AI defamation: your legal options when AI lies

AI defamation is a false statement of fact about an identifiable person, produced by an AI system and shown to other people, that damages that person's reputation. Whether it is legally actionable depends on five questions: who saw it, who is the right defendant, whether it states a fact, how much harm it caused, and which remedy you want. There is no single law for it. You choose between defamation law, data protection law and, in the United States, a platform immunity statute that may decide the case before it starts.

This guide is a map of those routes, built from the statutes and regulator pages named in the text, as read on the day this was written. It does not predict how a court will treat a generated answer. We read no court judgment for this guide, so we describe no lawsuit and no outcome, and where the sources do not settle a point we say so in plain words rather than guess. Nothing here is legal advice. A lawyer in the right country decides which route fits your facts, and some deadlines may already be running.

The practical steps that do not need a lawyer, such as reporting a response to a provider, fixing the web page an assistant read, and diagnosing why it was wrong, are covered in separate guides on this site and are not repeated here. This one starts where those end: the answer is false, it is damaging, it has been seen by others, and you want to know what the law can and cannot do.

Related readAI Answers About Job Candidates: Audit and Fix

One monththe usual time a controller has to answer a data protection request

Two monthsthe further extension allowed for complex or numerous requests

Serious harmthe threshold section 1 of the Defamation Act 2013 sets for a statement to be defamatory

Sources: Article 12(3) of the GDPR on EUR-Lex, and section 1 of the Defamation Act 2013 on legislation.gov.uk.

What counts as AI defamation, and what does not

A claim has to be about you, false, stated as fact and communicated to someone else. Most disappointing AI answers fail at least one of those, so sort yours first.

An invented fact that names you is the clearest case: a disqualification, a conviction, a bankruptcy that never happened. An outdated fact that is now false can qualify too. So can a merged identity, where the assistant attaches another person's record to your name: the words are about you as far as a reader can tell.

Three things are usually not defamation. A true but unflattering statement fails the truth test. A statement of opinion is treated differently from a statement of fact: the Defamation Act 2013 has an honest opinion defence, and the UK data protection regulator says an opinion is hard to call inaccurate. And a hedged sentence, such as an assistant saying it found no reliable information, asserts nothing that damages you.

Some assistants show a notice that answers may contain mistakes. Whether such a notice changes the legal position is a question for counsel. We found no primary source that settles it, so no one should tell you it either protects the provider or does not.

Five questions decide whether you have a case

Lawyers tend to work through the same five points in the same order. You can do a first pass yourself.

1. Publication. In general terms, defamation needs a statement communicated to at least one person other than you. A screenshot of your own private query shows one reader: you. Evidence that a client, employer or investor saw it changes the analysis.
2. Defendant. Candidates are the AI provider, a company that embeds the assistant in its own product, a person who repeated the answer, and the owner of the web page that contained the original falsehood. Each has different legal protection. The author of the source page is often the most straightforward defendant.
3. Fact or opinion. The statement must assert something checkable. "She was struck off" is checkable. "She seems unreliable" is not.
4. Harm. Systems differ on how much harm is needed. England and Wales sets a statutory threshold. Other systems ask different questions, and we did not read their sources for this guide.
5. Remedy. Decide what you actually want: the statement gone from that product, a correction, a retraction, damages, or an order against someone. The route that gets you the first rarely gets you the last.

England and Wales: the Defamation Act 2013 applied to a generated answer

The Defamation Act 2013 sets the tests below. Its extent across the nations of the UK was not checked for this guide, so a Scottish or Northern Irish reader should confirm with a local lawyer which rules apply.

Related readAI answers about doctors and lawyers: what to fix

Serious harm. Section 1 says a statement is not defamatory unless its publication has caused or is likely to cause serious harm to the reputation of the claimant. For a body that trades for profit, harm to reputation is not serious harm unless it has caused or is likely to cause serious financial loss. That second rule matters to companies, firms and sole traders trading through a business: the file needs financial evidence, not only hurt feelings.

Truth. Section 2 provides a defence where the imputation conveyed is substantially true. If an assistant exaggerated a real event, the question is whether the core of what it said is substantially true, and you may lose that argument.

Honest opinion and public interest. Section 3 requires that the statement be one of opinion, that it indicate its basis, and that an honest person could have held the opinion on that basis; the defence is defeated if the claimant shows the defendant did not hold the opinion. Section 4 requires that the statement be on a matter of public interest and that the defendant reasonably believed publishing it was in the public interest. Both defences are written around a human's state of mind. How they apply to machine output is not something we found settled in a primary source, so we make no prediction.

Website operators. Section 5 gives a website operator a defence for statements posted by others, which the claimant can defeat by showing three things together: they could not identify the poster, they gave the operator a notice of complaint, and the operator failed to respond as the regulations require. The defence is also defeated by malice. It is written for third-party posts, and we read no authority applying it to a generated answer. It is still a lever when the false claim began on a forum or review site.

Related readWrong company information in ChatGPT: how to fix

Who can be sued. Section 10 says a court has no jurisdiction to hear a defamation action against someone who is not the author, editor or publisher unless it is satisfied that it is not reasonably practicable to bring the action against those people. Whether an AI provider is an author or publisher of a generated answer is open on the sources we read.

