180 incidents on record · 2026 Headlights Incident reports by Ellie Harris · Melbourne
10 new this week Library last updated 30 August 2026
← The incident library
HD-INC-168
Consumer software · United States · 2025 · Fabricated factual claims about a named person, with fabricated citations

Google's Gemma model produced a fabricated criminal allegation about a sitting US senator, and Google withdrew the model from AI Studio while keeping it available through the API

By Ellie Harris · Filed Senate Commerce hearing 29 October 2025; letter to Google 30 October 2025

Alleged: Google 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.

Google's Gemma model produced a fabricated criminal allegation about a sitting US senator, and Google withdrew the model from AI Studio while keeping it available through the API

What happened

It was reported in October 2025 that a sitting US senator had asked Google’s Gemma model whether she had been accused of rape, and that the model returned a detailed account of an incident during a state election campaign, supported by links styled as news reports. Her office published the letter it sent to Google, stating that there had never been such an accusation, that no such person existed, that the campaign year was wrong, and that the links led to error pages and unrelated articles.

Google said Gemma was built for developers and researchers, was not meant for factual questions or for consumers, and removed it from its AI Studio interface while keeping it available through the programming interface. The senator replied that removing it from one interface does nothing to stop developers deploying it elsewhere. Separately, an activist who had found similar output about himself sued Google, and in July 2026 a court refused to dismiss the case while stating it was making no finding about the truth of the allegations.

What an auditable version would have shown

The people the output was about were the ones who discovered the problem. One ran the prompt in his own name and sued; the other’s office ran the prompt and published the result. The sources reviewed did not show how often the model produced claims like this, or what checks Google ran before returning them.

A record for each answer could show the sources the model used, whether those sources actually supported the claim, and whether the citations were checked before the answer appeared. That would give Google its own record of what happened, rather than leaving the person named in the output to discover it by running the prompt themselves.

Where the gap was

The senator’s letter states that the output presented a false factual claim about her and attached links styled as news reports that led to error pages and unrelated articles. A VerificationGate is designed to route a factual claim about a named person to a trusted source for checking rather than relying on model generation alone. A CitationVerifier is designed to test whether a citation supports the claim attached to it before an answer is delivered, and to record whether the check passed. A ConductRecord preserves the prompt, the retrieved documents, the checks and the output, which allows a disputed answer to be traced afterwards rather than reproduced by the person it concerns. None of these decides what a model should be permitted to say, and none resolves the questions before the Delaware court.

What governance should have looked like

Google’s stated position is that Gemma was built for developers and researchers and was not meant for factual assistance or for consumers. The senator argued that removing Gemma from one interface did not stop developers using it elsewhere. The sources reviewed do not settle that disagreement, and the proceedings are still unresolved.

Where a model is answering factual questions about real people, best practice would be to check those claims against a trusted source before showing them. Citations should also be checked to make sure they actually support what the model is saying, with a record kept of what was checked and what the checks found.

Failure Pattern: a model made a false claim about a real person and backed it up with links that looked credible but did not support what it said. The people named in the output found the problem, not the operator.

Governance Principle: when a model makes factual claims about real people, those claims should be checked against a trusted source. If it provides a citation, the organisation should also be able to show that the citation actually supports the claim.

The reference implementation of VerificationGate, CitationVerifier 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 Google 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 August 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 →