When a model says something false about a person

A new way to be defamed

Defamation law is old and it assumes a publisher. A model producing a false, damaging statement about a named person, to one user, in a conversation nobody else sees, does not fit that assumption neatly — and courts are now working out how it fits.

The cases are real. An Australian mayor threatened action in 2023 after a model described him as having been convicted in a bribery case in which he had in fact been the whistleblower. A US radio host sued after a model generated a fabricated legal complaint accusing him of embezzlement; a Georgia court granted the developer summary judgment in 2025, reasoning in part about whether a reasonable reader would treat the output as a statement of fact given the disclaimers and the circumstances. A European privacy group filed a complaint in 2025 on behalf of a Norwegian man about whom a model produced a false account of a serious crime, framing the issue not as defamation but as inaccurate personal data under data protection law.

That last framing may turn out to matter more than defamation. Data protection regimes require personal data to be accurate and give people a right to rectification, and they do not require proof of publication or of reputational damage.

Why models generate this specifically

The mechanism is the one from the third module and it explains why the false statements cluster around certain people.

A name plus a domain — a politician, a lawyer, an academic, a journalist — is exactly the kind of prompt where the model has strong statistical associations and weak specific knowledge. It produces the most plausible completion. For a person with a common name, the details of several people merge. For a person who appears in coverage of a scandal in any role — witness, investigator, whistleblower, defence counsel — the association between the name and the wrongdoing is present in the text, and the role can be lost.

So the systematically most exposed people are those who are named in public material near something bad without having done it. That is a specific, predictable population, and it includes journalists who covered a crime and lawyers who defended someone.

Who might be liable

Unsettled, and the candidates are these.

The developer, as the one whose system produced the words. Defences argued so far include that a conversational output with disclaimers is not a statement of fact, that there is no publication in the traditional sense, and that no intention or negligence exists in the required form.

The user who republishes it. This is the clearest case in most legal systems, and the one that should concern you personally. Take a false, damaging statement from a model, put it in an email or a post, and you have published it. Every ordinary defamation principle applies, and "an AI wrote it" addresses neither truth nor responsibility.

The deployer, where a business puts the output in front of customers — the Air Canada reasoning from the first module.

Intermediary protections that shield platforms from liability for user content were written for hosting other people's speech, and whether they cover speech a system generated is contested.

If it happens to you

Capture it properly. Full screenshots including the prompt, the complete answer, the date, the model and version. Ask again in a fresh conversation to see whether it reproduces, and capture that too. Do not rely on the conversation still being there.

Use the data protection route where available. In the EU, India, the UK and other regimes with rectification rights, a request that inaccurate personal data be corrected is faster and cheaper than litigation, and does not require you to prove damage. Address it to the provider's data protection contact.

Use the provider's process. Most have a mechanism for reporting factual errors about individuals, and they do act on the ones that reach them.

Take advice before publicising it. Amplifying the false statement in order to complain about it is a known way to make the harm worse.

If you are the user

One rule covers it. Never repeat a model's factual claim about a named living person without independent verification. Not in a message, not in a post, not in a report, not in a meeting. The cost of checking is two minutes and the cost of being wrong is somebody's reputation and possibly your own liability.