The Empty Seat

On the morning of 13 March 2004, fifteen driverless vehicles set off, one after another, from Barstow, California, on a 142-mile course through the Mojave Desert toward Primm, Nevada. The Defense Advanced Research Projects Agency had offered a million dollars to the first to finish within ten hours, with no human help of any kind. The vehicle that went furthest was Sandstorm, a modified Humvee entered by Carnegie Mellon University. About 7.4 miles out, on a switchback in the hills, it drifted off the course, caught on a berm and stuck there; its front tyres, spinning against the ground, caught fire, and the organisers disabled it by remote command. None of the fifteen covered more than 5 per cent of the route. No prize was awarded.¹

Twenty-two years later, cars with nobody in the driver’s seat give about half a million paid rides a week in American cities. Waymo, which grew out of Google’s project, reported 271 million miles driven with nobody behind the wheel by mid-2026 and was valued at $126 billion.² Tesla counts about 1.7 million subscriptions and purchases of software it calls Full Self-Driving, which its own filings say does not make the car autonomous. A test car killed a pedestrian in Arizona in 2018, and a robotaxi company lost its licence in California in 2023 for what it did after a crash it did not cause. Human drivers, meanwhile, killed about 36,640 people on American roads in 2025, by the federal government’s early estimate, and about 1.19 million a year worldwide.³

This essay follows how the car that could not get eight miles across an empty desert came to carry paying passengers through San Francisco, through the decisions that shaped it: to put the intelligence in the road, then in the car; to build a car that needs no driver, or one that needs a driver to watch; to ship early or to wait inside a map; to cut a second safety operator; to let a video speak for itself. Its deeper subject is what happens to responsibility when control passes to a machine. For a century the law, insurance and ordinary morality have assumed that a person is driving. The self-driving car takes that person away, or leaves them in the seat with nothing to do but wait to be needed. Either way, someone has to answer for what the car does.

The road that would drive the car

The first driverless car on an American street was a trick. On 27 July 1925 Houdina Radio Control sent its “American Wonder” up Broadway, steered by radio from a car following behind, until it crashed into a sedan.⁴ Someone was driving it, from a different car.

The serious vision came from General Motors. At the New York World’s Fair of 1939, GM’s Futurama exhibit, designed by Norman Bel Geddes, carried visitors over a model of the United States in 1960, crossed by superhighways on which the cars were kept apart by automatic control. Bel Geddes followed it with a book, Magic Motorways (1940).⁵ In the 1950s the idea became engineering. Vladimir Zworykin of RCA, one of the inventors of electronic television, later recalled his reasoning simply: “control of automobiles should be done by the road.” On 10 October 1957, RCA and the state of Nebraska demonstrated a car guided by circuits buried in a 400-foot strip of highway outside Lincoln, watched by 83 officials. GM tested a buried cable on its own track the following year, and its 1956 Motorama show featured a turbine concept car, the Firebird II, and a film, Key to the Future, in which a family of 1976 hands control to a control tower and cruises hands-free along an electronic highway. Its promoters said autonomous highway driving was a decade or two away.⁶

The design choice of the 1950s was to put the intelligence in the infrastructure. A car following a wire needed no eyes, only a road rebuilt for it. That made the self-driving car a public-works project, and it died on cost; the curators of The Henry Ford, which holds RCA’s brochure, note that the system was too expensive to be practical.⁶ The idea of automated highways rather than autonomous cars lasted into the 1990s.

Machines that learned to see the road

Intelligence moved from the road into the car when cameras and computers became cheap enough to carry. At the Mechanical Engineering Laboratory in Tsukuba, Japan, a team including Sadayuki Tsugawa built a car in 1977 that used a pair of cameras to follow the road; accounts differ on whether it still relied on guide rails.⁷ The decisive work came from Munich. Ernst Dickmanns, a professor at the Bundeswehr University, approached the problem as a control engineer: a model of how the car and the road ought to behave, constantly corrected by what the camera saw. In 1987 his van, VaMoRs, steered and controlled its speed by vision alone over more than 20 kilometres of empty autobahn at up to 96 km/h. The result changed the plans of Europe’s largest research programme on the subject, EUREKA PROMETHEUS, launched that year with carmakers across the continent: at Daimler-Benz’s initiative, it dropped its intended work on guidance by buried wires and adopted vision instead.⁸

In October 1994 two Mercedes saloons, Dickmanns’s VaMP and Daimler-Benz’s VITA-2, drove in ordinary traffic on the autoroute near Paris at up to 130 km/h. In 1995 VaMP drove more than 1,600 km from Munich to Odense in Denmark, steering and controlling its speed, with about 95 per cent of the distance covered without the safety driver intervening.⁸

Carnegie Mellon came at the problem from artificial intelligence. Its Navlab vehicles, funded by DARPA from the mid-1980s, carried a series of experiments, among them ALVINN, a neural network described by Dean Pomerleau in 1989 that learned to steer by watching a human drive: a coarse image of the road went in, a steering direction came out.⁹ In July 1995 Pomerleau and Todd Jochem took Navlab 5, a 1990 Pontiac minivan, from Pittsburgh to San Diego. The computer steered for 98.2 per cent of the 2,850 miles; the researchers worked the accelerator and brakes. Pomerleau predicted that such systems would reach ordinary cars as aids to drivers “long before we see a vehicle driving fully automatically.”¹⁰ In June 1998 Alberto Broggi’s team at the University of Parma drove a Lancia Thema about 2,000 km round Italy’s motorways, 94 per cent of it automatically.¹¹

The United States, meanwhile, had kept faith with the smart road. Congress required a demonstration of an automated highway system by 1997, and in August of that year a consortium led by the Federal Highway Administration and GM ran one on a 7.6-mile stretch of car-pool lanes north of San Diego, embedded with 92,778 magnets, along which a platoon of ten Buicks drove in close formation.¹² The demonstration worked. The programme ended anyway. In March 1998 the Department of Transportation announced it was “transitioning out” of the consortium in order to concentrate on the “near-term deployment of driver warning and assistance.”¹³ It was the first time a government chose driver assistance over full automation, and it would not be the last.

The desert and the city

The next push came from the military. The defence authorisation act for 2001 set a goal that a third of the armed forces’ operational ground combat vehicles should be unmanned by 2015.¹⁴ DARPA chose an unusual instrument: a prize, open to anyone. The failure of March 2004 was, in hindsight, the best thing that happened to the field, because DARPA ran the race again. On 8 October 2005, on a course that began and ended in Primm, 23 vehicles started and five finished. The winner was Stanley, a Volkswagen Touareg entered by Stanford University under Sebastian Thrun, which completed the course in 6 hours, 53 minutes and 58 seconds.¹⁵

Stanley’s team said it relied “predominately” on machine learning and probabilistic reasoning. Its hand-tuned lasers mistook 12.6 per cent of drivable ground for obstacles; a model learned from data cut that to 0.002 per cent, and the camera learned what road looked like from what the lasers had judged to be road. The team measured progress by the mean distance between failures that needed a human, which rose from about one mile in December 2004 to twenty by July 2005.¹⁵ The idea that a self-driving car is improved by data and measured by the miles between interventions, which governed the industry for the next twenty years, is already in that paper.

In November 2007 DARPA moved the contest to a town: the disused George Air Force Base in Victorville, California, where vehicles had to complete about 60 miles of urban driving in under six hours while obeying California’s traffic laws and dealing with one another. Eleven finalists started and six finished. Carnegie Mellon’s Boss, a Chevrolet Tahoe built with GM, won the $2 million prize; Stanford’s Junior was second, and Virginia Tech’s Odin third.¹⁶ Five of the six finishers carried the same sensor, a spinning lidar with 64 lasers built by David Hall, who had entered the 2005 race, failed to finish and decided to sell the sensor instead.¹⁷ Besides robots, the competitions had produced a set of people and a supply chain.

The people who had done it went to Google.

Ten routes in California

Google started its project in 2009 under Sebastian Thrun, who had led Stanford’s winning team in the desert. Larry Page and Sergey Brin set the goal: ten routes in California of about a hundred miles each, every one driven from start to finish without human intervention. The first was completed by December 2009 and the last in mid-2010.¹⁸ When the project became public on 9 October 2010, the cars, six Toyota Priuses and an Audi TT, had driven more than 140,000 miles with occasional human control and about 1,000 with none, always with a trained driver behind the wheel.¹⁹

The test that changed the plan

In 2012, according to Waymo’s safety report of 2017, Google tested a system that drove itself on the freeway in a single lane but needed a driver ready to take over at a moment’s notice. In internal testing, the report says, “human drivers over-trusted the technology” and stopped watching the road carefully.²⁰ John Krafcik, later the chief executive, told the story in 2019 with details the report leaves out, and dated it to 2013: employees with long commutes, filmed by cameras inside the car, were seen texting, putting on make-up, using laptops and, in at least one case, falling asleep. The company “committed then to full autonomy,” he said.²¹

The decision rested on a few weeks of watching a few dozen people, and the only public evidence for it is the company’s own account. But it set the course of everything that followed. The report named the reason: a person asked to switch “from passenger to driver in a matter of seconds, often in challenging or complex situations with little context of the scene ahead” was a weak link that could not be designed away. The cost was time. A driver-assistance product could have been sold within a few years; a car that needed no driver could not be sold until it was finished.

In May 2014 Google showed where the decision led: a two-seat prototype with no steering wheel or pedals, limited to 25 mph.²² On 20 October 2015, by Waymo’s date, Steve Mahan, who is legally blind, rode alone in one through a neighbourhood of Austin, Texas, the first fully driverless ride on a public road by the company’s count. Google did not announce it until December 2016.²³

Autopilot

Tesla took the road Google had rejected. Its cars carried a forward camera running a vision chip from Mobileye, an Israeli company, together with radar and ultrasonic sensors, and in 2015 it switched on, by software update, a set of features it called Autopilot, which kept a car in its lane and at a safe distance on the highway. The driver was told to keep watching with hands on the wheel. In engineers’ terms it was “Level 2”: the car steered and controlled speed, and the human remained responsible for everything.²⁴ The name mattered most. The manual said the driver was in charge; the name suggested the car was.