Single publication. Section 8 treats a cause of action over a later publication of substantially the same statement as accruing on the date of first publication, and the rule does not apply if the manner of the later publication is materially different. A generated answer is produced fresh each time it is asked, and whether that is one publication or many is not resolved by the sources we read. It bears on time limits. We did not read the limitation period itself, so ask about it in your first conversation with a lawyer.

Remedies. Section 13 lets a court that has found for a claimant order the operator of a website on which the statement is posted to remove it. That order is aimed at websites; the section we read does not describe an order directed at an AI product's output.

United States: Section 230 is the first question, not the last

In the United States, defamation claims are generally brought under state law, and we did not read a state statute or opinion for this guide. What we did read is 47 U.S.C. 230, on the Cornell Legal Information Institute site, because it often decides whether a claim against a platform can proceed at all.

Related readGoogle AI Mode reputation: audit what it says

Section 230(c)(1) says no provider or user of an interactive computer service is to be treated as the publisher or speaker of information provided by another information content provider. Section 230(f)(3) defines an information content provider as any person or entity responsible, in whole or in part, for the creation or development of information provided through the internet or another interactive computer service.

Those two definitions frame the argument. If an assistant's output is information provided by someone else, such as text lifted from a web page, the protection is at least arguable. If the system itself is responsible in whole or in part for creating or developing the sentence, it may be an information content provider as to that sentence, and the protection is harder to claim. We found no court opinion in this session that decides how the definition applies to generated text, so we do not state an answer. Anyone who tells you the question is settled, in either direction, has not shown you a source.

Two further points from the statute. Section 230(e)(3) says a state may enforce a state law consistent with the section, and that no cause of action may be brought and no liability imposed under state or local law that is inconsistent with it. And the exceptions in 230(e) cover federal criminal law, intellectual property, communications privacy law and certain sex trafficking claims. None is a defamation exception.

Two practical consequences follow for a US reader. The author of an original false page is the information content provider, not someone protected by 230(c)(1) as to that page, so a claim against that person can reach the source. And because the legal question is open and slow, the provider's own rights and legal-report routes, described elsewhere on this site, are usually the first move.

Related readWhat Siri, Alexa and Google Assistant say about you

Data protection: a route that does not need defamation

If you are in the EU or the UK, there is a second legal route that asks a different question. Defamation asks whether a statement hurt your reputation. Data protection asks whether personal data about you is inaccurate. The text we read is in the consolidated GDPR on EUR-Lex.

Accuracy and rectification. Article 5(1)(d) says personal data must be accurate and, where necessary, kept up to date, and that every reasonable step must be taken to ensure inaccurate data are removed or rectified without delay. Article 16 gives the data subject the right to obtain rectification of inaccurate personal data without undue delay, and the right to have incomplete data completed, including by a supplementary statement. That second right is a useful form of words for a person whose problem is a missing context rather than an invention.

Erasure. Article 17(1) lists grounds on which you can ask for erasure, including that the data are no longer necessary for the purposes for which they were processed (17(1)(a)) and that they have been unlawfully processed (17(1)(d)). Article 17(3)(a) says the right does not apply where processing is necessary for exercising the right of freedom of expression and information. That carve-out is why an accurate, lawful, public-interest report about you is not removed by asking.

Time. Article 12(3) requires the controller to act without undue delay and within one month of receipt, extendable by two further months where necessary given complexity and number of requests, with notice of the extension and reasons within the first month. Article 12(5) makes responses free of charge, with exceptions for manifestly unfounded or excessive requests, where the controller bears the burden of showing that.

Related readWhat Microsoft Copilot says about you: check and correct

What the UK regulator says. The ICO's right to rectification guidance says data is inaccurate if it is incorrect or misleading as to any matter of fact, that opinions are harder to treat as inaccurate where the record shows clearly that it is an opinion, and that the response period is one month. The page says the guidance is under review because of the Data (Use and Access) Act, so check it before you rely on it.

If the answer is no. Article 77 lets you complain to a supervisory authority, including in the Member State of your habitual residence. Article 79 gives a right to a judicial remedy, with proceedings brought where the controller has an establishment or where you habitually reside. Article 82(1) gives a person who has suffered material or non-material damage from an infringement a right to compensation from the controller or processor.

What we could not source is the key premise: whether a given provider is a controller for the text its model generates, and how regulators treat outputs. We read no regulator guidance on that, so read this section as the text of the rights, which is firm, and test their application to AI outputs with a specialist.

Choosing a route: a decision table

Match your facts to the route, and note the limit before you spend money.

Start with the cheapest route that can deliver what you actually want, and do not skip one that is a precondition for another: a rights request left unanswered past the one-month period is the natural basis for a regulator complaint. Never choose a route because it sounds powerful.

Related readChatGPT wrong information about my product: how to fix

Build the evidence file before anyone sends a letter

A lawyer's first questions are evidentiary.

