NHTSA: most Waymo school-bus passes came after the fix

NHTSA opened a recall query covering 2,895 Waymo vehicles on November 6. A majority of the reported school-bus passes happened after Waymo said its software fix was fully deployed on November 17, 2025.

That is the whole story in two sentences, and it is worse than it sounds. The fix was the recall. The recall was supposed to be the end of it. Eleven months later, the federal regulator is formally asking whether the fix did anything at all.

What NHTSA actually filed

I read the primary document, not just the coverage. NHTSA's Office of Defects Investigation published the resume for investigation RQ26006 on November 6, 2026. The facts, straight from it:

• Population: 2,895 vehicles running Waymo's 5th-generation automated driving system.
• The original recall: Waymo filed recall 25E084 on December 10, 2025, for a defect that could cause its vehicles to pass a stopped school bus with flashing red lights or an extended stop arm, after initially slowing or stopping.
• The remedy: a software update that Waymo said was completed for all affected vehicles on November 17, 2025.
• The problem: "Of the total reported incidents identified during PE25013, a majority of the incidents occurred after the November 17, 2025 remedy date."
• The action: NHTSA opened the recall query "to assess the adequacy of the remedy for Recall 25E084 and to identify the cause(s) of the conditions leading to the post-remedy reports."

Reuters first reported the filing on Monday, November 9. TechCrunch, Morningstar/Dow Jones, and Seeking Alpha all confirmed the same day. The story checks out.

A recall query is not a recall. It is the step before one — the regulator saying, in effect, "you told us this was fixed, and our data says otherwise, so prove it." It can lead to a second recall.

The timeline is the indictment

Run the dates:

• October 2025: NHTSA opens preliminary investigation PE25013 after footage shows a Waymo robotaxi maneuvering around a stopped school bus in Atlanta — crossing perpendicularly in front of the bus, then turning left around it, while the bus was unloading children with its stop sign extended and lights flashing.
• November 17, 2025: Waymo says the software remedy is complete on all affected vehicles.
• December 10, 2025: Waymo formally files recall 25E084 — 23 days after it says the fix was already deployed.
• November 6, 2026: NHTSA opens RQ26006 — 354 days after the fix was declared complete — because most of the violations happened after that date.

That 23-day inversion is the detail that sticks with me. The fix went out before the recall was filed. The recall, in other words, was paperwork ratifying a software push that had already happened. And the software push, per NHTSA's own count, did not stop the behavior it was meant to stop.

Austin Independent School District reported at least five such incidents at the time, according to a letter filed with NHTSA. The National Transportation Safety Board separately documented a January 12 incident in Austin in which a Waymo vehicle initially stopped for a school bus, then proceeded while the stop arms were still extended — six vehicles passed the bus; no crash occurred. NHTSA's resume lists 8 incident reports in its summary, with crash and injury counts marked confidential.

There is also a separate, still-open probe from January, after one of Waymo's vehicles struck a child near an elementary school in Santa Monica. That is a different investigation, but it is the same file drawer: children, school zones, and a system that was supposed to have learned.

The verification problem with software recalls

Here is the engineering question this story actually raises, and it is not "are robotaxis safe." It is: how does anyone verify a software fix for a behavior that only shows up at the edge?

A school bus with flashing red lights and an extended stop arm is a rare, high-stakes, highly specific scene. It is exactly the kind of scenario a fleet validates in simulation and in staged testing — and exactly the kind where simulation and the street diverge. Waymo's fix presumably passed whatever validation gate Waymo set for it. Then the fleet kept doing the thing, in Austin and Atlanta, in front of school districts that count violations and file letters.

This is the structural weakness of the software-update-as-recall model. With a hardware recall, the remedy is observable: a part is replaced, a dealer stamps a form. With a software recall, the remedy is a claim — "deployed to all affected vehicles on November 17" — and the only independent check is waiting to see whether the behavior recurs. NHTSA waited. It recurred. That is what a recall query is: the regulator discovering, a year later, that the verification loop was "ship it and watch."

The honest gap in the data is worth stating plainly. NHTSA says "a majority" of incidents came after the fix, but the resume does not publish the total incident count. "Majority" could mean 6 of 11 or 60 of 100, and those are very different fleets. The regulator knows the denominator; it has not shared it. Until it does, the precise failure rate is unconfirmed — but the direction is not. A fix that worked would produce zero post-remedy incidents, not a majority.

One more number for scale: the query covers 2,895 vehicles out of the 3,067 in the original recall — 94.4% of the recalled fleet is back under investigation. Waymo was valued at $126 billion in February. The company has not yet responded to TechCrunch's request for comment, and neither has Austin ISD.

What to watch

The query has two stated jobs: assess whether the remedy was adequate, and identify what caused the post-remedy reports. The second one is the interesting one. If the cause is a perception miss — the system not reliably detecting a stop arm — that is one kind of fix. If the cause is a planning decision — the system detecting the bus and choosing to proceed anyway — that is a much more serious one, because it means the "fix" tuned a behavior the stack still considers acceptable in some conditions.

Either way, the precedent matters beyond Waymo. Every AV operator now recalls by software push. If the first major test of that model ends with the regulator reopening the case a year later because the fix did not take, the next software recall will not be taken on the manufacturer's word. Expect NHTSA to start asking for the validation evidence up front — the test matrix, the pass criteria, the field monitoring plan — before accepting "deployed to all vehicles" as a remedy.

The fix was the fix. It did not fix it. That is the story.

Sources: NHTSA ODI Resume, Investigation RQ26006, opened November 6, 2026 (static.nhtsa.gov); TechCrunch, "Waymo robotaxis kept illegally passing school buses after software fix, feds say," November 9, 2026; Reuters, "US opens probe into Waymo vehicles over school bus safety violations," November 9, 2026; Morningstar/Dow Jones, November 9, 2026. The NTSB Austin incident and Austin ISD figures are as reported by TechCrunch and secondary coverage. Waymo and Austin ISD had not responded to TechCrunch's comment request at publication time. A recall query is an investigation into a prior remedy, not a new recall or a final finding of defect.