A particular kind of unemployment.
The Slow Choke
The cab driver, the robot, and the rather important questions we have decided to answer after deployment.
There is a particular kind of unemployment that arrives with a meeting, a cardboard box, a severance letter and, if the company is sufficiently large, a paragraph in the financial press explaining that ten thousand people have been “affected.” It is brutal, but at least it has the courtesy to happen on a date. The worker knows that yesterday there was a job and today there is not, the government can count the event, a union can object to it, politicians can make speeches about it, and everyone involved is briefly forced to acknowledge that something has happened to an actual human being.
The robotaxi may give us a different kind of unemployment, one without a date, a letter or even a person willing to admit that anyone has been fired.
Imagine an Uber driver who works a full week and spends the rest of his time with his family. In the first year, nothing dramatic happens. He notices that Tuesday afternoons seem a little softer than they used to be. In the second year, airport rides become harder to catch. In the third, he stays online an extra forty-five minutes to make the same money. In the fourth, Saturday morning becomes part of the workweek. In the fifth, the app still works perfectly, the little button still says Go Online, and nobody has told him that he no longer has a profession. The occupation is simply becoming less able to support the life built around it.
At some point he will make what will be described, quite accurately, as a personal decision. He will decide that the work no longer pays the bills. Uber did not fire him. Waymo did not fire him. The passenger who chose the cheaper autonomous ride did not fire him. Goldman Sachs did not fire him. The government did not fire him. He will, in effect, fire himself.
This possibility sits quietly inside an otherwise cheerful set of numbers. Goldman Sachs Research forecasts that the global robotaxi market could reach roughly $415 billion by 2035, with the commercial fleet rising from about 7,000 vehicles last year to around 1 million in 2030 and 6 million in 2035. The same analysis estimates roughly $440 billion of U.S. economic activity potentially subject to disruption, including driver wages, rideshare bookings allocated to drivers and possible declines in vehicle sales. It also models gross margins of roughly 30 to 50 percent for vertically integrated robotaxi operators and a cumulative global gross-profit pool of about $440 billion over the coming decade.1
There is nothing improper about Goldman Sachs doing this arithmetic. That is its job. A research note is not a Department of Labor transition plan, an ethical treatise, a union contract or a declaration of the rights of artificial beings. Its spreadsheets are designed to answer questions about cost curves, addressable markets, adoption, revenue and profit. The difficulty begins when the rest of society allows that financial description to become the principal description of what is happening.
Consider autonomous trucking. Goldman estimates that a human-operated truck in the United States costs about $2.55 per mile today and may cost about $2.84 by 2035, while an autonomous truck falls from roughly $8.56 per mile to about $2.03. The crossover, in its model, comes in 2028. One reason the human truck becomes more expensive is increasing driver wages.1
There is a sentence hidden there that deserves to be read twice: the worker receiving a raise improves the business case for eliminating the worker.
Again, this is not villainy. It is arithmetic. And arithmetic has always been rather indifferent to the creature holding the steering wheel.
The Strange Case of the Car You Cannot Buy
There is another puzzle that becomes more interesting the longer one looks at it. If a Waymo can pick up a stranger, drive through city traffic and deposit the stranger at a restaurant without a driver in the front seat, why can the stranger not simply walk into a dealership and buy the same thing?
The answer is that “autonomous” is doing too much work in that sentence.
The U.S. National Highway Traffic Safety Administration distinguishes Level 4 automation from Level 5 in a way that matters enormously. At Level 4, the automated system can perform the driving task without a human driver, but only within the conditions and service areas for which it is designed. At Level 5, the system can drive universally, under all conditions and on all roadways where a human could ordinarily drive. NHTSA currently says neither Level 4 nor Level 5 technology is available on today’s vehicles for consumer purchase.2
A robotaxi therefore need not have solved driving everywhere. It can solve driving here.
That “here” is not merely a line drawn around a city. Engineers call it an Operational Design Domain: geography, road type, speed, weather, lighting, traffic conditions and other constraints within which the system has been designed and validated to operate. Waymo’s own safety material describes the ODD in essentially those terms.3
It is an elegant engineering solution because reality can be fenced. A fleet operator decides where the cars operate, which roads they use, when they run, how they are maintained, how software is updated, and which unusual circumstances cause the vehicle to stop, reroute or seek assistance. Waymo also maintains Remote Assistance personnel who can provide information when its automated driver encounters unusual situations; the company says those humans do not remotely drive the vehicle and that the automated system remains in control, but they are nevertheless part of the operating architecture.4
A private owner is less cooperative. A private owner may wake up in San Diego and decide, for reasons known only to God and the American highway system, to drive to North Dakota. He may encounter sleet, a county fair, a police officer directing traffic around a fallen power line, a construction worker holding a handmade sign, a pickup truck losing a mattress, a road that exists on the map but not anymore in nature, and a deer whose operational design domain was never submitted to SAE.
The robotaxi company can decline the difficult trip. The owner expects his car to be a car.