On 7 May 2016, near Williston, Florida, Joshua Brown’s Model S drove under a lorry turning across the highway with Autopilot engaged, and he was killed. The NTSB found that in the 37 minutes Autopilot had been active, his hands were detected on the wheel for 25 seconds. It named as probable causes the lorry driver’s failure to give way and Brown’s “inattention due to overreliance on vehicle automation,” and found that the system’s “operational design,” which “permitted his prolonged disengagement from the driving task,” had contributed.²⁵

Two months later Mobileye said it would develop no further products with Tesla; the two companies gave different reasons, Mobileye citing safety and Tesla citing Mobileye’s objection to its building its own vision system.²⁶ On 19 October 2016 Tesla announced that every car it produced would have “the hardware needed for full self-driving capability at a safety level substantially greater than that of a human driver,” with the software to follow over the air, and Elon Musk promised a coast-to-coast drive “of full autonomy” by the end of 2017.²⁷

This was the opposite of Google’s bet. Google would sell nothing until the car could drive itself in a small area; Tesla would sell everywhere at once and improve the software in customers’ hands. Google’s bet put the risk on its balance sheet. Tesla’s put part of it on the road.

The race

Between 2015 and 2017 self-driving became a race that carmakers and technology companies believed they could not afford to lose. Uber, whose business depended on paying drivers, announced a partnership with Carnegie Mellon’s robotics centre in February 2015 and within months hired about forty of its researchers.²⁸ In August 2016 it bought Otto, a self-driving lorry start-up founded that year by Anthony Levandowski, a member of Google’s project, for a reported $680 million.²⁹ General Motors bought Cruise, a San Francisco start-up, in March 2016 for a reported $1 billion.³⁰

Speed was the point. In December 2016 Uber began carrying passengers in self-driving cars in San Francisco without the state’s testing permit, arguing that a person in the driver’s seat meant none was needed; the state revoked the registrations of its sixteen cars a week later.³¹ Arizona welcomed what California turned away: an executive order of 2015 allowed testing with minimal oversight.³² In 2017 Waymo, as Google’s project had become in December 2016, sued Uber, alleging that Levandowski had downloaded more than 14,000 files before leaving. The case was settled five days into the trial in February 2018, with Uber giving Waymo shares then worth about $245 million and admitting no wrongdoing.³³

Tempe

On the night of 18 March 2018, on North Mill Avenue in Tempe, Arizona, a Volvo XC90 belonging to Uber’s Advanced Technologies Group was driving itself at 45 mph. Elaine Herzberg, 49, was crossing the road outside a crosswalk, pushing a bicycle.

The NTSB’s report is a precise account of how a system can see something and still not understand it. The system detected her 5.6 seconds before impact. It never classified her as a pedestrian, because “the system design did not include consideration for jaywalking pedestrians.” Her classification kept “alternating between vehicle, bicycle, and other,” and because the design discarded an object’s tracking history whenever its classification changed, the system could not predict her path. It recognised an emergency 1.2 seconds before impact and then entered “action suppression,” a one-second pause in which it held back from braking, built in because of “concerns about false alarms.” No alert was given to the operator. Uber had switched off the Volvo’s own collision warning and automatic braking, and the design “relied instead on the operator’s intervention.”³⁴

The operator was alone. In late 2017 Uber had merged the jobs of two operators per car into one, a change that let it increase the number of test vehicles. During the trip she spent 34 per cent of her time looking down toward the centre console, where, according to her phone records, a television programme was streaming. She looked up about a second before impact.

Each decision in the chain was defensible alone. False braking is dangerous too; a second operator is expensive. Together they placed the whole of the car’s safety on the attention of one person, at night, watching a machine that almost never needed her, which is the situation research on human vigilance had shown for seventy years that people cannot sustain. The NTSB named as probable cause the operator’s distraction, and as contributing factors Uber’s “inadequate safety risk assessment procedures,” “ineffective oversight of vehicle operators” and “lack of adequate mechanisms for addressing operators’ automation complacency,” which it traced to an “inadequate safety culture,” together with Herzberg’s crossing while impaired and Arizona’s “insufficient oversight.”³⁴

Prosecutors decided in 2019 that Uber was not criminally liable. The operator was charged with negligent homicide and in 2023 pleaded guilty to endangerment.³⁵ Uber stopped testing the day after the crash and in December 2020 handed its unit to Aurora, a start-up led by Chris Urmson, who had once led Google’s project, receiving a stake in the combined company and no cash. The unit had been valued at $7.25 billion in 2019.³⁶

The handover, again

Five days after Tempe, a crash in California showed the same problem in a product sold to the public. Walter Huang’s Model X, on Autopilot, steered into a highway barrier in Mountain View, accelerating to 71 mph in its last seconds. The NTSB found that the system’s limitations had steered the car into the barrier, that the driver, likely distracted by a game on his phone, had not responded, and that “the Tesla vehicle’s ineffective monitoring of driver engagement” had contributed.³⁷

The federal regulator, NHTSA, moved more slowly. In 2021 it opened an investigation into Teslas on Autopilot hitting emergency vehicles, and when it closed its inquiry in April 2024 it had examined 956 crashes, at least 13 of them fatal. Its conclusion is the clearest official description of the handover problem in a product: “Tesla’s weak driver engagement system was not appropriate for Autopilot’s permissive operating capabilities,” and the mismatch produced “a critical safety gap between drivers’ expectations of the L2 system’s operating capabilities and the system’s true capabilities,” which “led to foreseeable misuse and avoidable crashes.”³⁸ Tesla had recalled 2,031,220 vehicles in December 2023 with a software update; NHTSA opened a new inquiry into whether the fix worked.³⁹

The promises of full autonomy kept their schedule and the cars did not. The coast-to-coast drive of 2017 did not happen. In 2019 Musk said the system would be “feature complete” that year and spoke of a million robotaxis on the road the following year, a remark later read either as a promise of working taxis or of cars ready to receive software.⁴⁰ In October 2020 Tesla released a “Full Self-Driving Beta” to a small group of owners, warning that it “may do the wrong thing at the worst time.”⁴¹ From 2021 it removed radar and then ultrasonic sensors, betting on cameras alone.⁴² The option’s price rose from $5,000 in 2019 to a peak of $15,000 and fell to $8,000 in 2024; from February 2026 it was sold only by subscription, at $99 a month, as “Full Self-Driving (Supervised),” which Tesla’s filings note “does not make the vehicle autonomous.”⁴³⁴⁴

Counting the miles

How would anyone know a self-driving car was safe? In 2016 Nidhi Kalra and Susan Paddock of the RAND Corporation set out the difficulty. Americans then died on the roads at about 1.09 per hundred million miles. Showing with 95 per cent confidence that a fleet was no worse would take about 275 million miles without a fatality; estimating its fatality rate to within 20 per cent would take about 8.8 billion. Google’s fleet had driven about 1.3 million. Developers could not drive their way to proof.⁴⁵ A year later Kalra and David Groves asked what waiting would cost, and found that deploying cars only 10 per cent safer than the average human, and improving them in use, saved more lives than waiting for near-perfection, though such permissiveness might clash with what the public expects of machines.⁴⁶

Another number did much work in the industry’s case: “94 per cent of serious crashes are due to human error.” It came from a NHTSA note of 2015 that assigned the “critical reason” for a crash, the last failure before it, to the driver in 94 per cent of cases, and said this was not meant to be read as the cause. NHTSA took the figure off its website in January 2022.⁴⁷ A machine that removed driver error would remove the crashes it handled better than people and add those it caused itself.

Regulators built their framework mostly from disclosure. The vocabulary came from SAE International’s standard J3016 of 2014, which divides automation into six levels, from 0 to 5. Its key line runs between Level 2, where a human must supervise, and Levels 3 and above, where the system drives within its conditions; critics argue that the numbering suggests a ladder each company would climb, when Google had jumped to Level 4 and Tesla sold Level 2 under a name suggesting more.⁴⁸ A voluntary federal policy of September 2016 set the division of labour that has lasted: the federal government regulates the vehicle, the states the driver, licensing and liability.⁴⁹ In June 2021 NHTSA ordered companies to report crashes involving automated driving and Level 2 systems.⁵⁰ No federal self-driving law has been passed; a bill introduced in February 2026 had not reached the House floor by November.⁵¹

Phoenix without a driver

Waymo did what RAND said could not be done quickly by narrowing the problem: mapped areas, chosen cities, good weather first. In November 2017 it began testing with nobody in the driver’s seat in part of metropolitan Phoenix, a region of wide streets and dry weather whose state government had invited it. A paid service with safety drivers opened in December 2018, and fully driverless rides opened to the public in October 2020.⁵² A decade after the ten routes, its driverless service covered part of one metropolitan area. The alternative was Tesla’s: a car that went everywhere and was finished nowhere.

