Waymo's self-driving cars are 94 percent safer than human drivers at preventing serious injuries, according to the company's own data. Your car insurance premium is not going down.
That contradiction is the real story of this week in autonomous vehicles. Last week, we tracked the commercialization thread: the expansion plans, the city-by-city rollouts, the quiet accumulation of permits and partners. This week, the edge cases surfaced — a pedestrian in Wuhan, a flood in Phoenix, and an insurance industry that sees the safety data but isn't repricing risk yet.
Start with Wuhan. A Baidu Apollo Go robotaxi made contact with a pedestrian at the intersection of Yingwu Avenue and Guobo Avenue in Hanyang district. The company's account, given to Yicai, is specific: the vehicle was starting from a green light when a person ran a red light. The contact was minor. Baidu says it cooperated with police immediately and accompanied the person to the hospital. No obvious external injuries; the person is under observation.
This is the first confirmed pedestrian contact for Apollo Go. That matters because Baidu has been the most aggressive operator in China. Wuhan is its flagship. The company holds the country's first batch of autonomous commercial licenses, granted when the Wuhan smart connected vehicle test zone opened on September 22, 2019. By September 2022, Wuhan and Chongqing issued the first fully driverless commercial pilot policies — allowing no safety driver, public roads, commercial service.
The scale is real. Baidu's Q1 earnings showed roughly 826,000 Apollo Go orders, up 25 percent year over year. Cumulative rides passed 6 million by mid-April. Robin Li told analysts that fully driverless orders in Wuhan exceeded 55 percent in Q1 and hit 70 percent in April, with a path to 100 percent. The company expects unit economics breakeven in Wuhan soon, and full profitability in the city by 2025.
Now the edge case. Li himself, in a separate interview, said something the company's promotional material doesn't emphasize: after 11 years, "we still cannot say this technology is mature." He called it a specialized scenario. That's the CEO of the company running the largest robotaxi fleet in China, speaking before the incident.
after 11 years, "we still cannot say this technology is mature." He called it a specialized scenario.
The incident response from Baidu was fast — police, hospital, statement. But the online reaction revealed the fault line. One commenter asked who goes to prison if a driverless car kills a pedestrian and is found fully at fault. Another suggested pedestrians who run red lights should be told the cars won't brake for them. A third asked about insurance fraud: "Will there be a lot of people faking accidents?"
Those questions don't have settled answers. And that's the point.
Jump to Phoenix. A Waymo robotaxi got stuck on a flooded street for more than ten minutes, caught in what witnesses described as an "AI loop" — inching forward, then backward, then forward again. Two Waymos were involved in the standoff, holding up traffic. One eventually passed. The other kept rocking, unable to find a way around the water. Video showed it nearly reversing into a car behind it.
This is not Waymo's first water problem. In April 2026, an unoccupied Waymo in San Antonio drove into floodwaters and was swept into Salado Creek. The company paused service there and filed a voluntary safety recall covering 3,791 vehicles. NHTSA's report noted a software defect: the cars would slow down but still proceed into untraversable flooded lanes on higher-speed roads. That's a significant loss-of-control risk.
A second recall followed in June, covering nearly 3,900 robotaxis, after more than a dozen incidents where Waymos drove past ramp closures into active freeway construction zones. Waymo briefly restricted freeway operations until an over-the-air update shipped.
These are not exotic scenarios. Flooding happens. Construction zones happen. The technology handles the routine 99 percent, then fails on the edge cases that human drivers handle with a glance and a judgment call.
Waymo's broader safety record is genuinely strong. That 94 percent reduction in serious injury crashes, measured over 220 million rider-only miles, is not trivial. The company serves more than half a million paid trips weekly across 11 U.S. metro areas. It just signed a partnership with Allianz Partners to build insurance, claims, and safety infrastructure for its European expansion into London and Munich.
Which brings us to the insurance gap. Alex Timm, founder of Root Inc., a publicly traded auto insurer, laid out the paradox on a recent industry panel. Autonomous vehicles are safer, but the insurance market hasn't repriced because AVs represent less than 1 percent of miles driven in the U.S. — roughly 360 million cumulative AV miles against 3.2 trillion human-driven miles annually. The $350 billion personal auto insurance market remains intact.
But the cost side is already shifting. A Waymo is a Jaguar I-PACE loaded with sensors and cameras, estimated at $175,000 per vehicle. If a lidar unit is damaged, Timm says, it's "almost impossible to repair practically." That dynamic — fewer crashes but much more expensive ones — is already affecting conventional cars too. The Highway Loss Data Institute found that driver-assist systems cut claim frequency by 13 to 39 percent, but the remaining claims cost far more. The average auto claim is now near $13,000.
Timm's projection: if AVs maintain their current safety performance, the typical premium — which he pegs at around $15,000 per six-month period for a typical customer — "could very well be halved or more." But not yet. Not at 1 percent mileage share.
Root's strategy is to get inside the vehicle. The company is partnering with OEMs that control the technology, embedding insurance directly into the car's systems. "They rarely crash, but they're also rarely in high-risk driving environments," Timm said. Pulling data directly from the vehicle would show how risk is actually changing.
The tension here is structural, not technological. The safety data says AVs are better. The insurance economics say the market can't adjust until the mileage share is large enough to matter. The edge cases — a pedestrian in Wuhan, a flood in Phoenix — say the technology still fails in ways that are hard to predict and expensive to fix.
Meanwhile, WeRide is running 12-hour livestreams of its robotaxi operations in Guangzhou. Two streams, 663 kilometers, 230,000 viewers. The company is showing the mundane reality: morning rush hour, evening peak, real passengers. No edits, no pauses.
That's the right instinct. Transparency builds trust. But the streams show the routine. The news shows the edge. And the insurance industry is watching both, waiting for the mileage to add up before it changes the price.