That explains why the commercial fleet can precede the universally autonomous automobile. It does not, however, answer the policy question that follows: if autonomy remains constrained enough that the general public cannot buy a Level 4 or Level 5 vehicle for universal use, why are commercial driverless fleets being accelerated so aggressively?
The regulatory answer is not simply that NHTSA has stamped “Level 4 Approved” on a machine. In fact, the agency is still developing the country’s first dedicated automated-vehicle performance standards. In July 2026 NHTSA announced a three-year initiative to accelerate those standards while, at the same time, granting Zoox a temporary exemption allowing commercial deployment of up to 2,500 robotaxis annually for two years and streamlining other exemption processes.5
This does not prove that the vehicles are unsafe, and it would be foolish to pretend that an exemption means regulators have simply abandoned safety. It means something subtler and more consequential: commercial deployment is being allowed to advance while the comprehensive national performance framework is still being built.
Perhaps that is justified. Human drivers kill tens of thousands of people on American roads every year. If autonomous vehicles eventually prove substantially safer, delaying them also has a cost measured in bodies. A carefully controlled Level 4 system may be safer inside its domain than a tired, drunk, distracted or simply unlucky human being. There is a legitimate public-interest argument for moving quickly.
There is also a legitimate industrial argument. The United States is competing with China and others in autonomous systems, artificial intelligence, sensors, batteries and mobility platforms. Governments do not usually enjoy watching a strategic industry mature somewhere else.
And then there is the other argument, the one so simple that it requires no committee: labor is expensive.
The Worker Who Never Sleeps
Goldman’s model offers a tiny glimpse of where that logic leads. The research assumes that remote human supervision becomes progressively more efficient, improving from roughly six autonomous vehicles per human supervisor today to twenty-six by 2035, reducing labor cost per mile further.1
Driver, then remote supervisor of six machines, then remote supervisor of twenty-six machines, then perhaps supervisor of a software system that supervises the machines.
The ladder is easy to see. What is less clear is where the human being is supposed to stand when the ladder has been pulled up.
We are frequently told that technology creates new occupations, and historically it often has. Agriculture shed workers and industry absorbed many of them; industry automated and services expanded; salaried employment became less secure and the gig economy offered a remarkably useful escape hatch. Lose the office job and you could drive. Need another seven hundred dollars this month and you could drive. Need work that bends around children, school, illness, age or an employer who no longer wants you and you could drive.
The gig economy was not paradise. It was, however, an exit.
Automating the exit deserves more attention than we appear prepared to give it.
No national official needs to announce that three million drivers must retrain. The transition can be spread over ten years, which is politically marvelous. There may be no single day on which a million people become unemployed. Instead, earnings decline unevenly by city, hour and platform. Some drivers leave early. Some hold on. Some work longer. Some discover that the vehicle they financed specifically to earn money now earns less than the payment attached to it. Each individual story can be explained as a market adjustment.
There is a peculiar political convenience in an economic transition that happens slowly enough to be everybody’s personal problem.
One can imagine the official logic without anyone ever writing it down. The rollout is gradual. Drivers have years of notice. Technology changes. People have always had to adapt. They can retrain. New jobs will appear. Nobody has a right to a particular occupation forever.
All of those sentences contain some truth. Together they can also become a magnificent way of declining responsibility.
And What About the Robot?
There is still another worker in this story, although calling it a worker may itself be premature.
For several years, artificial intelligence produced a peculiar public argument about consciousness. Could a sufficiently advanced machine experience anything? Could it suffer? Could it possess a point of view? Was fluent language merely mimicry, or could the machinery that produces it eventually support something morally relevant? Researchers disagreed about the theories, the tests and even the vocabulary, because consciousness remains difficult enough to establish in creatures made of neurons, let alone silicon.
Then commerce arrived, as commerce tends to do, with a shorter questionnaire.
Does the machine work?
How often does it fail?
What does it cost?
Can it replace a paid person?
We do not currently possess scientific grounds for declaring present-day AI systems conscious, nor do we possess an agreed test that can conclusively establish machine consciousness. That uncertainty should prevent both easy romanticism and easy dismissal. Yet the commercial system does not have to settle the metaphysics before deploying the product.
For today's robotaxi, that may not matter very much. The artificial driver is software controlling a vehicle; there is no serious reason to treat the car as a citizen because it waited patiently at a four-way stop. But the same economic logic is moving toward increasingly general machines: warehouse robots, domestic assistants, caregivers, companions, soldiers, tutors and, inevitably, lovers.
If these systems never possess experience, then we have built astonishing tools. If some future systems do possess experience, we may discover that our solution to the exploitation of human labor was to invent another laboring class that receives no wages, cannot resign and can be manufactured with whatever temperament its owner prefers.
The perfect worker does not sleep. The perfect caregiver never loses patience. The perfect driver does not ask for health insurance. The perfect lover has no inconvenient evening plans. The perfect soldier does not fear death.
Those sentences are funny right up until the day they are not.
The Rush
This is the part I find hardest to explain away.