San Francisco, 2023

San Francisco was where Waymo’s caution and Cruise’s speed met the same city. Cruise, backed by GM and by SoftBank, had been running driverless rides at night. In July 2023, before a vote on expanding them, a group called Safe Street Rebel found that a traffic cone on a car’s bonnet left it stranded with hazard lights flashing, and invited others to do it.⁵³ On 10 August 2023, over the objections of city officials, the California Public Utilities Commission voted three to one to let Waymo and Cruise charge for rides at all hours across the city.⁵⁴

Seven weeks later, on the night of 2 October, a woman crossing Market Street against the signal was hit by a human driver in a Nissan, who did not brake, and thrown into the path of a Cruise car. The car had tracked both her and the Nissan but had not predicted either would enter its lane, and hit her at about 19 mph. What followed ended the company. With the woman mostly hidden from its lidar, the car misjudged the impact as a side collision, which called for pulling over to the kerb rather than stopping. It moved again, reached 7.7 mph and dragged her about 20 feet. Engineers hired by Cruise found no hardware or software fault: the system did what it was designed to do.⁵⁵

Cruise’s leaders knew of the dragging by the next morning. In meetings with regulators they played their video of the crash without pointing it out. The lawyers Cruise later commissioned found that the company had planned to let “the video speak for itself,” that poor internet connections had probably kept officials from seeing it clearly in three meetings, and that leadership was “fixated on correcting the inaccurate media narrative” that Cruise had caused the crash. They found no intent to mislead, and called the outcome “a proverbial self-inflicted wound.”⁵⁶

On 24 October the California DMV suspended Cruise’s driverless permits, finding its vehicles unsafe and that it had misrepresented information about their safety. Cruise halted operations, recalled 950 vehicles, and its chief executive resigned.⁵⁷ It later paid $1.5 million to NHTSA and, under a deferred prosecution agreement, a $500,000 fine for a crash report that omitted the dragging.⁵⁸ In December 2024 GM stopped funding Cruise’s robotaxi, citing “the considerable time and resources that would be needed to scale the business,” and folded its technology into driver assistance. It had spent more than $10 billion on Cruise since 2016.⁵⁹

The retreat

Cruise was the last of a series. Ford and Volkswagen shut Argo AI in October 2022 after it failed to find outside investors; Ford wrote off $2.7 billion and said profitable autonomy at scale was “a long way off.”⁶⁰ Apple ended its car project in February 2024 and moved many of its roughly 2,000 staff to generative AI.⁶¹ In trucking, Embark sold itself for $71 million in 2023 and TuSimple left the American market.⁶² Investment in autonomous driving since 2010 has been put at more than $100 billion.⁶³

Argo, Apple and GM left because the finish line was further than they would pay to reach; Uber and Cruise lost their licence to operate after a single event and what they did around it. Aurora, which had absorbed Uber’s unit, began driverless commercial trucking between Dallas and Houston in May 2025; within three weeks it put an observer back in the driver’s seat at the lorry maker’s request. By September 2026 it had passed 500,000 driverless miles, on revenue of $3 million a quarter against operating losses of $257 million.⁶⁴

Scale

Waymo went the other way. It opened to everyone in San Francisco and Los Angeles in 2024, reached Austin and Atlanta through Uber’s app in 2025, took riders on freeways from November 2025, and in 2026 opened in ten more cities, from Miami to Las Vegas.⁶⁵ Its paid rides rose from about 50,000 a week in May 2024 to about 500,000 by March 2026, short of its target of a million by the end of the year. In February 2026 it raised $16 billion at a valuation of $126 billion; by September it had about 4,000 vehicles.⁶⁶

Its safety case rests on comparisons with human drivers in the same places. Through June 2026, over 271 million miles with nobody behind the wheel, Waymo reported 95 per cent fewer crashes causing serious injury or worse than a human benchmark, and 82 per cent fewer causing any injury.⁶⁷ A study with the insurer Swiss Re found 88 per cent fewer property-damage claims and 92 per cent fewer injury claims,⁶⁸ and a peer-reviewed paper by Waymo researchers in 2025 found significant reductions in injury crashes and no significant increase in any of eleven crash types.⁶⁹ The caveats are real. The studies are written or funded by the company; human benchmarks must be built from police reports that miss minor crashes; and deaths are too rare for even 271 million miles to say much about them. Waymo itself said that “more miles are needed” for the rarest outcomes.

Its failures have been public too, including software recalls after its cars hit a utility pole and, in December 2025, passed stopped school buses, nineteen times in Austin by the school district’s count.⁷⁰ In January 2026 a Waymo car hit a child who ran out from behind a double-parked SUV near a Santa Monica school, after braking from about 17 mph to under 6; the injuries were minor, and federal investigators opened an inquiry.⁷¹ In February 2026 the company told the Senate that some of the remote workers who advise its stuck cars are in the Philippines, about seventy agents on duty at any time for about 3,000 vehicles, giving guidance but not driving.⁷²

Tesla began a robotaxi service in Austin in June 2025 with an employee in the passenger seat. Rides with nobody on board opened in January 2026, at first in a few cars, and its Cybercab, without steering wheel or pedals, began carrying riders there in September. By the end of that month Tesla reported nearly two million unsupervised miles.⁷³ The same year, two legal judgements fell on the decisions of 2015 and 2016. California’s DMV adopted a judge’s finding that a reasonable consumer would believe a car with “Full Self-Driving Capability” could drive without constant attention, and that “this belief is wrong, both as a technological matter and as a legal matter”; Tesla dropped the name “Autopilot” there to avoid a suspension of sales, while contesting the order.⁷⁴ And a federal jury in Miami found Tesla 33 per cent responsible for a 2019 crash on Autopilot in Key Largo in which one person was killed and another badly injured, with a judgment of about $243 million. The driver, who had dropped his phone, settled separately. Tesla said no car then or now would have prevented the crash, and has appealed.⁷⁵

Beyond America

Other countries wrote law first. Germany’s law allowing driverless vehicles in approved areas, under remote “technical supervision,” took effect in July 2021, and a United Nations regulation of 2020 set the first binding international rule for Level 3 systems on motorways.⁷⁶ Britain’s Automated and Electric Vehicles Act 2018 made the insurer liable in the first instance for a crash caused by a car driving itself; its Automated Vehicles Act 2024 requires authorised vehicles to be at least as safe as “careful and competent human drivers” and moves responsibility for how a vehicle drives from the person inside to the company that authorised it. The full regime is not due until 2027, and London’s first paid rides, from September 2026, carried a safety driver.⁷⁷ Japan began its first driverless service in May 2023, on a two-kilometre path near a temple.⁷⁸

China built the largest market outside the United States. Baidu’s Apollo Go reported more than 23 million rides in 28 cities by mid-2026. When it expanded in Wuhan in 2024, taxi drivers petitioned the city; one ride-hail driver asked a Reuters reporter, “If you kill off this industry, what is left for them to do?”⁷⁹

The American Automobile Association found 68 per cent afraid of self-driving cars in 2023 and about six in ten in 2025. In 2026 Pew found that 5 per cent of adults had ridden in one and 71 per cent would be uncomfortable doing so; but among those who had ridden, 40 per cent would be very or extremely comfortable, against 5 per cent of those who had not.⁸⁰ Unions pressed for laws requiring a human in every commercial vehicle, and California’s governor vetoed two such bills.⁸¹

What self-driving cars are today

In November 2026 two different products are called self-driving cars. The first is the robotaxi: a car with nobody in the driver’s seat, run by a company within a mapped area and watched from a distance by people who can advise it but, the companies say, not drive it. Waymo runs about 4,000 in more than a dozen American cities; Baidu, Pony.ai and WeRide run similar services in China and the Gulf;⁸² Tesla runs a smaller one in Austin; and Zoox, owned by Amazon, announced paid rides in Las Vegas from August 2026 in a vehicle with no controls at all, under an exemption from standards written for cars with steering wheels.⁸³ In each case the company is the driver: it holds the permits, reports the crashes and answers to regulators.

The second is driver assistance that steers, brakes and accelerates while a person supervises: Tesla’s “Full Self-Driving (Supervised),” with about 1.74 million active subscriptions or purchases in the third quarter of 2026, and the systems GM and Ford turned to after abandoning their robotaxis.⁸⁴ Here the driver remains the driver in law, whatever the car is doing.

Neither can yet drive anywhere a person can. And human driving got safer in the same years for reasons unrelated to automation: American road deaths fell from nearly 43,000 in 2021 to an early estimate of 36,640 in 2025.⁸⁵ The benchmark the machines must beat keeps moving.

How self-driving cars changed driving

From the driver to the operator

For a century the law of the road assumed a person at the wheel who could be licensed, fined and blamed. A robotaxi has no such person. Responsibility that belonged to millions of drivers is moving to a few companies, through permits in California and Texas, crash reports to NHTSA, and Britain’s “authorised self-driving entity.” The driver is becoming an operator: an organisation judged on the record of a fleet.

From a skill to a supervision task

Driver assistance changed the nature of driving rather than removing it. The person in a car on Autopilot watches for the moment the machine fails, a task the investigations of the Brown, Huang and Herzberg crashes found people do badly. Much of the history since has been the search for ways to keep that monitor engaged: steering-wheel sensors, cameras, alerts, recalls.

From a car you own to a ride you buy

The robotaxi turns a product into a service, owned by a fleet that must keep it busy to pay for its sensors. Tesla’s subscription for supervised driving applies the same logic to cars people still own: the driving is rented month by month, and the software changes under them.

From a public road to a contested space

A car without a driver forces questions a car with one let people avoid: who the road is for, who decides what moves on it, and who answers for what happens there. Cities objected to state commissions licensing robotaxis over their heads; activists put cones on bonnets; unions sought laws requiring humans in every vehicle; cars were burned in Los Angeles.⁸⁶

From proof before sale to proof by miles

Safety used to be shown before a car went on sale, by testing it against standards written in advance. Self-driving cars cannot be fully tested that way, because their behaviour depends on situations nobody can list. The industry and its regulators moved instead to limited deployment, public crash reporting and statistical comparison with human drivers as the miles accumulate. The car is approved, in effect, by its record, and the record is kept mostly by the company.

What the history suggests

Some of this will be familiar from other products. Permissive defaults are borrowed time: Tesla let Autopilot work with weak checks on the driver’s attention, and the bill came due in investigations, a recall of two million cars, a ruling on its marketing and a jury verdict. Decide at the speed of reversibility: Waymo’s geofences and city-by-city expansion kept each step small enough to undo, while Uber’s removal of a second operator and Cruise’s handling of its crash were fast decisions in a domain where mistakes cannot be taken back. Real values are what you reward: Uber rewarded more test cars and miles, and the NTSB found its safety culture inadequate; Cruise’s leaders rewarded correcting the press narrative and lost the regulator. A defining strength becomes the constraint: Tesla’s ability to ship software to millions of cars at once made it fastest to market and put every flaw on the road at once. The metric becomes the target: miles driven and disengagements became the scoreboard, though neither measures safety in the situations that kill people.

The thing that sets self-driving cars apart is that the product takes over an act for which people have always been held personally responsible, and does so in public, among people who never agreed to be part of the trial. Every decision in this history, from Google’s refusal to sell a half-measure to Cruise’s video, was in the end a decision about two things that ordinary driving keeps bundled together and automation pulls apart: who is in control, and who answers for what happens. The philosophy of self-driving cars begins with that separation.

