Hallucinations by large language models raise a simple question: who answers when an AI wrongly attributes an offence to an identified person? Walters v. OpenAI sets out the US answer; French press law strikes a very different balance.
I need helpUnder French law, defamation requires the public allegation of a specific fact harming the honour of an identifiable person. An erroneous AI-generated answer can therefore, in theory, meet that definition, provided it was made public and not merely displayed to a single user. The victim does not have to prove the publisher's bad faith, unlike the US actual-malice standard, but must act within three months of publication.
The rise of large language models (LLMs) has opened up an unprecedented field of uses… and risks. Among these risks, “hallucinations” - those responses that invent facts - raise a cardinal question: who is legally responsible when a tool like ChatGPT wrongly attributes delinquent behavior to an identified person?
A decision rendered in May 2025 by the Superior Court of Gwinnett County (Georgia) in the case of Mark Walters v. OpenAI, L.L.C. \\ [1.\")\\] provides concrete insight from the American side: the court granted “summary judgment” to OpenAI, meaning it ended the litigation in its favor, without trial, based on the evidence and applicable law.
We propose here a structured analysis of this judgment, followed by a comparison with French defamation law (Law of July 29, 1881), whose balances are quite different.
On May 3, 2023, Frederick Riehl, editor-in-chief of Ammoland (a news/advocacy website on the Second Amendment), received a press release and the complaint from the Second Amendment Foundation (SAF) against the Attorney General of Washington State (SAF v. Ferguson). Wishing to write an article, he asked ChatGPT to summarize passages from the complaint; for the pasted sections, the summary was accurate.
At other times, the tool provides inaccurate responses, a phenomenon that technical literature now refers to as “hallucinations.”
The judge expressly notes that, by their generative nature, major LLMs can produce information contradicting the source, and recalls this technical vocabulary.
Crucial point: Riehl was aware of these limitations (he had already observed “flat-out fictional” responses) and, on May 3, he had accepted ChatGPT’s terms of use, which warned him: probabilistic output, potentially inaccurate results, necessity of human review. He also saw several on-screen warnings stating that “ChatGPT may produce inaccurate information.”
The person pointed out by these erroneous responses, Mark Walters, is a well-known media figure in the “Second Amendment” sphere: daily radio host, author, public personality within pro-gun organizations. These elements form the basis of the court’s analysis of his “public figure” status (at least limited).
The court grants summary judgment to OpenAI; in other words, no material question of fact warranted a trial.
The “reasonable reader” filter is decisive: Riehl was not a layperson; he knew ChatGPT’s limitations and had explicit warnings before his eyes, urging him to verify responses in case of editorial use.
In this context, the Court considers that the incriminating outputs did not, in law, convey a defamatory meaning likely to engage OpenAI’s liability: for a reasonable user in these circumstances, the output had limited and conditional informational value, requiring verification. The materiality of these warnings is documented in the file (Terms of Use, interface messages, etc.).
Even assuming potential reputational harm, the action fails on two levels: (i) negligence (ordinary standard of a “reasonable publisher” under Georgia law) and \\*\\*(ii) the aggravated degree of actual malice required for a public figure.
Regarding negligence, the Court finds that OpenAI deployed reasonable organizational and technical measures: design and training aimed at reducing hallucinations, clear and repeated warnings for the user, recommendation of human review before editorial use. These elements defuse the allegation of a failure of due diligence in the provision of the tool and its user interface.
Regarding \"actual malice\", the plaintiff had to prove -with “clear and convincing” evidence- that the publisher knew the assertions were false or acted with deliberate disregard for the truth. The Court, on the contrary, notes the absence of evidence that anyone at OpenAI was aware of a specific hallucination concerning Mr. Walters, or intended to ignore identified risks for this particular case. The record rather shows proactive steps to recall uncertainty and frame sensitive uses.
The US decision creates no immunity for generative AI: it merely finds no qualified fault in the context of use examined.
AI defamation: US law and French law
Legal basis
Burden of fault
Role of warnings
Limitation period
Defences
| Criterion | US law | French law |
|---|---|---|
| Legal basis | Common law defamation, filtered through the First Amendment | Article 29 of the Act of 29 July 1881 |
| Burden of fault | Negligence, and actual malice for a public figure | Mental element presumed once the statement is published |
| Role of warnings | Decisive in assessing the meaning perceived by a reasonable reader | Considered as context, with no automatic exemption |
| Limitation period | Generally one to three years depending on the State | Three months from publication |
| Defences | Claimant must prove qualified fault | Truth defence or good faith, to be established by the defendant |
In French law, defamation is defined by Article 29, para. 1 of the Law of July 29, 1881: “any allegation or imputation of a fact that harms the honor or reputation” of a person (physical or legal). The incriminated fact must be precise, verifiable, and the identification of the person possible.
