31 AUG 2026 — Occupational safety filings show autonomous vehicle test drivers sustained more than two dozen injuries across 2024 and 2025, caused by hard braking and sudden movements by the vehicles they were supervising. Every robotaxi safety argument so far has been about passengers and pedestrians. The only injuries anyone has counted are to the workers inside.
What the filings show
Reporting by Sean O'Kane at TechCrunch, drawing on data submitted to the United States Occupational Safety and Health Administration and interviews with current and former workers, documents more than two dozen injuries to test drivers at Waymo and Zoox over the two years. The injuries are sprains, strains and whiplash.
The mechanism is consistent and unglamorous. The vehicle brakes hard or swerves without warning, and the driver's body absorbs the force. Whiplash is the recurring injury, and some workers were off for months.
Two incidents show the range. On 5 January 2025 a Zoox vehicle driving autonomously in San Francisco braked harshly and injured the shoulder and upper arm of a contractor riding inside, who spent 16 days on restricted duty. In Phoenix last September a test driver was unable to work for 157 days after an autonomous vehicle braked suddenly with no obstacle in its path.
Why the number is a floor
The count covers employers subject to OSHA reporting. Test drivers at other developers are very likely being injured too, and their employers may be exempt from the requirement that produced this data.
Injuries below the recording threshold do not appear at all, and a jolt that produces stiffness rather than lost time is exactly the kind that goes unlogged.
The 16 Waymo injuries are not recorded against Waymo
This is the structural fact underneath the story, and it is documented rather than inferred.
The 16 injuries tied to Waymo depot locations in three cities were reported by Transdev, the company that employs and manages Waymo's test drivers. Zoox reported as many as eight itself, and the San Francisco case involved a contractor riding in the vehicle.
So an injury caused by Waymo's software, in a Waymo vehicle, at a Waymo depot, appears in the safety record of a staffing company. That is a lawful and entirely ordinary consequence of how contract labour is structured, and it means the obvious research method, looking up an AV developer's injury record, returns a number that is not the number. Anyone comparing developers on published safety data is comparing their contracting arrangements as much as their software.
The more than two dozen figure is what remains after several reporting filters. Treating it as the total would be the standard error with regulatory data, which measures what must be reported rather than what happened.
The safety debate has been about the wrong people
Every public argument about autonomous vehicles concerns third parties. Disengagement rates, pedestrian detection, whether a camera-only stack is sufficient, how many miles before removing the driver — we covered the sensor-redundancy version of that argument in New Jersey's bill requiring a camera plus two further modalities, and the regulatory-vacuum version in London, where the rules have not been written.
None of those frameworks counts the driver as a person who can be hurt. The safety driver appears in AV regulation as a control, a fallback that makes testing permissible, and controls are not usually thought of as having necks.
This framing has a consequence. Regulators count disengagements — the times a human intervenes. They do not count the times the vehicle acts abruptly and the human simply takes the jolt. The two are different events with different causes. A system that brakes hard unnecessarily produces no disengagement and no incident report, and it produces whiplash.
Hard braking is a design decision, not a malfunction
This is the part that will not be fixed by better perception.
The Phoenix case is the one to hold onto: the vehicle braked with nothing in its path. Faced with an ambiguous obstacle, or with a sensor return it cannot resolve into an object, an autonomous system chooses between braking hard and continuing. Braking is the safe error in every framework anyone would write down, because the cost of an unnecessary stop is discomfort and the cost of not stopping can be a fatality. Every developer tunes towards braking, and every developer should.
The occupants absorb the difference. A human driver telegraphs a stop: they lift off, the nose dips, the passenger braces. An automated system goes from cruise to maximum deceleration with no physical warning at all, which is why the injuries are whiplash rather than anything else. Bracing is what protects a neck, and there is nothing to brace against.
This is not a bug to be ironed out before launch. It is the current behaviour of a correctly tuned system, and it will persist in commercial service, where the passenger in the seat may be untrained, unprepared, elderly, pregnant or holding a child.
What the region should take from it
The data comes from American suburbs. Traffic in Southeast Asia is denser and less predictable, which suggests the braking problem would be worse there rather than better.
Motorcycles filtering between lanes, informal stops, unmarked crossings and street vendors all produce exactly the ambiguous-obstacle situations that trigger conservative braking. A stack tuned in Phoenix or Chandler will brake more often in Ho Chi Minh City or Jakarta, not less, and the resulting ride quality is a passenger-experience problem before it is a safety one.
The regulatory point follows. A jurisdiction writing AV rules now can require reporting of hard-braking events and occupant injuries alongside collisions, at essentially no cost, because the vehicles already log deceleration continuously. Nobody currently asks for it. That is a cheap requirement to add before deployment and a difficult one to add afterwards, which is the same sequencing lesson that surveillance procurement keeps teaching.
The labour question underneath
Test driving is presented as a transitional job, which shapes how it is treated.
The work is monotonous, requires sustained attention with nothing to do, and is widely understood by both parties to be training a system that will replace it. That combination is unusual, and it weakens ordinary workplace-safety mechanisms. Complaining about a job you have been told is temporary is a different calculation from complaining about a career.
The injuries are the visible part. They surfaced because OSHA has a reporting threshold, and everything below that threshold is invisible by construction.
View author profile → · Editorial policy
