Who is responsible when an AI chatbot makes up false claims about a real person, and what can that person do?

What should I check, step by step?

1. Save the output the moment you see it: export or copy the full conversation, the question that produced it, the product and model shown, and the date and time. A screenshot alone is the weakest form because it's easy to edit.

2. Keep the conversation's share link if the product offers one, but don't rely on it alone: a share link can stop working if the conversation is deleted. Hash the exported file and get an independent timestamp on the hash, so you can show later the copy wasn't changed.
3. Record who else saw it. Defamation needs publication to someone other than you; in Walters the only reader was the journalist who prompted it, and that mattered. Customers, partners or a search results page that showed it to the public are a different situation.
4. Write down the harm as it happens: lost contracts, cancelled orders, messages from people who believed it. Walters conceded he had no actual damages, which was one of the three reasons he lost.
5. Ask the company to correct it, in writing, and keep the request and any reply. Georgia law required a correction or retraction request before seeking punitive damages, and the Starbuck cases put weight on repeated notices that went unfixed.
6. In the EU or EEA, you can file a data protection complaint under the GDPR's accuracy and rectification rights, as noyb did in Norway, alongside or instead of a defamation claim.
7. Check again over time and keep each result. Answers change between model versions, and a later model that uses web search may stop repeating the claim, so your saved copy may be the only record.
8. If you run an AI product that answers questions about people, keep a server-side record of each answer (request, model, output and a hash), so you can show what your product actually said when someone complains.
9. Talk to a lawyer before suing or publicly accusing the company. Defamation rules, public-figure status and deadlines differ by state and country.

Can you sue OpenAI if ChatGPT makes up false claims about you?

Journalist Frederick Riehl asked ChatGPT to summarize a real lawsuit. After he gave it a link it said it couldn't open, it produced a different, invented summary saying gun-rights activist Mark Walters had embezzled from the Second Amendment Foundation. Walters sued OpenAI for libel.

• Judge Tracie Cason granted OpenAI summary judgment for three independent reasons. First, a reasonable reader in Riehl's position wouldn't have taken the output as actual fact: ChatGPT had warned him it couldn't access the link, the terms of use and repeated disclaimers said output can be inaccurate, and Riehl knew the claims were false. Second, Walters couldn't show negligence or actual malice; the court pointed to OpenAI's efforts to reduce errors and its warnings to users. Third, Walters conceded he had no actual damages, hadn't requested a correction or retraction as Georgia law requires before punitive damages, and couldn't rely on presumed damages.
• So the ruling says little about a case where the false claim reached the public, the reader believed it, and real losses followed. Eugene Volokh, who wrote about the case, described the result as turning on the peculiar facts.

What AI defamation lawsuits have there been so far?

Volokh's September 14, 2026 count lists nine US libel-by-AI cases since 2023: Walters (summary judgment for OpenAI); Battle v. Microsoft (sent to arbitration and resolved there); Starbuck v. Meta (settled); Keene v. Google and Starbuck v. Google (motions to dismiss denied, so the cases go forward); LTL LED v. Google, the Wolf River Electric case (going forward after the federal court sent it back to state court); and Murray v. Alphabet and two Grybniak cases against Google and xAI (filed, no judicial action yet).

• Starbuck v. Meta alleged Meta's AI falsely said the activist took part in the January 6 Capitol riot, and that Meta didn't fix it after notice. It settled in August 2025, with Starbuck advising Meta on bias in its AI. Wolf River Electric, a Minnesota solar company, alleges Google's AI Overview said the state attorney general had sued it for deceptive practices when none of the cited sources said so, and that customers cancelled contracts. Unlike Walters, the output there reached the public alongside sources that made it look reliable. These allegations are unproven.
• In Europe, noyb's March 2025 complaint to Norway's Datatilsynet says ChatGPT told Arve Hjalmar Holmen he had murdered two of his sons and been sentenced to 21 years, mixing in real details such as his hometown and the number of his children. It asks the regulator to order the output deleted, the model adjusted and a fine imposed. noyb lists the case as pending.

How do I prove what an AI chatbot said about me?

Every one of these cases needs the same thing before any legal question: what exactly the AI said, in response to what, and when. The person harmed usually has only a screenshot or a share link, and the AI company's copy is under its control and may be kept only for a limited time.

• A copy that was hashed and timestamped when it was saved can be checked later by anyone: change one character and the hash no longer matches. That doesn't prove the model produced it, but it does show the copy hasn't changed since that time, which a bare screenshot can't.