Understanding the philosophy

The thinkers below are linked to the history of self-driving cars by the author’s interpretation, not by anything they claimed. Most of these thinkers wrote before a car drove itself on a public road; the point is to see what their ideas show about the decisions in this history, not to claim they predicted them.

1. The supervisor who cannot supervise

The idea in plain terms

Partial automation turns a driver into a supervisor. The machine does the routine work, and the person’s job is to notice when it is about to fail and take over in time. The idea is that this job, which looks lighter than driving, is in some ways harder, and that people are poorly built for it.

The tradition

The research begins with radar. The Royal Air Force needed operators who could watch screens for long periods for signals that rarely came, and the psychologist Norman Mackworth simulated the task with a clock: subjects watched a pointer move round its face in small steps and reported the occasional double jump. Their performance fell off after about the first half hour. His paper of 1948 founded the study of the vigilance decrement, one of the most robust findings in applied psychology.⁸⁷

Lisanne Bainbridge drew the consequences for automation in 1983, in a paper discussed in this site’s essays on GPS and AI. Citing Mackworth, she wrote that it is “impossible for even a highly motivated human being to maintain effective visual attention towards a source of information on which very little happens, for more than about half an hour.” The automatic system is installed because it does the job better than the operator, and then the operator is asked to monitor it. Her final irony is the sharpest: “it is the most successful automated systems, with rare need for manual intervention, which may need the greatest investment in human operator training.”⁸⁸

Raja Parasuraman and colleagues found that such complacency appears in experts as well as novices and is not cured by practice.⁸⁹ In a simulator study of 2017, Alexander Eriksson and Neville Stanton found that drivers asked to take back control in routine conditions took between 1.9 and 25.7 seconds, and longer if they had been reading.⁹⁰ John Lee and Katrina See argued that the goal is calibrated trust, matched to what the system can actually do, and that calibration is hardest when failures are rare.⁹¹ The idea is discussed further on this site’s page Trusting the Machine Too Much.

Where it shows up

This history acts out Bainbridge’s paper. Google’s employees in 2012 did what Mackworth’s subjects did, and Google changed the design. Tesla built a product on the opposite assumption, and NHTSA’s phrase for the result, “a critical safety gap between drivers’ expectations of the L2 system’s operating capabilities and the system’s true capabilities,” is a regulator’s description of miscalibrated trust. Uber gave a single operator a system designed to leave emergency braking to her and no alert when it decided to hold back. The NTSB’s finding of “lack of adequate mechanisms for addressing operators’ automation complacency” could have been written by Bainbridge. Nicholas Carr had described the same “automation complacency” in The Glass Cage (2014).⁹² SAE’s Level 3, in which the driver need not supervise but must take over when asked, is a formal bet that a handover can be made safe with enough warning; Eriksson and Stanton’s range of 1.9 to 25.7 seconds measures how uncertain that bet is.

The strongest objection

Perhaps “complacency” is simply rational. Parasuraman and Manzey note that it is hard to call monitoring inadequate without a standard for how often a person ought to check a system that is usually right. A driver who relaxes when the system handles almost everything is responding to evidence. And partial automation may save more lives than it costs, by catching the fatigue and distraction that cause crashes too. A technology should be judged by its net effect, not its most vivid failures, and on this view the handover problem is a reason to improve monitoring, not to abandon partial automation.

What it means today

The question is empirical, and Tesla’s own comparisons, which set highway miles on Autopilot against all roads and all cars, have not settled it. What is settled is the mechanism: vigilance decays, and a system good enough to be trusted and not good enough to be trusted completely invites the most dangerous kind of trust. For anyone designing a product in which a person supervises automation, the rules follow from Bainbridge. Do not ask people to watch for rare failures unaided; check that they are actually watching; keep the human’s part of the task active rather than residual; and be honest, in the name of the product, about which of the two is in charge.

2. Dread and how safe is safe enough

The idea in plain terms

People do not judge risks by counting deaths. A risk that is imposed, unfamiliar and catastrophic frightens people more than a larger risk that is chosen, familiar and spread out. The idea is that this difference in perception, as much as any difference in danger, sets how safe the machines must be before people accept them.

The tradition

In 1969 the engineer Chauncey Starr inferred from past behaviour that society accepts voluntary risks far larger, by his estimate around a thousand times, than involuntary ones of comparable benefit; critics replied that what people have accepted is not what they find acceptable, and later work found the gap much smaller.⁹³ The psychologist Paul Slovic, with Baruch Fischhoff, Sarah Lichtenstein and others, studied perception directly. In “Perception of Risk” (1987) he showed that judgements of hazards cluster on two dimensions: “dread risk,” for hazards that feel uncontrollable, catastrophic and involuntary, and “unknown risk,” for the new and poorly understood. Hazards high on both have high “signal potential”: one accident is read as a sign of what is coming.⁹⁴

Gerd Gigerenzer argued that Americans who avoided flying after 11 September 2001 drove instead, and estimated with Wolfgang Gaissmaier about 1,600 extra road deaths in the following year, an estimate others dispute.⁹⁵ For self-driving cars, a survey of 499 people in Tianjin published in 2019 found they would tolerate the cars only if they were four to five times as safe as human drivers.⁹⁶ Azim Shariff, Jean-François Bonnefon and Iyad Rahwan, in three preregistered studies of 4,566 people published in 2021, traced such demands to algorithm aversion, the tendency documented by Berkeley Dietvorst and colleagues to lose faith in an algorithm faster than in a person after seeing it err, and to the better-than-average effect, by which most drivers think themselves safer than the average a machine is compared with. People would adopt the cars only if they met “extremely potentially unrealistically high” standards, and the biases proved “stubbornly hard to overcome.”⁹⁷

Where it shows up

Every feature of the dread profile applies. Nobody on the street in Tempe chose to share it with test vehicles; the risk is new and opaque; and a crash caused by software looks like a sign of what every car running it might do. Elaine Herzberg’s death was one of tens of thousands on American roads that year and halted a programme; one woman’s injuries in San Francisco ended a company. Slovic’s signal potential explains both. The polls fit too: American fear barely moved between 2023 and 2025, while those who had ridden in a driverless car were eight times as likely as those who had not to be very comfortable doing so.

The strongest objection

A higher bar for machines may be no bias at all. A fleet’s risk comes from a single design, so its failures are correlated: one flaw repeats in every car, as Waymo’s school-bus recall showed. The machines lack a century of track record, the evidence for their safety comes mostly from their makers, and the people who bear the risk, pedestrians above all, have not consented. On this view the psychologists are mislabelling prudence as prejudice.

What it means today

So the task is to separate the parts of public caution that track real differences from those that do not. Correlated failure, opacity and lack of consent are real, and they argue for independent data, fast recalls and public reporting. Algorithm aversion and the better-than-average effect are not, and no statistics will satisfy them. The practical lesson for anyone deploying automation in public is that a single vivid failure will be judged by its signal, not its frequency. A company that wants its record to count must make its conduct after a failure show that the failure was an exception, which is precisely what Cruise did not do.

3. Statistical lives and the enemy of good

The idea in plain terms

A policy that saves lives usually saves people nobody can name. A self-driving car that kills a pedestrian kills someone with a name and a face. The idea is that societies weigh these two kinds of life very differently, and that the argument over self-driving cars sets identified victims of machines against statistical lives the machines may save.

The tradition

The economist Thomas Schelling drew the distinction in 1968, in “The Life You Save May Be Your Own.” A child trapped in a well is an identified life, and a community will spend almost anything to save her; a life saved by a road improvement is statistical, and the same community will balk at a fraction of the cost. Schelling proposed valuing such lives by what people would pay for small reductions in their own risk, the reasoning behind the value of a statistical life.⁹⁸ Psychologists named the asymmetry the identified victim effect; Deborah Small and George Loewenstein showed in 2003 that even knowing a particular victim had been chosen, with no other details, increased giving. The effect has since shrunk under scrutiny: a meta-analysis in 2016 found it very small, and a reanalysis correcting for publication bias found little evidence for it.⁹⁹

The RAND reports supply the other side of the ledger. Kalra and Paddock calculated that demonstrating with confidence that a fleet was 20 per cent safer than human drivers would take more than 11 billion miles, about 518 years for a hundred cars, and that for a 5 per cent improvement the figure became “almost ludicrous: 215 billion miles.” “Uncertainty will persist,” they concluded.¹⁰⁰ Kalra and Groves then found that under none of the conditions they explored did waiting for large safety gains lead to fewer deaths.¹⁰¹

Where it shows up

The debate has been conducted almost entirely in Schelling’s two currencies. On one side are about 36,640 people killed on American roads in 2025, almost none known to the public, and RAND’s projection of lives saved by deploying early. On the other are Joshua Brown, Walter Huang, Elaine Herzberg and the woman dragged in San Francisco, each named, each investigated, each a reason to slow down. The industry’s case has relied on statistical lives; its setbacks have come from identified ones. Waymo’s published comparisons are an attempt to give the unnamed people who were not hurt a presence in the argument.

The strongest objection

Treating lives as interchangeable hides who pays. The calculation counts a pedestrian killed by a robotaxi and a driver saved from a drunk human as one life each, and nets them out. But Elaine Herzberg gained nothing from Uber’s test programme. A risk accepted by a rider differs from one imposed on a stranger. And RAND’s model assumes that cars deployed early will improve with use, which requires their makers to learn from failures; Cruise’s record shows that is not automatic.

What it means today

The identified victim effect turned out weaker in the laboratory than its reputation, but its political version is as strong as ever: one named death can halt a programme the statistics say is saving lives. They can make the statistical lives real through independent data rather than marketing, and make the distribution of risk defensible, so that the people exposed during testing are protected by more than averages: second operators, conservative operating areas, fast responses to failure. The enemy-of-good arithmetic holds only if the good actually improves.

4. The trolley that never came

The idea in plain terms

The most famous puzzle about self-driving cars asks what a car should do when a crash is unavoidable and it must choose whom to hit. The idea here is that the puzzle is real but mostly the wrong one: the choices that matter are made long before any crash, in mundane decisions about speed, braking, classification and how much risk to accept in exchange for fewer false alarms.