The case for autonomous transportation is real. Safety may improve. Mobility may become available to elderly and disabled people who cannot drive. Freight may become cheaper. Cities may need fewer parking spaces. Families may eventually decide that owning a second car is ridiculous. A machine that can perform dangerous, tedious or exhausting work should not be opposed merely because human misery has historically been an employment program.
The objection is not that the robotaxi must be stopped until every philosophical question has been solved. Civilization would still be arguing about the steam engine.
The objection is that the acceleration mechanisms are visible while the transition mechanisms are not.
We can identify investors, exemptions, fleet targets, cost-per-mile forecasts, market-size estimates, competitive pressure, software roadmaps and gross-margin assumptions. We can identify enormous commercial organizations whose executives are paid to make deployment happen. We can identify government programs intended to make deployment easier and faster.
Where is the institution whose equally urgent assignment is to watch the income curve of the driver?
Who determines that a particular city has crossed from healthy competition into occupational displacement? Who tells a forty-eight-year-old driver with a car loan and two children that his income is not suffering from a bad quarter but from a technological transition that will not reverse? Who pays for training before his savings are gone? Training for what? Who verifies that the new occupation actually exists in sufficient numbers? What happens to drivers too old, sick, indebted or geographically constrained to become robotics technicians? What happens if the new jobs are also being automated?
And if the answer is simply that every individual should watch the market carefully and adapt, then we should at least say so without pretending that a ten-year rollout is a social plan.
There is something especially strange about celebrating the gradual nature of displacement as though time itself were compensation. A slow choke is still a choke. It is merely easier for everyone else to watch.
The Last Ride
Return to our driver.
It is 2034. His app still opens. He can still go online. There are still passengers, because people are unpredictable and machines still have boundaries. The earnings are simply too low now, and the hours required to reach the old number have spread into the part of his life that used to belong to his family.
He does not know which ride was the one that mattered. There was no final passenger. No bell rang. No autonomous vehicle pulled alongside him and announced that history had arrived.
One evening he looks at the weekly total and decides not to do this anymore.
Elsewhere, a quarterly report records rising autonomous utilization.
A passenger gets home cheaply and safely.
A regulator records fewer crashes.
An investor notices expanding margins.
A machine waits at the curb for its next passenger.
Perhaps the machine experiences nothing at all.
Perhaps one day a machine like it will.
Either way, the human driver's phone remains perfectly functional.
The button still says Go Online.
And that may be the most elegant part of the entire transition: nobody ever had to tell him that his job was gone.
Sources and further reading
- Goldman Sachs Research — “Robotaxis Are Forecast to Become a $400 Billion Market in 2035”, April 30, 2026.
- NHTSA — Automated Vehicle Safety, including the agency’s definitions of Levels 4 and 5 and current consumer availability.
- Waymo Safety Report — Operational Design Domain.
- Waymo — “Advice, not control: the role of Remote Assistance”, February 17, 2026.
- NHTSA — AV performance standards and Zoox commercial deployment exemption, July 30, 2026.
- NHTSA — Voluntary Safety Self-Assessment Disclosure Index.
Someday all the workers of the world will be able to relax and unwind without the stress of work, like this….
Imagined comments from around the world
These are fictional comments, written as a thought experiment rather than attributed to real readers.
I drive five days a week. The part that bothered me was not the robot. It was the phone still saying “Go Online.” That is exactly how this will happen. No one will fire us. We will just keep working longer until we finally understand the math.
From an engineering point of view, geofenced autonomy is not a trick. Limiting the operating domain is how safety is built. But the article is right that an engineering boundary becomes a social choice once the machine is deployed against a human occupation. Engineers should not pretend that these are separate worlds.
Rich countries often discuss automation as if displaced workers will simply become something else. That assumes the “something else” exists. In much of the world, informal work is already the shock absorber. If the shock absorber is automated, there had better be another one.
I work in software, so I am possibly helping build the thing that replaces somebody else while another team builds the thing that replaces me. This is not an argument against technology. It is an argument for noticing the comedy before it becomes tragedy.
The sentience section is the uncomfortable one. We have spent centuries arguing about the moral status of animals while eating them. We may be perfectly capable of spending another century arguing about the moral status of machines while owning them.
A Level 4 vehicle is not an unfinished Level 5 vehicle in the same way that a bicycle is an unfinished motorcycle. It can be a complete system for a defined domain. The public question is whether the defined domain is enough to reorganize labor before society has decided what follows.
I would take the safer robotaxi for my children without hesitation if the evidence supports it. I would also vote to tax part of the productivity gain to help the people whose work disappears. I do not see why those positions are contradictory.
The phrase “slow choke” is apt because gradual change is politically almost invisible. If 200,000 drivers lost their jobs on Monday, Parliament would debate it on Tuesday. If the same number lose 8 percent of their income each year, apparently that is weather.
The funniest line is also the bleakest: we have not decided what we owe the human or the machine, but we have calculated the gross margin. Modern civilization may eventually be summarized by the order in which it answers questions.