110 incidents on record · 2026 Headlights Incident reports by Ellie Harris · Melbourne
10 new this week Library last updated 27 July 2026
← The incident library
HD-INC-102
Transport · United States · 2023 · Safety-critical automation

Tesla recalled more than two million cars after US regulators found Autopilot's safeguards against driver misuse were inadequate

By Ellie Harris · Filed NHTSA defect investigation 2021 to 2023

Alleged: Tesla 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.

Tesla recalled more than two million cars after US regulators found Autopilot's safeguards against driver misuse were inadequate

What happened

It was reported that in December 2023 the United States National Highway Traffic Safety Administration concluded a defect investigation into Tesla’s Autopilot, which it had opened in 2021, and found that the system’s methods for keeping drivers engaged were inadequate. Tesla recalled more than two million vehicles in the United States, its Model S, 3, X and Y built between October 2012 and December 2023, in what was reported to be the largest recall in the company’s history. The regulator found that Autopilot “can provide inadequate driver engagement and usage controls that can lead to foreseeable misuse”, meaning a driver could treat a driver-assistance feature as though it were full self-driving and stop paying attention to the road.

According to the agency, since 2016 it had opened dozens of special investigations into crashes in which Tesla driver-assistance systems were suspected of being in use, with at least seventeen deaths among them. The recall remedy was delivered largely as an over-the-air software update that made warnings more prominent and would disengage Autosteer if a driver kept ignoring prompts to pay attention. In April 2024 the regulator opened a further query into whether that fix went far enough, after continued reports of crashes, and that inquiry remained open.

What an auditable version would have shown

A driver-assistance system sits between the car and the person, and the safety question is whether the person stayed responsible for the driving the feature was never certified to do on its own. An auditable version keeps that on the record: what the system was doing at each moment, whether the driver was engaged, what warnings were given and how the driver responded, and whether the feature was being used outside the conditions it was designed for. With that record, “the driver misused it” and “the safeguards were inadequate” stop being competing stories, because the engagement data shows what actually happened in the seconds before a crash, and the rate of misuse across the fleet shows whether the controls were working as a standing number rather than a case-by-case argument.

Where the gap was

The system allowed a level of driver disengagement that its own design could not safely support, and the controls meant to prevent that were, the regulator found, not enough. A ConstraintGate encodes the operating limits the feature was built for and enforces them, escalating warnings and disengaging when a driver stops paying attention, so misuse is caught by the system rather than tolerated until a crash. A ConductRecord keeps the account of driver engagement and system state over time, so that after an incident the sequence can be reconstructed from evidence, and a MetricRecord turns misuse and disengagement into a fleet-wide figure that shows whether the safeguards are actually working. The regulator’s finding was about the controls and monitoring around how people used the system rather than the driving software itself, so the gap sat in the safeguards and the records, not the automation alone.

What governance should have looked like

When automation takes over part of a safety-critical task but still needs a human ready to take back control, the whole system has to be judged on whether it keeps that human engaged, and it should be able to show, from its own data, that it did. Warnings that can be ignored are not a safeguard, and a claim that drivers were misusing the system is only as good as the record of what the system did to stop them. The recall framed foreseeable misuse as a design and monitoring problem for the maker to address, and the evidence for whether it did is in the engagement record.

Failure Pattern: a driver-assistance system permitted foreseeable driver disengagement its design could not safely support, and its safeguards and records were inadequate to prevent or reconstruct the resulting misuse.

Governance Principle: a safety-critical system that relies on a human staying engaged should enforce the limits it was designed for and keep a record of engagement and system state, so misuse can be prevented and, if it occurs, reconstructed from evidence.

The reference implementation of ConstraintGate and ConductRecord 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 Tesla 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 →