Waymo Runs 500,000 Rides a Week and Recalled 3,871 Robotaxis in June. Both Are True.

The Scale Is Real

Waymo is no longer a pilot. In late March 2026 it said it was giving 500,000 paid rides a week across ten US cities — tenfold growth from 50,000 a week in May 2024. Through March it had accumulated 220.6 million rider-only miles, with nobody behind the wheel. Co-CEO Tekedra Mawakana has set a target of one million weekly rides by the end of 2026.

The safety claim is correspondingly large: a 13x reduction in serious-injury-or-worse crashes compared with human drivers over the same road miles. That figure comes from Waymo’s published crash data against matched human benchmark rates, and taken at face value it describes a system meaningfully safer than the median driver.

The Regulatory Record Is Also Real

In June 2026, Waymo recalled 3,871 vehicles running its fifth-generation driving system. It had found at least 13 instances of robotaxis driving into highway sections closed for construction — six in Phoenix in April, seven in the San Francisco Bay Area on 18 May. In its filing, Waymo said the Phoenix vehicles “did not recognize and drove past ramp closure signs into pre-planned freeway construction zones.” It restricted freeway driving while it developed the software fix.

Separately, NHTSA opened an investigation in October 2025 after a Waymo in Atlanta drove around a school bus stopped with red lights flashing and its stop arm deployed. The probe was later expanded. The Austin school district logged at least 20 such violations during the 2025–26 school year; Atlanta identified at least six. The NTSB has opened its own investigation and has characterised the vehicles’ behaviour around the buses as illegal.

And on 23 January 2026, a Waymo struck a nine-year-old crossing midblock in a Santa Monica school zone during drop-off. The child ran out from behind a double-parked SUV; the vehicle braked hard from about 17 mph and hit at roughly 6 mph, causing minor injuries. NHTSA and the NTSB are both investigating, NHTSA specifically assessing whether the vehicle exercised appropriate caution given its proximity to a school during drop-off hours.

None of these investigations has reached a finding. They are open questions, not conclusions.

Why Both Sets of Facts Hold

The instinct is to treat these as contradictory. They are not, and understanding why is the useful part.

The safety statistic is a population average. Human drivers crash constantly and ordinarily, and much of it is fatigue, distraction and intoxication — three things software does not have. Beating that average across 220 million miles is a genuine achievement.

The investigations concern specific failure modes: scenarios the system handles wrongly whenever it meets them. A system can be far better on average and still have a category it reliably gets wrong, because the average is dominated by millions of unremarkable miles where nothing happens.

Averages describe expected harm across a fleet. Failure modes describe the tail — and regulators, courts and parents at a school gate care about the tail.

The Fix That Did Not Hold

Here is the detail that turns this from an argument into evidence.

Waymo said it had identified the software issue behind the school-bus incidents and implemented updates by 17 November 2025. By 1 December, Austin had issued its 20th citation — after the fix.

That is what a correlated failure looks like. Human error is distributed: one driver, one mistake. An autonomous fleet’s gap exists in every vehicle simultaneously and fires wherever the scenario appears. The same property makes it fixable — one patch reaches the whole fleet at once — but only if the patch is right. When it is not, the failure persists everywhere at once, too.

Construction zones make the point from the other side. NHTSA closed an earlier 14-month Waymo investigation in July 2025 without action, citing recalls; that probe had also covered problems entering construction zones. A year later, construction zones produced a 3,871-vehicle recall.

What to Watch

• Time from incident to remedy. Construction-zone failures ran from April to a June recall. That interval, across future defects, is the real safety metric for a fleet that shares one brain.
• Whether fixes are verified in the field, not just shipped. The school-bus record shows the difference between deploying a patch and closing a failure mode.
• What the four open investigations conclude. A confirmed recognition gap in a scenario with statutory stopping rules is a different regulatory object from a crash-rate debate.

If you are judging autonomous driving honestly, hold both numbers at once and stop asking which is the real story. The fleet average tells you what to expect from a million rides. The open investigations tell you what the system still cannot do. Companies quote the first, regulators act on the second, and the gap between them is where this industry will be decided.