Doctrine and jurisprudence emphasize the precise nature of the imputed fact: it must have content capable of proof and adversarial debate. Failing this (vague statements, opinions), the qualification shifts to insult. Courts assess the meaning of statements taking into account the intrinsic and extrinsic elements of the message (context, references, associated media).
A major difference: in France the victim need not prove the author's bad faith, it is for the author to establish good faith or the truth of the facts.
Defamation is an intentional offense. However, the “intent to defame” in the subjective sense -the desire to harm- is not a constitutive element that the victim would have to prove: the moral element results from the will to utter the incriminating statements; it is not required to demonstrate “malice” in the American sense. The burden shifts, after the offense is constituted, to means of exoneration available to the author.
Practical consequence: unlike the American standard of actual malice, the victim does not have to prove that the author knew it was false or was indifferent to the truth. The mere act of publicly alleging a precise fact harming honor is sufficient, it then being up to the defendant/accused to invoke their justifications.
Where American law sometimes conditions the very existence of liability on the victim’s proof of actual malice, French law constitutes the offense more easily, then opens avenues of exoneration for the author.
Ultimately, the Walters v. OpenAI judgment creates neither general immunity for generative AI, nor an exceptional regime: the court simply recalls that, for a public figure, defamation is not characterized if, when placed in its context of use, the output of an LLM does not convey a defamatory meaning for a reasonable reader, that no fault -and a fortiori no actual malice- is demonstrated, and that no compensable damage is established.
The visible warnings and the design oriented towards human verification weighed heavily in this assessment, as did the user’s profile and the absence of evidence of an internal knowledge of falsity.
Facing a campaign of fake reviews?
Our lawyers audit the disputed reviews, preserve the evidence and handle the removal and compensation proceedings. Every enquiry is answered within 24 hours.
Contact our lawyersResponding to a defamatory AI answer
01Preserve evidence of the answer
Have the model's output formally recorded by a judicial commissioner, with the prompt used, the date, the timestamp and any share link: a generative answer cannot be reproduced identically.
02Check whether the allegation is public
Establish whether the statement travelled beyond a private exchange, through a shared conversation link, a republished screenshot or editorial coverage: publicity governs the characterisation as public defamation.
03Characterise the facts
Distinguish the allegation of a specific fact, which is defamation, from a vague judgement, which is insult, and consider alternative grounds based on the data-accuracy principle.
04Notify the model's publisher
Send a reasoned formal notice seeking correction, blocking of the answer and, where relevant, rectification of the data concerning the person targeted.
05Act within three months
Bring civil or criminal proceedings before the short limitation period applicable to press offences expires, failing which the claim is inadmissible.
Deadlines to remember
3 months
Limitation period for defamation from publication
EUR 12,000
Maximum fine for public defamation of a private individual
1881
The 29 July Act governing press offences
24 h
INFLUXIO response time to any enquiry
Our lawyers handle disputes involving AI-generated content, from formally recording the disputed answer to claiming compensation. See also our artificial intelligence lawyer and defamation lawyer pages. Contact the firm: we answer every enquiry within 24 hours.
An AI is not a legal person and cannot be sued. The question is the liability of the model's publisher and, where relevant, of whoever disseminates the answer. Under French law, publicly circulating an answer alleging a specific fact harmful to someone's honour can amount to defamation.
In principle no. Public defamation requires dissemination to an indeterminate audience. An answer read by a single user will rather fall under non-public defamation, punished as a minor offence. Everything changes once the conversation is shared through a public link, republished or picked up in an article.
They weighed heavily in the US decision under review, where the judge held that an informed user could not reasonably read a defamatory meaning into the answer. Under French law those warnings form part of the context assessed by the judge, but they are not an automatic defence.
No. Under French law the mental element flows from the intention to publish the statement; the victim need not show malice in the US sense. It is then for the defendant to raise the truth defence or good faith. That difference makes the claim markedly more accessible than in the United States.
Data protection offers a complementary route: the accuracy principle allows rectification or erasure of inaccurate information about an identified person. Depending on the case, invasion of privacy and ordinary civil liability may also apply.
Three months from first publication for a defamation claim, which makes immediate evidence-gathering essential. Claims based on data protection or civil liability follow longer periods, but the date of the evidence remains decisive.
Related publications
About the author
Partner
Admitted to the Paris Bar, Maître Raphaël MOLINA is a co-founding partner of INFLUXIO and has specialized in intellectual property law and digital law for several years.
Contact
Would you like to schedule a meeting or get a quote?
We respond within 24 hours.