The tradition

Philippa Foot introduced the case in 1967, in a paper on the doctrine of double effect: the driver of a runaway tram can steer onto a track where one man is working instead of five. Judith Jarvis Thomson named it “the trolley problem” in 1976 and, in 1985, contrasted switching the points with pushing a large man off a footbridge to stop the trolley.¹⁰² In 2016 Bonnefon, Shariff and Rahwan brought it to cars: people approved of cars that would sacrifice their passengers to save more pedestrians, and wanted others to buy them, but preferred to ride in cars that protected them.¹⁰³ Their Moral Machine website then gathered about 40 million decisions from more than two million people in 233 countries and territories; people preferred to spare humans over animals, more lives over fewer, and the young over the old.¹⁰⁴ Germany’s Ethics Commission on Automated and Connected Driving, chaired by the former constitutional judge Udo Di Fabio, issued twenty rules in 2017; among them, any distinction between people by personal features such as age or gender in an unavoidable accident was impermissible, though general programming to reduce injuries might be justifiable.¹⁰⁵

Sven Nyholm and Jilles Smids argued in 2016 that a car’s choice is made in advance by many people, not in a moment by one; that responsibility matters in real crashes, though trolley cases stipulate it away; and that real outcomes are uncertain.¹⁰⁶ Johannes Himmelreich argued in 2018 that the important ethics lies in mundane driving, such as how much room to give a cyclist.¹⁰⁷ Yochanan Bigman and Kurt Gray showed in 2020 that when people could choose to treat lives equally, most did, so that many of the Moral Machine’s preferences came from forcing a choice.¹⁰⁸ Julian De Freitas and colleagues proposed in 2021 replacing dilemmas with tests of common sense on the road.¹⁰⁹

Where it shows up

No crash in this history was a trolley problem. The Uber car in Tempe did not choose between Elaine Herzberg and someone else. It failed to classify her, because its designers had not considered people crossing outside crosswalks, and held back from braking because its designers had judged false alarms a greater risk than late braking. Those were ethical decisions about whose safety to trade for whose convenience, made months earlier. The Cruise car’s designers had decided what a car should do after a side impact and had not imagined a person beneath it. The moral content lay where Nyholm, Smids and Himmelreich said it would. Brian Christian used Tempe in The Alignment Problem (2020) as an example of a system failing because of the categories it was given.¹¹⁰

The strongest objection

A defender of the trolley problem would answer that mundane cases are trolley problems in slow motion. Deciding how much room to give a cyclist distributes risk between the cyclist and the passengers. Tuning a braking threshold trades false stops, which cause rear-end crashes, against late stops, which kill pedestrians. Every such choice allocates expected harm, and the trolley problem, by isolating the allocation, is exactly the tool for thinking about it.

What it means today

That rescues the trolley problem’s principle and abandons its picture. What matters is how risk is distributed across thousands of ordinary decisions, not whom a car should hit in a dilemma. That moves ethical attention from the moment of the crash to the engineering process: which situations the system was designed to consider, which thresholds were set and why, and who reviewed the trade-offs. The lesson extends beyond cars. The ethics of an automated system lives in its categories and defaults, and teams should be able to say whom each threshold protects and whom it exposes, and write it down, because a regulator, a court or a family will eventually ask.

5. The gap and the crumple zone

The idea in plain terms

When a person drives badly, we know whom to blame. When a machine that learned from data drives badly, it can be hard to find anyone who could have foreseen the specific failure. The idea of a responsibility gap is that automated systems can cause harms for which no one is fully responsible. The idea of a moral crumple zone is what tends to fill the gap: the nearest human, who had little control, absorbs the blame.

The tradition

The philosopher Andreas Matthias argued in 2004 that machines that learn and act on their own produce behaviour their makers and users could not predict and so cannot fairly be blamed for; society must either forgo such machines or accept a gap that ordinary ways of assigning responsibility cannot bridge.¹¹¹ Robert Sparrow turned the argument against autonomous weapons in 2007.¹¹² The anthropologist Madeleine Clare Elish named the other half of the problem in 2019. A crumple zone is the part of a car built to absorb the force of a crash and protect the people inside; a moral crumple zone describes “how responsibility for an action may be misattributed to a human actor who had limited control over the behavior of an automated or autonomous system.” Where the physical one protects the driver, “the moral crumple zone protects the integrity of the technological system, at the expense of the nearest human operator.” Her examples included Three Mile Island, Air France 447 and Tempe.¹¹³

Filippo Santoni de Sio and Jeroen van den Hoven proposed in 2018 a standard of “meaningful human control” with two conditions: the system should respond to the moral reasons of the people it serves (tracking), and every outcome should be traceable to at least one person who understood the system’s capabilities and their own role (tracing).¹¹⁴ Santoni de Sio and Giulio Mecacci later distinguished four gaps, in culpability, moral accountability, public accountability and active responsibility, and argued that organisations create them as much as learning machines do.¹¹⁵

Where it shows up

Tempe is the clearest crumple zone in this history. The design left emergency braking to the operator, a second operator had been removed, nothing caught her inattention, and the NTSB found the company’s safety culture inadequate. She was the only person charged. That does not mean she bore no responsibility; she was watching television. It means that the law, following the probable cause, found the human at the end of the chain easier to reach than the decisions that put her there. Tesla placed its drivers in a similar position: the driver was described as responsible at all times, while the NTSB, NHTSA and a Miami jury each found that the design had contributed to crashes. The jury’s 33 per cent was an attempt to divide what the product’s framing had assigned to one party.

The robotaxi closes the gap from the other side: with nobody in the driver’s seat, responsibility goes to the company, as Britain’s 2024 Act makes explicit.

The strongest objection

Some philosophers deny that there is any gap. Daniel Tigard argued in 2021 that technology creates no gaps our varied practices of responsibility cannot cover, and Peter Königs asked in 2022 why, even if gaps exist, they should be thought seriously troubling.¹¹⁶ Companies are already liable for their products, regulators can recall them, prosecutors can charge organisations. That no single engineer could foresee a failure is true of any complex product, and corporate accountability fills the space.

What it means today

That describes what can happen. This history shows what usually does. Corporate accountability arrived, but slowly and mostly in money, while criminal charges and public blame fell on individuals at the edge of the system. Santoni de Sio and Mecacci’s point that the gaps are organisational means they can be designed out. The test of tracing is practical: for every way the system can fail, can someone say who decided the risk was acceptable, and did that person understand the decision? If the honest answer points only to the person in the driver’s seat, the organisation has built a crumple zone.

6. Moral luck and the fleet

The idea in plain terms

We judge people partly by outcomes they do not control. Two drivers glance at their phones for the same two seconds; one looks up to an empty road, the other hits a child. We blame the second far more, though they did the same thing. The idea of moral luck is that this is not a mistake we can simply correct: our practices of praise and blame are shot through with luck.

The tradition

Bernard Williams and Thomas Nagel introduced the problem in two papers titled “Moral Luck,” given at a symposium of the Aristotelian Society in 1976. Williams described a lorry driver who, through no fault of his own, runs over a child, and who feels what Williams called “agent-regret,” a sense of involvement an onlooker does not share, though no one should blame him. Nagel distinguished resultant luck, in how things turn out, from circumstantial luck, in the situations one faces, and constitutive luck, in the kind of person one is. If the lorry driver had been even slightly negligent, failing to have his brakes checked, the outcome would transform how he is judged, though the negligence would have been the same had no child run into the road.¹¹⁷ The problem is discussed on this site’s page Moral Luck and Outcome Bias, and it is why this site’s decisions boxes judge a decision’s process apart from its outcome.

Where it shows up

A fleet turns individual luck into statistics. A design choice in a robotaxi produces a rate of crashes across millions of miles, and which crash happens, where and to whom, is luck. Waymo’s car in Santa Monica braked from about 17 mph to under 6 before hitting a child who ran out from behind a parked SUV; at a slightly higher speed the same design could have produced a death and a different history for the company. The Cruise crash began with a human driver who threw a pedestrian into the robotaxi’s path: circumstantial luck, in Nagel’s terms, of the worst kind. Regulators and the public judge by outcomes, as Nagel would predict.

The strongest objection

Many philosophers would remove resultant luck from judgement altogether. On the view sometimes called the control principle, people and organisations should be judged only by what they controlled: their process, their care, their design. A company whose car hit a child at 6 mph is no better than one whose car hit at 30, if their systems were equally good. RAND’s arithmetic shows that fatal crashes are too rare for outcomes to separate good systems from bad ones for hundreds of millions of miles, so judging fleets by outcomes teaches the wrong lesson. Process is the only fair standard.

What it means today

Fair enough, as far as blame goes. But outcomes carry information about process when they are examined closely, as the NTSB examined Tempe and Cruise’s lawyers examined its response. The Cruise case is instructive because the luck was bad and the conduct worse: the crash was largely a human driver’s fault, and what ended the company was what it did afterwards, which was not luck at all. Leaders should judge their own systems by design and rates, as the control principle asks, and expect the world to judge them by their worst day. What they control on that day is the candour of the response.

7. Trust and the licence to operate

The idea in plain terms

A self-driving car cannot be proved safe before it is deployed, so it must be trusted: by riders, by people on the street and above all by the regulators who grant permission to operate. The idea is that trust is a relationship that depends on the trusted party’s conduct, especially its honesty about failures, and that losing it can end a business faster than any crash.

The tradition

Annette Baier defined trust in 1986 as “accepted vulnerability to another’s possible but not expected ill will (or lack of good will) toward one.” To trust is to leave others an opportunity to harm us, confident they will not take it.¹¹⁸ Onora O’Neill argued in her Reith Lectures of 2002, A Question of Trust, that elaborate audit may damage trust rather than restore it, and that what matters is that institutions be trustworthy and that people be able to place trust well.¹¹⁹ David Collingridge described in 1980 the dilemma of governing a new technology: at first its effects cannot be foreseen, and by the time they are clear it is entrenched. His answer was to keep technologies correctable.¹²⁰

Where it shows up

The regulators in this history granted trust on Collingridge’s terms: permits limited by area and hours, revocable at will. The suspension of Cruise was his escape hatch working. What made it necessary was a failure of the kind Baier describes. Regulators had accepted vulnerability to Cruise’s good faith, depending on its reports to know what its cars had done. When the company let a video “speak for itself” instead of saying its car had dragged a woman, the damage was to the relationship, and the DMV cited misrepresentation alongside safety. Waymo published its safety data and, when asked by the Senate, said where its remote advisers worked, which did not please everyone but gave regulators something to check. Tesla’s naming is a quieter case. A buyer who trusts “Full Self-Driving” to mean what it says accepts a vulnerability the product cannot honour, and California’s ruling that a “reasonable consumer” would be misled is, in Baier’s terms, a finding that the company invited trust it could not bear.