1. Capture the answer in fullSave the full text, the exact question, the product and version, the date and time, and any sources the product cited. Repeat over several days, since output changes.
2. Show who else saw itCollect written statements from anyone who saw the answer or told you about it: what they saw, when, on which product.
3. Prove the truthGather documents that show the real position: registers, certificates, letters, dated records.
4. Document the harmKeep dated records of lost work, withdrawn offers, cancelled meetings and messages that mention the claim. For a business, add the revenue figures.
5. Trace the sourceOpen every cited link and search for the same sentence elsewhere. Note which pages say it, who runs them and when they were published or changed.
6. Log what you have already askedKeep copies of every report and rights request, with dates, ticket numbers and replies. Silence past the response period is part of the record.

Never crop out a date or present new answers as old ones: credibility is most of what a letter has.

A worked scenario: one false claim, three routes

This is an invented scenario with generic parties, to show the order of work. It is not a real case.

A chartered surveyor trading as a small firm in England learns from a prospective client that an AI assistant told the client she had been struck off. She was never disciplined. The client withdrew the instruction, and two other enquiries have gone quiet.

Assumptions: the firm trades for profit, the prospective client will write down what they saw, the assistant cited no source, and the surveyor has her professional body's register entry showing she is in good standing.

1. Day 1. She captures the answer in several phrasings, asks the client for a short dated statement, and prints the register entry. She finds the claim nowhere else, which points to a generated error rather than a copied one.
2. Days 2 to 3. She sends the provider a rights request that quotes the sentence, states it is false, attaches the register entry, and asks for rectification of the inaccurate personal data. She notes the date, because the one-month period in Article 12(3) runs from receipt, and the UK regulator's guidance uses the same period.
3. Week 2. She lists her losses with dates. Because she trades for profit, section 1(2) means the harm must be serious financial loss, so she needs figures: the withdrawn instruction and the quiet enquiries, each tied to a date.
4. Week 4 to 5. If the provider has not answered or refuses, she has a clear paper trail. She instructs a solicitor with the file, who decides whether to send a letter before claim and whether to complain to the regulator. The solicitor tests the open points above: publisher status, single publication and limitation.

The scenario ends at the letter on purpose: what follows depends on facts and law that no source we read settles.

The windows are experience ranges, not promises, apart from the GDPR periods. We read no court source for court timing.

Routes that fail, and risks you should price in

Legal action over a false AI answer fails in predictable ways.

The answer is true or substantially true. Section 2 of the Defamation Act 2013 protects a substantially true statement, and Article 17(3)(a) protects expression and information. An accurate, lawful, public-interest account of something you did is not removable by any route in this guide, and we do not help try. The honest response to a true report is context, not a complaint.

Only you saw it. A claim built on one private screenshot has no audience.

Related readAI due diligence on executives: what to trust

Harm is thin. Under section 1, discomfort is not serious harm, and a trading business needs serious financial loss.

The wrong defendant. A claim against a provider whose status as publisher is contested, while the real source is a forum post by an identifiable person, spends money on the hardest target. Section 10 of the 2013 Act is one example of why defendant choice matters.

A second story. Public litigation, or even a loud letter, can create new coverage of the false claim, and new coverage becomes new text that assistants read. This is the Streisand effect. A private letter and a quiet rights request avoid it. A press release does not.

Intimidation and delay. A letter that threatens a complainant or a reviewer with a claim you do not intend to bring can damage your own position. And limitation periods, with the single publication rule in section 8, mean time can run from first publication, which is why the first lawyer call should ask about it.

A legal route does not replace source correction and monitoring. A court order against one defendant leaves other assistants and other copies of the claim untouched, so the pages those systems read still need accurate, current material. Our AI search reputation page describes how that work fits alongside a legal file.

What to do before you instruct anyone

Reputation Partners is not a law firm and does not give legal advice. Our part is the evidence and the sources: what AI assistants say, which pages they read, and what accurate material exists. A solicitor or attorney decides the legal route.

Questions people ask

Can I sue an AI company for defamation?

In principle a false, damaging statement of fact can be actionable, but the sources we read do not settle how courts treat generated answers. The defendant, the publisher question, the harm threshold and, in the United States, Section 230 all matter. Take the evidence file to a lawyer in your country before deciding.

Does an AI warning label protect the provider?

We found no primary source that settles it. Some assistants show a notice that answers may contain mistakes. How much weight such a notice carries, set against how the answer was presented and who saw it, is a question for counsel in your country. Do not rely on anyone who says it settles the matter either way.

Can I claim compensation under data protection law?

Article 82(1) of the GDPR gives a person who has suffered material or non-material damage as a result of an infringement a right to compensation from the controller or processor. The damage has to result from an infringement, so a rights request alone does not create a claim. Whether an infringement occurred, who the controller is for a generated answer, and what amount a court would award are for specialist advice.

What if the answer is true but unflattering?

Then defamation law offers little. Section 2 of the 2013 Act protects substantially true statements, and Article 17(3)(a) of the GDPR excludes processing needed for freedom of expression and information. The useful response is context: your own account, later developments and stronger accurate sources, never an attempt to hide a lawful, accurate public record.