The robotaxi took 20 years to arrive: the lesson nobody applies to AI's other promises

THE NUMBER THAT MATTERS. Our thesis is simple. The robotaxi works now, but only after two decades, many disappointments and billions of dollars spent. That pace is not an exception. It is the norm when AI leaves the screen and enters the street. The numbers from Waymo, Alphabet's self-driving unit, show it clearly. In December 2025 the company passed 20 million trips with public riders, 15 million of them in 2025 alone. In March 2026 it announced 500,000 paid rides a week across ten US cities. A year and a half earlier it was doing about 100,000. In February it raised $16 billion at a $126 billion valuation. Its stated goal is more than one million weekly rides by the end of 2026. As of mid-October it had not published a figure newer than March's, so that goal remains unconfirmed.

TWENTY YEARS IN THREE ACTS. Act one was the demo. In 2004, no vehicle finished the self-driving race run by DARPA, the US military research agency. In 2005 a Stanford car called Stanley won it. In 2009 Google started the project that became Waymo. For years the problem looked almost solved. In 2019 Elon Musk even predicted a million Tesla robotaxis by 2020. Act two was the winter, from 2018 to 2022. Solving 90% of driving situations proved easy. The remaining 10% (roadworks, floods, unpredictable pedestrians, police officers waving traffic through) cost more than everything else. Engineers call these 'edge cases': rare situations a system never saw during training. Act three is today's slow rollout, city by city. Our reading: AI did not fail for lack of intelligence. It failed for lack of reliability in rare situations. Only miles, validation and patience fix that.

WHAT THE SAFETY DATA SAYS (AND DOESN'T). Rigor matters here. Waymo's December 2024 study with the reinsurer Swiss Re covered 25.3 million miles. According to that work, Waymo cars generated 88% fewer property-damage claims than human drivers and 92% fewer bodily-injury claims. The company's safety dashboard goes further. Based on 271 million driverless miles through June 2026, Waymo says it has 95% fewer crashes causing serious injury or death than an average human driver in the same cities. The data is solid and consistent, but it has limits. The company produced or commissioned it. Serious crashes number only a few dozen, so each percentage has a wide margin of error. And the figures only describe the areas where Waymo operates, with their weather, streets and restrictions. They say nothing about a Chicago snowstorm or a roundabout in Madrid. The recent record is not spotless either. In 2026 Waymo issued two software recalls: one for flooded roads and one for freeway roadworks, which covered 3,871 vehicles and gives a sense of its fleet size. In October NHTSA, the US road-safety regulator, opened an investigation into its cars passing stopped school buses. According to local reports, the Austin school district alone has issued about 20 citations this school year. Waymo has announced another voluntary recall. At a Senate hearing in February, it also acknowledged that about 70 remote-assistance agents, some of them in the Philippines, help its cars when they get stuck. The company maintains that those agents give guidance but do not drive. Our reading: the evidence suggests the robotaxi is already safer than the average human in its own territory. It is not yet proven safer everywhere.

TWO STRATEGIES, ONE WARNING. Tesla is taking a different route: cameras only, cheap cars and fast scale. According to the company, it now runs rides with no safety monitor on board in six cities in Texas and Florida, and its AI chief cites 380,000 miles without notable incidents. Waymo has logged roughly 700 times more driverless miles. Tesla's fleet size is disputed. In June, Texas had 42 Tesla robotaxis registered, and independent trackers counted between 20 and 85 active vehicles before the first Cybercabs (the model with no steering wheel) arrived. Tesla does not publish official figures. The warning came from Cruise, General Motors' unit. In October 2023 one of its cars ran over a pedestrian and dragged her about 20 feet, after another vehicle had thrown her into its path. Cruise was fined for withholding details from regulators. In December 2024 GM shut the business after more than $10 billion in losses. In China, Baidu's Apollo Go says it reached peaks of 350,000 weekly rides in March 2026, a scale comparable to Waymo's. Our reading: in the physical world, one badly handled incident can wipe out years of work. The trust of regulators and the public is as important an asset as the model.

THE LESSON FOR THE REST OF AI. Today we hear promises much like the self-driving pitch of 2015. Agents that will do an employee's job. Humanoid robots in every factory. Automated medical diagnosis. The robotaxi suggests three rules for judging them. First: a demo is not a deployment. What matters is the failure rate in rare situations, measured at scale. Second: the final stretch needs invisible human infrastructure. Waymo needs remote assistants and fleet crews, just as software agents will need someone to check their work. Third: progress comes zone by zone. Waymo did not promise to 'drive everywhere'; it started with one Phoenix suburb and expanded from there. Language models work the same way: they do well on well-defined tasks and fail on open-ended ones. Be wary of any timeline that skips these stages. But don't give up during the winter, either. Those who declared the self-driving car dead in 2020 were wrong.

IMPLICATIONS: JOBS AND THE HORIZON. In the short term, the robotaxi raises real problems. Millions of people in the US drive for a living: taxi and ride-hail drivers, couriers and truckers. Their transition will be hard and uneven, and optimism alone won't solve it. Open questions also remain about who is liable for a crash and who oversees remote operators. The international rollout will stay slow. In London, Waymo has not yet confirmed public driverless rides. Tokyo is aiming for 2027 and Toronto for 2028. In the long term, the horizon is worth the patience. The WHO estimates about 1.2 million road deaths a year, and the vast majority are caused by human error. If today's data holds up outside the test zones, self-driving cars could save hundreds of thousands of lives a year and give people back hours of their time. That is abundance in the literal sense: time and lives. That future will arrive late and in pieces, just as the robotaxi did. That is probably how most of AI's promises will arrive.