Every trip in a self-driving Waymo has the same dangerous moment. The robotaxi can successfully shuttle you to your destination, stopping carefully at every red light and dutifully following the speed limit. But at the very end, you, a flawed human being, will have to place your hand on the door handle, look both ways, and push the door open.
From mid-February to mid-August of this year, Waymo’s driverless cars were involved in three collisions that came down to roughly identical circumstances: A passenger flung their door open and hit somebody passing by on a bike or scooter. That’s according to an independent analysis of crash reports the company has disclosed to the government, which found that most of the 45 serious accidents involving Waymos were the fault of other motorists or seemingly an act of God. (In one case, a pickup truck being towed in front of a Waymo came loose and smashed into the vehicle.) None were definitively the fault of Waymo’s actual self-driving technology.
Waymo, an AI company that is part of Google, loves to brag about its safety record. In a recent report tracking 96 million miles of fully autonomous rides, Waymo says its cars have been involved in 91 percent fewer accidents resulting in a “serious injury or worse” than cars driven by an average human over the same distance. Experts I spoke with had quibbles with Waymo’s comparisons but agreed that the company has an undeniably strong safety record. “They have not obviously been at fault for any fatalities,” Phil Koopman, an driverless-car expert at Carnegie Mellon, told me. Humans may not always do our job, but by and large, Waymo’s machines are doing theirs.
In a world of AI, it turns out that a 5,000-pound Jaguar SUV may be less concerning than an interactive text box. The AI boom has led OpenAI and many other companies to rush out their products, sometimes with disastrous results: Gemini has engaged in bondage scenarios with adolescent users, Elon Musk’s Grok recently went full Nazi for a few hours, and OpenAI is mired in a pending wrongful-death lawsuit after ChatGPT allegedly played a role in a teen’s suicide. (OpenAI declined to comment; Musk has posted that Grok was “manipulated” into going on an anti-Semitic rampage; and Google, which runs Gemini, has said that it has enacted additional safeguards to protect kids.)
“I like to tell people that if Waymo worked as well as ChatGPT, they’d be dead,” Bryant Walker Smith, a self-driving-car expert at the University of South Carolina School of Law, told me. Imagine if, instead of turning left at a stop light, a robotaxi decided to blast the stereo and start doing figure eights. Waymo pokes a hole in Silicon Valley’s prevailing ethos, especially in the AI age: Move fast and break things. Mark Zuckerberg has said that the risk of “misspending a couple hundred of billion dollars” on AI is smaller for Meta than risking a future in which his company is “out of position.” If you slow down, you might fall behind in building world-changing “superintelligence.”
The stakes are more immediately obvious for self-driving vehicles than for chatbots: Every day, more than 100 Americans die in car crashes, more than a 9/11’s worth of fatalities per month. Yet Waymo’s self-driving competitors have also seemed to adopt the ethos that is animating other AI companies. Over the summer, Tesla rolled out its own robotaxi service in a penis-shaped swath of Austin. (Elon Musk is going to Elon Musk.) In their first month on the streets, driverless Teslas got into three crashes while logging roughly 7,000 miles. Nationally, Waymo’s fleet racks up many more miles every day. Meanwhile, Tesla’s “Full Self-Driving” feature, which is built into many of the company’s cars, has been linked to numerous deaths, and the company is facing lawsuits alleging false advertising. Just this week, two Democratic senators called for safety regulators to investigate the “Full Self-Driving” feature. (Tesla and Elon Musk did not respond to requests for comment.)
Or consider Cruise, a start-up that was quickly expanding nationally in 2023. In October of that year, one of its robotaxis struck a pedestrian in San Francisco after she was hit by another car and thrown in front of the vehicle. The Cruise vehicle dragged her more than 20 feet as it attempted to pull over. (The woman was seriously injured but survived.) Perhaps Cruise got unlucky and Waymo hasn’t. But even at the time, driverless-car experts were not surprised. A Cruise whistleblower had sent a letter to safety regulators in California alleging that the company’s cars weren’t up to snuff, as The Wall Street Journal reported then. After Cruise’s accident in 2023, its parent company, GM, halted its robotaxi business for good. (GM declined to comment.)