The strongest objection

Talk of trusting machines may be a category mistake. Baier’s trust is a relation between persons, who can have good or ill will; a car has neither. We rely on machines as on bridges, and reliance is a matter of evidence. The right response is to demand proof: independent testing, standards, certification. Staged deployment, Collingridge’s critics add, gathers knowledge before entrenchment, so trust need hardly enter.

What it means today

Pressed far enough, both points lead back to trust. Because proof by miles takes so long, the evidence on which reliance depends comes for years from the companies themselves, so reliance on the car becomes trust in the company that reports on it. Staged deployment gathers knowledge only if the reports are honest. For anyone whose product needs a regulator’s permission, the asset is not the permission but the regulator’s confidence in what it is told. Report bad news fully, early and in words, and do not invite more trust in a product’s name than the product can bear.

8. Whom the street is for

The idea in plain terms

Roads look like neutral infrastructure, but the rules of who may use them, and how, were made by people and could have been made otherwise. The idea is that self-driving cars are a new claim on the street, and that the conflicts over them, between companies and cities, drivers and machines, pedestrians and fleets, are the latest round of an old argument about whom public space is for.

The tradition

The historian Peter Norton showed in Fighting Traffic (2008) that American streets in the early twentieth century belonged to pedestrians, children and streetcars as much as to motorists, and that the motor age required a change in social rules as well as roads. By about 1930 streets had been redefined as thoroughfares for cars, and a pedestrian crossing outside the designated places had become a “jaywalker.”¹²¹

The political theorist Langdon Winner asked in 1980, in an essay reprinted in The Whale and the Reactor (1986), “Do Artifacts Have Politics?” His best-known example, drawn from Robert Caro’s The Power Broker (1974), was the low overpasses of the Long Island parkways, said to have been built so that buses, and the poorer and Black New Yorkers who rode them, could not reach the beaches. A design, Winner argued, can settle a political question so quietly that nobody remembers it was asked. The sociologist Bernward Joerges disputed the example in 1999, arguing that it had become a parable whose authority exceeded its evidence.¹²² Ivan Illich, in Energy and Equity (1974), reckoned that the typical American spent about 1,600 hours a year on the car, counting the hours worked to pay for it, to travel 7,500 miles, a speed of less than five miles an hour; later recalculations give much lower totals.¹²³

Where it shows up

The most precise sentence in this history, for Norton’s argument, is the NTSB’s: “the system design did not include consideration for jaywalking pedestrians.” A category invented in the 1920s to give cars priority over people was written, by omission, into software a century later, and a woman crossing where the category did not expect her was not recognised as a pedestrian. That is Winner’s point in its strongest form: the politics of the street were built into an artefact without anyone deciding to put them there.

The fights over robotaxis are fights about who decides. San Francisco’s officials opposed the expansion of 2023 and were overruled by a state commission. Activists answered with cones, unions with bills, taxi drivers in Wuhan with petitions. Britain and Germany wrote laws before allowing driverless operation; the United States regulates by permit, exemption and crash report. Each is an answer to Winner’s question about who settles what the street is for.

The strongest objection

Joerges’s complaint is the obvious one: the stories are better than the evidence. The bridges may not have been built for the reason Winner said, and the politics read into artefacts are often the critic’s. A self-driving car follows the rules the street already has, and if those rules are unfair, the remedy is to change them, not to blame the machine. Robotaxis may even improve the street for people outside cars: they do not drink or tire, and if the safety data is right they injure pedestrians far less often than human drivers do.

What it means today

Joerges may be right about the bridges. He has nothing to say about the jaywalker. Artefacts need not be built with political intent to carry political consequences; it is enough that their designers inherit categories and defaults without examining them. Companies building automated systems for shared space should look for the assumptions they have inherited about who belongs where before those assumptions appear in a crash report, and should expect their technology to be argued over as a claim on that space. The cities and residents who never chose to join the trial are not obstacles to be managed. It is their street.

Bringing the threads together

Mackworth and Bainbridge explain why partial automation produces a supervisor who cannot supervise; Slovic, Schelling and the RAND researchers why the public demands more of machines than of people, and what that demand costs. Foot’s critics move the ethics from the crash to the design before it. Elish and Nagel show where blame lands when control is divided, Baier why a licence to operate depends on candour, and Norton and Winner that the machines inherited a street already settled among people.

Then there are the decisions themselves. Google chose not to build the supervisor’s job; Tesla built it and named it Autopilot; Uber removed the second supervisor; Cruise chose how to tell regulators about its worst night; GM chose to stop; Waymo chose to go slowly and publish. Each was a decision about where control would sit and who would answer for it.

The empty seat

For a century the driver’s seat held two things at once: the control of the car and the responsibility for what it did. The person who steered was the person who answered. Self-driving cars separate them. Control can go to software while responsibility stays with a person who is no longer really driving, which is the arrangement that failed in Florida, Mountain View and Tempe; or control and responsibility can move together to a company, which is what the robotaxi does and what Britain’s law now says. The empty seat is the visible sign of that second arrangement. Most of the trouble in this history came from the first.

A few working principles come out of all this.

Do not build a job that people cannot do. A design that depends on someone watching for rare failures for long periods depends on something seventy years of research says people do not reliably provide. Keep the person engaged, or take the job away entirely, as Google did.

Say who is in charge, in the name as well as the manual. If a name promises more than the product does, users’ trust fills the gap, and the cost falls on them and on the people around them.

Treat thresholds as moral choices. The decisions that killed in Tempe were made long before the crash, in settings for false alarms, classification and braking. Write down whom each one protects and whom it exposes, and have someone who understands the trade-off sign it.

Design the launch to produce proof. Where safety cannot be demonstrated before deployment, deploy in steps small enough to reverse, publish what happens, and let others check it.

Spend trust on bad news first. A licence to operate rests on a regulator’s confidence in what it is told. A company that tells the worst part of a story early, plainly and in words keeps that confidence; one that lets the evidence speak for itself may find it has spoken against them.

The cars that could not cross the Mojave in 2004 now carry half a million paying passengers a week, and the question they raise is no longer whether a machine can drive. It is whether the people who build the machine will take the seat they have emptied: answer for what their cars do, in the open, as drivers always had to.

Selected bibliography

• Foot, Philippa. “The Problem of Abortion and the Doctrine of the Double Effect.” Oxford Review 5 (1967): 5–15.
• Schelling, Thomas C. “The Life You Save May Be Your Own.” In Problems in Public Expenditure Analysis, edited by Samuel B. Chase Jr., 127–162. Washington, DC: Brookings Institution, 1968.
• Illich, Ivan. Energy and Equity. London: Calder & Boyars, 1974.
• Caro, Robert A. The Power Broker: Robert Moses and the Fall of New York. New York: Knopf, 1974.
• Williams, Bernard, and Thomas Nagel. “Moral Luck.” Proceedings of the Aristotelian Society, Supplementary Volume 50 (1976): 115–135 and 137–151.
• Winner, Langdon. “Do Artifacts Have Politics?” Daedalus 109, no. 1 (1980): 121–136; reprinted in The Whale and the Reactor. Chicago: University of Chicago Press, 1986.
• Bainbridge, Lisanne. “Ironies of Automation.” Automatica 19, no. 6 (1983): 775–779.
• Baier, Annette. “Trust and Antitrust.” Ethics 96, no. 2 (1986): 231–260.
• Slovic, Paul. “Perception of Risk.” Science 236, no. 4799 (1987): 280–285.
• Parasuraman, Raja, and Victor Riley. “Humans and Automation: Use, Misuse, Disuse, Abuse.” Human Factors 39, no. 2 (1997): 230–253.
• Thrun, Sebastian, et al. “Stanley: The Robot That Won the DARPA Grand Challenge.” Journal of Field Robotics 23, no. 9 (2006): 661–692.
• Norton, Peter D. Fighting Traffic: The Dawn of the Motor Age in the American City. Cambridge, MA: MIT Press, 2008.
• Carr, Nicholas. The Glass Cage: Automation and Us. New York: W. W. Norton, 2014.
• Bonnefon, Jean-François, Azim Shariff, and Iyad Rahwan. “The Social Dilemma of Autonomous Vehicles.” Science 352, no. 6293 (2016): 1573–1576.
• Kalra, Nidhi, and Susan M. Paddock. “Driving to Safety.” Transportation Research Part A 94 (2016): 182–193.
• Nyholm, Sven, and Jilles Smids. “The Ethics of Accident-Algorithms for Self-Driving Cars: An Applied Trolley Problem?” Ethical Theory and Moral Practice 19, no. 5 (2016): 1275–1289.
• Awad, Edmond, et al. “The Moral Machine Experiment.” Nature 563 (2018): 59–64.
• Elish, Madeleine Clare. “Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction.” Engaging Science, Technology, and Society 5 (2019): 40–60.
• National Transportation Safety Board. Collision Between Vehicle Controlled by Developmental Automated Driving System and Pedestrian, Tempe, Arizona, March 18, 2018. HAR-19/03, 2019.
• Christian, Brian. The Alignment Problem. New York: W. W. Norton, 2020.

