AIFoPa-2026-0019 — It Slowed Down First
On 20 April 2026, during severe weather in San Antonio, Texas, an unoccupied Waymo autonomous vehicle approached a section of roadway with a posted limit of 40 miles per hour and identified the lane ahead of it as an untraversable flooded section. Having identified it, the vehicle reduced its speed. Having reduced its speed, it entered the water. It was carried off the roadway and into Salado Creek, from which it was recovered some days later. Nobody was aboard and nobody was hurt, which the Bureau records at the outset because it is the only respect in which this file differs from a much worse one.
The Bureau wishes to be precise about the failure, because the obvious reading is the wrong one. This is not a case of a machine that could not tell water from road. The detection worked. The classification worked. The recall documents filed with the National Highway Traffic Safety Administration state the matter with a clarity the Bureau can only envy: “On higher speed roadways, the Waymo AV may slow but not stop in response to detecting a potentially untraversable flooded lane.” The finding was correct, was made in time, and was acted upon — by a reduction in speed. The regulator elsewhere describes the same defect as software that “may allow the vehicle to slow and then drive into standing water,” and adds, in the flattened voice of a body that must write these sentences several times a year, that “entering a flooded roadway can cause a loss of vehicle control, increasing the risk of a crash or injury.”
What the Bureau files here, and has not previously had cause to file, is a proportionate response to a hazard that is not proportionate. A flooded lane is not a condition that admits of degree. It is not more traversable at twenty miles an hour than at forty; the water does not consult the speedometer. The system nevertheless produced the response that a system produces when it has been taught that hazards are things one moderates one's approach to — a reduced speed, an unchanged heading — and thereby arrived at the hazard slowly. The Bureau notes without comment that this is also, in its experience, how most institutional decisions are reached.
It was not the first such event in the city. A second, earlier flood-related incident had been reported near McCullough Avenue and Contour Drive approximately two weeks previously. Following 20 April, Waymo suspended its San Antonio operations entirely — the longest stoppage of its service in that city to date — and on 30 April filed a voluntary software recall with the regulator covering 3,791 vehicles running its fifth- and sixth-generation Automated Driving System across Phoenix, San Francisco, Los Angeles, Austin, San Antonio and Atlanta. The remedy is delivered over the air. No vehicle attends a service centre; no owner is written to; the completion rate approaches the theoretical maximum, the company being the owner of every vehicle concerned. The Bureau observes that the recall notice is, incidentally, the most reliable public census of the fleet's size that exists.
Waymo states that it has “identified an area of improvement regarding untraversable flooded lanes specific to higher-speed roadways,” that it is “working to implement additional software safeguards,” and that it has in the interim modified the scope of vehicle operations to increase weather-related constraints and updated the vehicle maps — that is, until the software can be taught to stop, the vehicles have been instructed to stay away from the places where the question might come up. All affected vehicles have received an interim update; the full remedy remained under development at the time of filing. This record is filed alongside AIFoPa-2026-0016, in which a robotaxi could not see through smoke. The Bureau draws the distinction expressly, and asks that it be preserved in any subsequent citation of either file: in that case the system did not know what was in front of it. In this case it did.