable output of a fleet doing exactly what it was designed to do, measured against a shared resource that no one in the system is responsible for.
This is the pattern I spend the book on. W. Edwards Deming showed it in factories long before any of this. Optimize each part on its own and the whole can still degrade, because the parts interact in ways no single part can see. A support agent and a billing agent inside a company can each be flawless and still act on contradictory assumptions about the same customer. Three hundred robotaxis can each park perfectly rationally and still congest a city. The mechanism is identical. The only new thing is that it now runs at the speed and scale of software, on a public street.
Where the analogy breaks, and why the break is the interesting part
My book is mostly about a different situation. Many systems, built by many teams, with no shared owner, colliding inside one company. Waymo is close to the opposite. One company, one software stack, one central fleet manager. The cars are not incoherent with each other. They are all perfectly coherent with Waymo’s goal. The incoherence is between the fleet and the city.
That difference does not weaken the parallel. It sharpens it. Inside a single enterprise, the cost of local optimization eventually lands back on the enterprise itself. It pays its own complexity debt, later and with interest. In the robotaxi case, the company captures the efficiency and the public absorbs the cost. The fleet gets the shorter pickup times. The churchgoers get the blocked garage. The externality lands outside the firm, which means the firm has no natural reason to see it or price it.
Which is where the parking ticket returns, transformed. The tickets are not the failure in this story. The tickets are the city’s answer to it. Austin cannot rewrite Waymo’s software, so it does the only thing available to an outsider. It attaches a dollar figure to each incoherent act and bills it back. A tow-away citation is a coordinating institution forcing a cost back onto the party that created it, because that party will not absorb a cost it cannot see on its own dashboard.
In the book I call that pricing the incoherence, making the party that creates a coordination burden bear the cost it imposes on everyone else. It is one of only a few places you can intervene when you cannot redesign the system directly. Austin is doing it with a parking-enforcement officer and a public complaints database. It may be crude, but it is the right instinct. When you cannot fix the system, you can at least make it pay for the mess, so the incentive to stop finally reaches someone who can.
The instinct is widely shared, which is its own small piece of evidence. I read the comments under the article, and the readers who were not busy mocking it reached, unprompted, for exactly this lever. Charge a flat $5,000 fee for a driverless tow. Impound the car and make the company pay storage, same as a person would. Hold them to the standard you and I are held to. Nobody in that thread proposed debugging Waymo’s curb detection, because none of them can. They proposed raising the price of the behavior, which is the one move available to an outsider who cannot see inside the system and cannot change it. Pricing is what is left when coordination is out of reach.
A note on fairness, since it matters. Waymo pays these tickets like any other driver, and its spokesman said the company expects no special treatment. It contests some citations and has had a couple dismissed. None of that is evasion. It is a company behaving reasonably inside a system that has not yet given it a better way to behave. The point is not that Waymo is careless. The point is that even a careful, centrally managed, genuinely safer-than-human fleet produces coordination costs its own metrics will never show. That is the part that should worry anyone deploying autonomous systems anywhere.
The version of this that has not happened yet
One last thought, and I will flag it clearly as speculation rather than something the article reports.
Today Austin has one large fleet parking itself on the curb. The same article names two more operators already here, Tesla’s Robotaxi and Amazon’s Zoox. Imagine the near future where three or four fleets, each centrally coherent, each optimizing its own vehicles against the same finite curb, all share one city. None of them is incoherent on its own terms. Each is a model citizen by its own dashboard. Together they compete for the same few feet of pavement outside the same church at the same 10 a.m. service, and no one owns the result.
That is the multi-owner version of the trap, and it is the one that looks most like the enterprise problem I actually write about. Many capable systems, no shared view of the whole, a commons that quietly degrades while every participant is behaving well. When it arrives, the city will reach for the same tool it is using now, only harder. It will try to price the congestion, because pricing is what is left when you cannot coordinate the systems directly and you cannot see inside any of them.
There is a sharper edge to a single fleet that is worth one more sentence, because it cuts against the intuition that central control is safer. A fleet of 300 identical cars does not only optimize together. It fails together. Every vehicle runs the same software and leans on the same positioning inputs, so a single upstream fault does not hit one car, it hits all of them at once and in the same way. I made this point about software agents in a recent post: two agents drawn from the same model share the same blind spot, so the redundancy between them is nominal. Here it is 300 machines sharing one blind spot instead of two. Homogeneity buys clean coordination on a good day and correlated failure on a bad one. That is not an argument against central control. It is a reminder that the thing which makes a fleet coherent is the same thing that can make it fail in unison.
The safest car on the road parks in the fire lane. The fleet that adds no traffic blocks the garage. Every decision was locally correct, and the street got worse anyway. That is not a story about cars. It is the story of what happens to any organization, or any city, that fills up with capable systems faster than it builds the means to keep them coherent.
The gap between capable systems and coherent ones is the subject of my book, Coherence, arriving this Fall. If you want to follow the thinking as it develops, join the list at coherise.com. The one-page decision tool from the book is the first thing I send.