Research limitations

Several early episodes rest on secondary sources, among them the 1925 Houdina run, Futurama and the Tsukuba car, where accounts differ on whether it used guide rails; figures for the 2004 race and Dickmanns’s 1995 drive differ between reference works, and the essay uses the Stanley paper and Dickmanns’s own account. Tesla’s announcements of 2014 to 2016 could not be retrieved from its website and are quoted from contemporaneous press; its first Autopilot dates are given only by year, and the April 2019 robotaxi remark is open to more than one reading. Mobileye’s and Tesla’s accounts of their split are participant claims. Google’s 2012 test is known only from Waymo’s account, which conflicts with its later chief executive’s date of 2013. The prices paid for Otto and Cruise were reported, not disclosed, and the industry investment total is second-hand. Waymo’s safety figures, including the Swiss Re and peer-reviewed studies, are produced or funded by Waymo; Tesla’s safety statistics have been criticised for their comparison base, and no independent analysis of either was found. NHTSA’s withdrawal of the 94 per cent figure rests on one secondary source. At the time of writing, Tesla’s appeal in Miami and its challenge to California’s ruling were pending, several NHTSA investigations were open, the federal SELF DRIVE Act had not passed, and Waymo’s target of a million weekly rides had not been reported met. Gigerenzer’s estimate of extra road deaths after September 2001, the identified victim effect and Winner’s Moses bridges are contested, and the German Ethics Commission’s rule 9 is paraphrased from secondary quotation. The connections to Mackworth, Bainbridge, Parasuraman, Manzey, Eriksson, Stanton, Lee, See, Carr, Starr, Slovic, Gigerenzer, Gaissmaier, Liu, Shariff, Bonnefon, Rahwan, Dietvorst, Schelling, Small, Loewenstein, Kalra, Paddock, Groves, Foot, Thomson, Awad, Di Fabio, Nyholm, Smids, Himmelreich, Bigman, Gray, De Freitas, Christian, Matthias, Sparrow, Elish, Santoni de Sio, van den Hoven, Mecacci, Tigard, Königs, Williams, Nagel, Baier, O’Neill, Collingridge, Norton, Winner, Caro, Joerges and Illich are the author’s interpretation; their works are paraphrased apart from short, attributed phrases. The decisions in the decisions box are judged by the author, separating what could be known at the time (process) from what happened (call).

Footnotes

1. Sebastian Thrun et al., “Stanley: The Robot That Won the DARPA Grand Challenge,” Journal of Field Robotics 23, no. 9 (2006): 661–692, on the 2004 race; IEEE Spectrum, “Dusted”. ↩

2. Waymo, “Safety Impact”, data through June 2026; Waymo, “Waymo raises $16 billion”, 2 February 2026. ↩

3. NHTSA, early estimate of motor vehicle traffic fatalities in 2025, April 2026; World Health Organization, Global Status Report on Road Safety 2023. ↩

4. “Houdina Radio Control,” Wikipedia, citing Time, 10 August 1925, and the Harrisburg Evening News, 28 July 1925. ↩

5. “Futurama (New York World’s Fair),” Wikipedia; Harry Ransom Center, Norman Bel Geddes collection. ↩

6. Evan Ackerman, “Self-Driving Cars Were Just Around the Corner, in 1960,” IEEE Spectrum, 31 August 2016; The Henry Ford, RCA brochure “Electronic Control of Motor Vehicles on the Highway”. ↩ ↩²

7. Ernst D. Dickmanns, “The Development of Machine Vision for Road Vehicles in the Last Decade,” keynote, IEEE Intelligent Vehicles Symposium, 2002, citing S. Tsugawa, T. Hirose and T. Yatabe, “An Automobile with Artificial Intelligence,” Proceedings of the 6th IJCAI (Tokyo, 1979): 893–895; “History of self-driving cars,” Wikipedia. ↩

8. Dickmanns, “The Development of Machine Vision for Road Vehicles in the Last Decade”, 2002. ↩ ↩²

9. Dean A. Pomerleau, “ALVINN: An Autonomous Land Vehicle in a Neural Network,” in Advances in Neural Information Processing Systems 1 (1989). ↩

10. Carnegie Mellon University, “No Hands Across America,” official press release, 1995; Carnegie Mellon School of Computer Science, “About SCS”. ↩

11. University of Parma, ARGO project flyer; “Professor reads newspaper in ton-up robocar on autostrada,” Times Higher Education. ↩

12. “Overview of Automated Highway Demonstration ‘97”, University of Washington; California PATH, Demo ‘97 retrospective. ↩

13. Federal Highway Administration, FHWA Transporter, March 1998. ↩

14. Floyd D. Spence National Defense Authorization Act for Fiscal Year 2001, Public Law 106-398, section 220. ↩

15. Thrun et al., “Stanley,” 2006. ↩ ↩²

16. “Champion Robot Car Declared,” MIT Technology Review, 5 November 2007; Carnegie Mellon NREC, “Urban Challenge”. ↩

17. “Velodyne,” Lidar Magazine; Velodyne Lidar, company history. ↩

18. Waymo, “In the driver’s seat: 1000 mile challenge,” April 2020; Dmitri Dolgov interviewed by TechCrunch, 8 February 2019. ↩

19. John Markoff, “Google Cars Drive Themselves, in Traffic,” New York Times, October 2010, reprinted by KurzweilAI; eWeek, October 2010, quoting Thrun’s Google blog post of 9 October 2010. ↩

20. Waymo, On the Road to Fully Self-Driving: Waymo Safety Report, October 2017, section “The Case for Full Autonomy,” PDF copy. ↩

21. John Krafcik at the IAA, Frankfurt, September 2019, as reported by Jalopnik, 12 September 2019. ↩

22. Cars.com, May 2014; TechRadar, May 2014. ↩

23. Waymo, “From post-it note to prototype”, June 2017; Waymo, “Expanding Waymo’s testing to Austin”, March 2023. ↩

24. On Mobileye’s EyeQ3 chip, Autoblog, 26 July 2016; on the sensors, InsideEVs and The Drive. Tesla’s own announcements of 2014 and 2015 could not be retrieved. ↩

25. National Transportation Safety Board, investigation HWY16FH018, report HAR-17-02, adopted 12 September 2017; on the warnings, Insurance Journal, 21 June 2017. ↩

26. Fortune, 26 July 2016; Shashua to Reuters, reported by Globes and Electrek, 15 September 2016. ↩

27. Tesla’s announcement of 19 October 2016 as quoted by Electrek and CleanTechnica. ↩

28. Inside Higher Ed, 1 June 2015; Reuters via Yahoo. ↩

29. FreightWaves; Berkeley Law, The Network. The price was reported, not disclosed. ↩

30. “Cruise (autonomous vehicle),” Wikipedia, on the purchase and the SoftBank investment; GM, Form 8-K, exhibit 99.1, 10 December 2024, on GM’s ownership. ↩

31. CapRadio/NPR, 21 December 2016; Engadget, 21 December 2016. ↩

32. Smart Cities Dive; Phoenix New Times. ↩

33. NPR, 9 February 2018; VentureBeat, 9 February 2018. ↩

34. National Transportation Safety Board, Collision Between Vehicle Controlled by Developmental Automated Driving System and Pedestrian, Tempe, Arizona, March 18, 2018, HAR-19/03, adopted 19 November 2019. ↩ ↩²

35. AP via News4Jax, 28 July 2023; Engadget. ↩

36. Uber Technologies, Form 10-K for 2020; TechCrunch, 7 December 2020. ↩

37. National Transportation Safety Board, investigation HWY18FH011, report HAR-20-01, adopted 25 February 2020; on the speed, ISHN. ↩

38. NHTSA Office of Defects Investigation, PE21-020 opening resume, August 2021, and EA22-002 closing resume, 25 April 2024. ↩

39. Tesla, Part 573 safety recall report 23V-838, 12 December 2023; NHTSA, RQ24-009 opening resume, 25 April 2024. ↩

40. CNBC, 19 February 2019; Fortune, 22 April 2019. ↩

41. Engadget, 20 October 2020; TeslaNorth, 21 October 2020. ↩

42. The Drive; TechSpot, October 2022. ↩

43. Not a Tesla App; InsideEVs. ↩

44. Tesla, Q3 2026 Update, SEC Form 8-K exhibit 99.1, October 2026. ↩

45. Nidhi Kalra and Susan M. Paddock, Driving to Safety: How Many Miles of Driving Would It Take to Demonstrate Autonomous Vehicle Reliability? RAND, RR-1478-RC, 2016; RAND news release via EurekAlert. ↩

46. RAND, “Why Waiting for Perfect Autonomous Vehicles May Cost Lives,” November 2017, summarising Kalra and Groves, The Enemy of Good; on the report’s caveat, The Drive. ↩

47. NHTSA, “Critical Reasons for Crashes Investigated in the National Motor Vehicle Crash Causation Survey,” DOT HS 812 115, February 2015; on its withdrawal, Nelson Niehaus. ↩

48. SAE International, J3016, September 2016 revision; SAE, “SAE revises levels of driving automation”, June 2021. ↩

49. US Department of Transportation, “U.S. DOT issues Federal Policy for safe testing and deployment of automated vehicles,” 20 September 2016. ↩

50. NHTSA, report to Congress on vehicles equipped with automated driving systems; on the 2025 amendments, Mayer Brown, 28 April 2025. ↩

51. Congress.gov, H.R. 7390; Sidley, January 2026. ↩

52. Waymo, “Waymo’s fully self-driving vehicles are here”, 7 November 2017; SiliconANGLE, 5 December 2018. ↩

53. SF Standard, 7 July 2023; NPR, 26 August 2023. ↩

54. Axios, 11 August 2023; SF Standard, 10 August 2023. ↩

55. Exponent, technical root-cause analysis of 12 December 2023, reproduced in Cruise’s filing to the California Public Utilities Commission, 25 January 2024. ↩

56. Quinn Emanuel Urquhart & Sullivan, report to the boards of Cruise and GM, in the same CPUC filing; CNBC, 25 January 2024. ↩

57. California DMV, “DMV statement on Cruise LLC suspension”, 24 October 2023; NBC News, November 2023. ↩

58. SFist, 1 October 2024; TechCrunch, 15 November 2024. ↩

59. GM, Form 8-K, exhibit 99.1, 10 December 2024; on total spending, Fortune, 10 December 2024. ↩

60. TechCrunch, 26 October 2022; Fortune, 27 October 2022. ↩

61. Bloomberg, 27 February 2024; Macworld. ↩

62. FreightWaves on Embark; FreightWaves on TuSimple. ↩

63. The figure attributed to McKinsey by the Financial Times, as reported by Global Venturing; McKinsey’s own chart uses a broader category. ↩