Compared with its robotaxi competitors, “Waymo has moved the slowest and the most deliberately,” Smith said—which may be a lesson for the world’s AI developers. The company was founded in 2009 as a secretive project inside of Google; a year later, it had logged 1,000 miles of autonomous rides in a tricked-out Prius. Close to a decade later, in 2018, Waymo officially launched its robotaxi service. Even now, when Waymos are inching their way into the mainstream, the company has been hypercautious. The company is limited to specific zones within the five cities it operates in (San Francisco, Phoenix, Los Angeles, Austin, and Atlanta). And only Waymo employees and “a growing number of guests” can ride them on the highway, Chris Bonelli, a Waymo spokesperson, told me. Although the company successfully completed rides on the highway years ago, higher speeds bring more risk for people and self-driving cars alike. What might look like a few grainy pixels to Waymo’s cameras one moment could be roadkill to swerve around the very next.
That’s not to say that Waymo’s extreme caution has resulted in perfection. Experts I spoke with pointed out some limitations in how the company compares itself to human drivers. For one, the average car on the road is more than a decade old, which makes it less safe than Waymo’s fleet of new, regularly serviced vehicles. Waymos also have glitched in ways that do not show up in aggregated crash data: There are anecdotes of the robotaxis driving into a flooded street, getting stuck in wet cement, and blocking two firetrucks. Last week, cops in the Bay Area pulled over a Waymo for making an illegal U-turn. “Since there was no human driver,” the police department wrote on Facebook, “a ticket couldn’t be issued (our citation books don’t have a box for ‘robot’).”
But such examples make up a vanishingly tiny portion of rides from a company that has come to dominate the world of robotaxis. The overwhelming majority of driverless-car rides in the United States on any given day happen in a Waymo. Bonelli told me it now completes “hundreds of thousands of paid rides per week.” Parents bring home their newborns in Waymos. High-schoolers head to first period in Waymos. And many more of the company’s robotaxis are about to hit the roads: The company recently won approval to start testing airport pickups in San Francisco and San Jose, California; next year, Waymo plans to debut its service in Washington, D.C.; Nashville; Miami; Dallas; and Denver.
Robotaxis are only as good as the driving data they have ingested. Building a driverless car that can roll through a cul-de-sac is relatively easy; building one that is ready for every possible scenario it might face on the road is much harder. Thanks to Google’s enormous war chest, Waymo has had the luxury to spend 16 years raking in data to create more precise self-driving cars, all with the hope of an eventual payoff. Waymo has been around for longer than Lyft, Instagram, and Snapchat—yet it has still never turned a profit. The robotaxi business is part of a unit of Alphabet called “Other Bets,” which lost a staggering $1.25 billion from April to June.
Self-driving cars still have a lot to prove when asphalt gives way to snow and black ice. As it enters new cities, Waymo says it is preparing for any possible weather. “We have been in Detroit in the snow and Seattle in the rain and thunderstorms in Miami,” Waymo’s co-CEO Tekedra Mawakana recently told the Atlantic contributing writer Derek Thompson. Because Waymo’s cars lack “someone to physically scrape off the icy build up,” Bonelli, the Waymo spokesperson said, “we implemented preventive measures to each of our sensors to maintain a clean view of its surroundings.” Waymos also can’t avoid the highway forever. On city streets, a confused robotaxi can just pull over. Robotaxis traveling at highway speeds can’t easily do that.
At any point, the company’s luck could run out. “We could see a fatality caused by the technology tomorrow,” Koopman said. For now, if a single death is tied back to an errant line of Waymo code, it could imperil the company’s future. As Waymo conquers one city and highway at a time, however, it may eventually become something like transportation infrastructure. Alphabet’s vision for Waymo points to a future in which its fleet of robotaxis is part of every aspect of travel. Why drive yourself to work when you could just order a Waymo? Why waste your Saturday getting a tune-up for your RAV4 when a driverless Waymo doesn’t need a chaperone? “They want to be Uber but also Toyota, the car dealer, Exxon, Pep Boys, State Farm, your transit agency, GMAC, UPS, Michelin and more—all combined,” Brad Templeton, an early Waymo consultant, has written.
Like other AI companies, Waymo wants to rule the world. Chatbots that cure cancer might still be a work in progress, but Waymos are already taking over cities. In the areas of San Francisco where it operates, Waymo is ferrying more riders than Lyft. Every new rollout of safe Waymos is one step toward a future in which the company is embedded in American life. Whether slow or fast, the AI endgame is all the same.
The post Move Fast and Break Nothing appeared first on The Atlantic.
