90 incidents on record · 2026 Headlights Incident reports by Ellie Harris · Melbourne
10 new this week Library last updated 13 July 2026
← The incident library
HD-INC-043
Transport · United States · 2023 · Unconstrained / manipulated action

A Cruise robotaxi dragged a pedestrian twenty feet, and the company left the dragging out of what it told regulators

By Ellie Harris · Filed 2 October 2023

Alleged: Cruise (General Motors) developed or deployed the AI system implicated in this incident. Details are drawn from public reports; parties are presumed innocent of any wrongdoing not established by an official finding.

A Cruise robotaxi dragged a pedestrian twenty feet, and the company left the dragging out of what it told regulators

What happened

On the night of 2 October 2023, at Fifth and Market streets in San Francisco, a woman crossing the road was struck by a human-driven car whose driver fled. The impact threw her into the path of a Cruise robotaxi, a driverless car operated by General Motors’ autonomous-vehicle unit. The Cruise car braked and stopped, but then attempted a pull-over manoeuvre with the woman still beneath it, dragging her about twenty feet before coming to rest with a wheel on her. She was severely injured.

The initial collision was not Cruise’s doing. What followed was. In the days after, regulators said Cruise had not given them the full picture: in a meeting with federal officials the morning after the crash, and in a report filed that day, the company did not disclose that its vehicle had dragged the woman. Cruise later provided the video, but the omission became the centre of the story.

The California Department of Motor Vehicles suspended Cruise’s driverless permit on 24 October 2023, citing risk to the public and a failure to represent the incident fully. The California Public Utilities Commission followed. Cruise paused its operations across the country. A subsequent federal process produced penalties: the National Highway Traffic Safety Administration imposed a $1.5 million penalty for failing to report the crash fully, and the US Justice Department imposed a $500,000 fine, with Cruise accepting responsibility for filing a false report. Cruise reached a settlement with the injured woman.

What an auditable version would have shown

A driverless car generates a complete record of what it sensed and did. The failure here was not that the record did not exist; it was that what reached the regulator was partial, and the most damaging fact, the dragging, was the part missing. An auditable version makes the incident record a signed, tamper-evident artefact: the full sequence of perception, decision and motion, fixed at the time and verifiable by a third party, so that what the operator hands a regulator can be checked against what the vehicle actually did. With that, an account that omits the dragging does not survive contact with the record. The point of the record is precisely to remove the operator’s discretion over which parts of an incident the authorities get to see.

Where the gap was

There were two failures, and the second is the one this library exists for. The first was a control failure: after a collision, the vehicle executed a pull-over with a person underneath it, a manoeuvre that should be impossible when a vulnerable road user may be in contact with the car. A ConstraintGate is the control on that, refusing a movement that has not cleared a check for a person in or under the vehicle’s path. The second was an accountability failure: the company controlled the narrative given to regulators, and used that control. A ConductRecord, signed and tamper-evident, is the control on that, because it makes the full conduct of the vehicle independently verifiable and removes the option of reporting a tidier version. A record the operator can edit is not evidence. A record the operator cannot edit is.

What governance should have looked like

Autonomous vehicles will have accidents, including ones they did not start. What a regulator and a victim need is not a promise that nothing will go wrong, but the ability to know exactly what happened and to trust that knowledge. The lesson of this incident is that the technical record and the reported record diverged, and that the divergence favoured the company. The fix is to make the conduct record the source of truth: complete, signed at the time, and verifiable by the regulator directly, so that what the vehicle did and what the company says it did are the same thing. The hardest part of automated accountability is not capturing the data. It is making sure the operator cannot choose which of it counts.

The reference implementation of ConductRecord and ConstraintGate is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.

Sources

The mailing list

Fresh incident reports every week. One email to match.

We add new incidents to the library regularly, and send a single short email each week with what's new. The library stays free and open; this is just how you keep up with it.

No tracking. Unsubscribe in one click.

The record

An auditable system would have produced a signed, tamper-evident record the moment this happened: what the system did, the version that did it, the basis it acted on, and the action taken, and Cruise (General Motors) could have produced it on demand.

This is the record the system as deployed did not produce in a signed, auditable form.

What this teaches
Capture what happened when it happens
What the system did, the version that did it, the basis it acted on, and the action taken, recorded at the moment, not reconstructed after.
Sign it, so no one has to trust the record-keeper
A tamper-evident entry. Edit it later and the signature breaks. The record does not ask for the benefit of the doubt.
Make it verifiable by anyone
A court, a regulator, a customer's lawyer can check the record themselves, without taking the company, or us, at our word.

Headlights summarises publicly reported AI incidents. All summaries are independently written, attributed to their original sources, and intended for research and educational purposes. Allegations are identified as such until established through official findings.

Last reviewed June 2026. This report is based on the sources listed above and reflects information available at the time of review; later developments may not be captured. Where a person is described as charged with or alleged to have done something, that allegation is unproven unless a conviction or a court or regulatory finding is stated. Headlights publishes journalism and commentary, not legal advice.

Want to write back?

Direct to my inbox.

ellie@useheadlights.com →