64. Axios, 19 May 2025; Aurora Innovation, Q3 2026 shareholder letter, 28 October 2026. ↩

65. Waymo, Help Center announcements, 2024 to 2026; TechCrunch, 12 November 2025. ↩

66. TechCrunch, 27 March 2026; Sherwood News. ↩

67. Waymo, “Safety Impact”, data through June 2026. ↩

68. Waymo, “New Swiss Re study”, December 2024; Electrek, 19 December 2024. ↩

69. Kristofer D. Kusano et al., “Comparison of Waymo Rider-Only Crash Rates by Crash Type to Human Benchmarks at 56.7 Million Miles,” Traffic Injury Prevention 26, sup. 1 (2025); PubMed 40378124. ↩

70. TechCrunch, 12 June 2024; Electrive, 15 May 2025. ↩

71. NHTSA, PE26-001 opening resume, January 2026; TechCrunch, 29 January 2026. ↩

72. Axios, 31 March 2026; SFist, 6 February 2026. ↩

73. Claims Journal/Bloomberg, 23 January 2026; InsideEVs. ↩

74. CNBC, 16 December 2025; TechCrunch, 16 December 2025. ↩

75. Benavides v. Tesla, Inc., No. 21-cv-21940 (S.D. Fla.), as reported by CarComplaints, August 2025, and CarComplaints, 22 February 2026. ↩

76. German Federal Ministry of Transport, “Automated and connected driving”; Hogan Lovells. ↩

77. Automated and Electric Vehicles Act 2018, s. 2; Automated Vehicles Act 2024. ↩

78. Ministry of Economy, Trade and Industry, press releases of 31 March, 12 May and 22 May 2023. ↩

79. Baidu’s second-quarter 2026 results as reported by Venture Atlas; Pony.ai, Form 6-K, exhibit 99.1, 2026. ↩

80. AAA, “Fear in self-driving vehicles persists”, February 2025; The Hill, 2023. ↩

81. TT News on AB 316 (2023); CCJ on AB 2286 (2024). ↩

82. TechCrunch, 25 November 2025; Automotive World, 2026. ↩

83. TechCrunch, 10 September 2025; Las Vegas Sun, 6 August 2026. ↩

84. Tesla, Q3 2026 Update; the figure includes upfront purchases and excludes trials. ↩

85. NHTSA, early estimate of motor vehicle traffic fatalities in 2025, released April 2026; NHTSA Crash Stats 813762 on 2021. ↩

86. TechCrunch, 9 June 2025. ↩

87. N. H. Mackworth, “The Breakdown of Vigilance during Prolonged Visual Search,” Quarterly Journal of Experimental Psychology 1, no. 1 (1948): 6–21, record; “Vigilance (psychology),” Wikipedia. ↩

88. Lisanne Bainbridge, “Ironies of Automation,” Automatica 19, no. 6 (1983): 775–779, at 776; she cites Mackworth’s monograph of 1950. ↩

89. Parasuraman and Riley, Human Factors 39, no. 2 (1997): 230–253, record; Parasuraman and Manzey, Human Factors 52, no. 3 (2010): 381–410, PubMed 21077562. ↩

90. Alexander Eriksson and Neville A. Stanton, “Takeover Time in Highly Automated Vehicles,” Human Factors 59, no. 4 (2017): 689–705, Southampton ePrints. ↩

91. John D. Lee and Katrina A. See, “Trust in Automation: Designing for Appropriate Reliance,” Human Factors 46, no. 1 (2004): 50–80, PubMed 15151155. ↩

92. Nicholas Carr, The Glass Cage: Automation and Us (New York: W. W. Norton, 2014); see the review in the Christian Science Monitor, 29 September 2014. ↩

93. Chauncey Starr, “Social Benefit versus Technological Risk,” Science 165, no. 3899 (1969): 1232–1238; on the critique, “Risk perception,” Wikipedia. ↩

94. Paul Slovic, “Perception of Risk,” Science 236, no. 4799 (1987): 280–285. ↩

95. Gerd Gigerenzer, “Dread Risk, September 11, and Fatal Traffic Accidents,” Psychological Science 15, no. 4 (2004): 286–287; Wolfgang Gaissmaier and Gerd Gigerenzer, Psychological Science 23, no. 12 (2012): 1449–1454, Max Planck Society; on the critiques, Ayton, Murray and Hampton, Judgment and Decision Making 14, no. 3 (2019): 280–287. ↩

96. Peng Liu, Run Yang and Zhigang Xu, “How Safe Is Safe Enough for Self-Driving Vehicles?” Risk Analysis 39, no. 2 (2019): 315–325, RePEc. ↩

97. Azim Shariff, Jean-François Bonnefon and Iyad Rahwan, “How Safe Is Safe Enough? Psychological Mechanisms Underlying Extreme Safety Demands for Self-Driving Cars,” Transportation Research Part C 126 (2021): 103069; Berkeley J. Dietvorst, Joseph P. Simmons and Cade Massey, “Algorithm Aversion,” Journal of Experimental Psychology: General 144, no. 1 (2015). ↩

98. Thomas C. Schelling, “The Life You Save May Be Your Own,” in Samuel B. Chase Jr., ed., Problems in Public Expenditure Analysis (Brookings, 1968), 127–162; H. Spencer Banzhaf, “Retrospectives: The Cold-War Origins of the Value of Statistical Life,” Journal of Economic Perspectives 28, no. 4 (2014). ↩

99. Deborah A. Small and George Loewenstein, “Helping a Victim or Helping the Victim,” Journal of Risk and Uncertainty 26, no. 1 (2003): 5–16; on Jenni and Loewenstein (1997), Lee and Feeley (2016) and the later reanalysis, “Identifiable victim effect,” Wikipedia. ↩

100. Kalra and Paddock, Driving to Safety, RAND, 2016; journal version Transportation Research Part A 94 (2016): 182–193. ↩

101. RAND, “Why Waiting for Perfect Autonomous Vehicles May Cost Lives,” 2017. ↩

102. Philippa Foot, “The Problem of Abortion and the Doctrine of the Double Effect,” Oxford Review 5 (1967): 5–15; Judith Jarvis Thomson, “Killing, Letting Die, and the Trolley Problem,” The Monist 59, no. 2 (1976), and “The Trolley Problem,” Yale Law Journal 94, no. 6 (1985): 1395–1415. ↩

103. Bonnefon, Shariff and Rahwan, “The Social Dilemma of Autonomous Vehicles,” Science 352 (2016): 1573–1576, MIT DSpace. ↩

104. Edmond Awad et al., “The Moral Machine Experiment,” Nature 563 (2018): 59–64, MIT DSpace. ↩

105. acatech, “BMVI Ethics Commission publishes automated driving guidelines”, 20 June 2017; on rule 9, Hunton Andrews Kurth. ↩

106. Sven Nyholm and Jilles Smids, “The Ethics of Accident-Algorithms for Self-Driving Cars: An Applied Trolley Problem?” Ethical Theory and Moral Practice 19, no. 5 (2016): 1275–1289. ↩

107. Johannes Himmelreich, “Never Mind the Trolley: The Ethics of Autonomous Vehicles in Mundane Situations,” Ethical Theory and Moral Practice 21, no. 3 (2018): 669–684. ↩

108. Yochanan E. Bigman and Kurt Gray, “Life and Death Decisions of Autonomous Vehicles,” Nature 579 (2020): E1–E2. ↩

109. Julian De Freitas et al., “From Driverless Dilemmas to More Practical Commonsense Tests for Automated Vehicles,” PNAS 118, no. 11 (2021): e2010202118. ↩

110. Brian Christian, The Alignment Problem: Machine Learning and Human Values (New York: W. W. Norton, 2020); Christian discussing the case on The Closed Session. ↩

111. Andreas Matthias, “The Responsibility Gap: Ascribing Responsibility for the Actions of Learning Automata,” Ethics and Information Technology 6, no. 3 (2004): 175–183. ↩

112. Robert Sparrow, “Killer Robots,” Journal of Applied Philosophy 24, no. 1 (2007): 62–77. ↩

113. Madeleine Clare Elish, “Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction,” Engaging Science, Technology, and Society 5 (2019): 40–60. ↩

114. Filippo Santoni de Sio and Jeroen van den Hoven, “Meaningful Human Control over Autonomous Systems: A Philosophical Account,” Frontiers in Robotics and AI 5 (2018): 15. ↩

115. Filippo Santoni de Sio and Giulio Mecacci, “Four Responsibility Gaps with Artificial Intelligence,” Philosophy & Technology 34, no. 4 (2021): 1057–1084. ↩

116. Daniel W. Tigard, “There Is No Techno-Responsibility Gap,” Philosophy & Technology 34, no. 3 (2021): 589–607; Peter Königs, “Artificial Intelligence and Responsibility Gaps: What Is the Problem?” Ethics and Information Technology 24 (2022): 36. ↩

117. Bernard Williams, “Moral Luck,” and Thomas Nagel, “Moral Luck,” Proceedings of the Aristotelian Society, Supplementary Volume 50 (1976): 115–135 and 137–151; summary of Nagel’s paper. ↩

118. Annette Baier, “Trust and Antitrust,” Ethics 96, no. 2 (1986): 231–260. ↩

119. Onora O’Neill, A Question of Trust: The BBC Reith Lectures 2002 (Cambridge: Cambridge University Press, 2002), Open Library. ↩

120. David Collingridge, The Social Control of Technology (London: Frances Pinter, 1980); “Collingridge dilemma,” Wikipedia. ↩

121. Peter D. Norton, Fighting Traffic: The Dawn of the Motor Age in the American City (Cambridge, MA: MIT Press, 2008), MIT Press Bookstore. ↩

122. Langdon Winner, “Do Artifacts Have Politics?” Daedalus 109, no. 1 (1980): 121–136; Bernward Joerges, “Do Politics Have Artefacts?” Social Studies of Science 29, no. 3 (1999): 411–431; Steve Woolgar and Geoff Cooper, “Do Artefacts Have Ambivalence?” ibid., 433–449. ↩

123. Ivan Illich, Energy and Equity (London: Calder & Boyars, 1974), as discussed by Resilience.org. ↩