Liability for AI-Generated Content: What Regulators and Courts Have Actually Said

If a draft came from an AI tool and later proved false, misleading or copied, uncertainty about responsibility is normal. The authorities reviewed here have not created one global rule for AI-generated content. They have applied existing advertising, misrepresentation, professional-conduct and copyright principles to particular facts, while leaving major questions unresolved.

There is no single rule for AI-generated content liability

The legal result depends on what was published, where it appeared, who relied on it and which jurisdiction's law applies. An inaccurate product claim, a chatbot's statement about a refund, a fabricated court citation and an image resembling protected artwork raise different legal questions.

The table below is Need to Know AI's synthesis of the authority boundary. It records what each source establishes and, equally importantly, what it does not establish.

US regulators apply ordinary advertising rules to AI-assisted claims

The FTC's published position is that the use of AI does not create an exemption from existing consumer-protection law. In a joint statement with other US agencies, the FTC said it would use its existing authority against unfair or deceptive practices involving automated systems. The statement does not say that every erroneous output produces liability or identify one party as responsible in every AI supply chain. See the FTC's joint agency statement on AI.

For advertising, the FTC describes the underlying test without making it AI-specific. Its small-business guidance says advertisements are assessed for truthful and non-deceptive claims, supporting evidence, materiality and the overall impression conveyed to a reasonable consumer. That means an AI-written sentence is not evaluated under a separate truth standard merely because software drafted it. See the FTC's Advertising FAQs for small businesses.

This authority supports a limited conclusion: the origin of advertising copy does not displace the FTC's ordinary framework. It does not establish the outcome of a private lawsuit, determine responsibility between a business and its software supplier, or resolve the effect of particular contractual terms.

UK advertising guidance places the final advertisement in the advertiser's sphere

The UK's ASA and CAP say that advertising rules apply regardless of how an advertisement was created. Their June 2026 guidance points to the Stripe & Stare ruling and says the advertiser remained primarily responsible for compliance even though Google's automated tools produced the advertisement. See ASA and CAP guidance on AI and deepfakes.

That statement is unusually direct, but its scope matters. The ASA administers the UK advertising codes. Its guidance does not adjudicate damages for defamation, copyright infringement, negligent advice or breach of contract, and it does not determine how responsibility would be divided between an advertiser, agency, platform and AI provider in civil litigation.

The ASA has also explained that disclosing AI use would not cure an advertisement that is misleading in substance. Its guidance says a misleading claim cannot be rescued simply by identifying AI as the method used to make it. See ASA and CAP guidance on disclosure of AI in advertising.

The EU AI Act addresses transparency, not every form of liability

Article 50 of the EU AI Act creates specific transparency rules rather than a universal liability code for published output. Among other provisions, it covers disclosure of deepfakes and AI-generated or manipulated text published to inform the public about matters of public interest. The text includes a defined exception for that public-interest text where it has undergone human review or editorial control and a natural or legal person holds editorial responsibility. See Article 50 in Regulation (EU) 2024/1689.

The European Commission says Article 50's transparency obligations started applying on 2 August 2026. It also explains that a limited transition applies to the machine-readable marking obligation for certain systems placed on the market before that date. See the Commission's Article 50 questions and answers.

Editorial responsibility in Article 50 has a specific function within that provision. It would overstate the source to treat the phrase as a general EU judgment about who pays damages whenever reviewed content is false, defamatory or infringing. Other EU and national laws may apply, but the facts and relevant cause of action would matter.

One Canadian tribunal rejected the idea that a website chatbot stood alone

Moffatt v Air Canada is a narrow but concrete decision about inaccurate information delivered through a business's own website. A customer relied on a chatbot's statement about applying for a bereavement fare after travel. The British Columbia Civil Resolution Tribunal, deciding on 14 February 2024, found that Air Canada had not taken reasonable care to make the representation accurate. On the airline's argument that it was not answerable for the chatbot, the tribunal member wrote: “In effect, Air Canada suggests the chatbot is a separate legal entity that is responsible for its own actions. This is a remarkable submission. While a chatbot has an interactive component, it is still just a part of Air Canada's website. It should be obvious to Air Canada that it is responsible for all the information on its website.” See the official decision in Moffatt v Air Canada, 2024 BCCRT 149.

The tribunal treated the chatbot as part of the airline's website and ordered Air Canada to pay a total of $812.02, comprising $650.88 in damages, $36.14 in pre-judgment interest and $125 in tribunal fees. This was a provincial tribunal decision based on a particular commercial relationship and a negligent-misrepresentation claim. It was not a higher-court ruling, a class action or a general judgment about generative AI products.

The careful reading is therefore limited: in that dispute, the business did not avoid responsibility by separating its chatbot from the rest of its website. Extending that result to another jurisdiction, a different claim or a third-party publishing platform would require analysis beyond what the decision itself supplies.

Courts have treated fabricated AI citations as the filer's work product

US federal courts have imposed consequences when professionals submitted fabricated AI-generated authorities without adequate verification. In Mata v Avianca, lawyers filed nonexistent cases and false quotations generated by ChatGPT. The court said using a reliable AI tool for assistance was not inherently improper, but existing procedural rules gave lawyers a gatekeeping role over the accuracy of their filings.

The court imposed a US$5,000 penalty jointly on the lawyers and their firm, alongside notification requirements. The sanction concerned the lawyers' conduct, including continuing to stand by false material after its authenticity was questioned, rather than a general ruling that every user is liable for every model output. Read the court's opinion and order on sanctions.

For regulated professions, this case illustrates a source-backed boundary rather than a universal rule. Professional and procedural duties can continue to apply when AI assists with a document, but Mata does not decide the obligations of marketers, retailers or other businesses publishing outside court proceedings.

Copyright ownership and copyright infringement are different questions

A business can face uncertainty on both sides of copyright: whether it owns an output and whether that output copies protected expression. These questions are related, but an answer about one does not resolve the other.

In the United States, the Copyright Office's January 2025 report says material generated wholly by AI is not copyrightable. It says protection can cover human-authored expression, creative selection or arrangement, or sufficiently creative human modifications, while prompts alone do not ordinarily provide the required control over expressive elements. See the US Copyright Office report on copyrightability.

The US Court of Appeals for the District of Columbia Circuit subsequently held that the Copyright Act requires human authorship, while stressing that this does not prevent protection for work made with AI assistance. The Supreme Court declined to review the case in March 2026, leaving the appellate decision in place without issuing its own opinion on the merits. See the D.C. Circuit decision in Thaler v Perlmutter and the Supreme Court docket.

The infringement question remains less settled. The Copyright Office's pre-publication Part 3 report says models sometimes produce material that replicates or closely resembles protected works and that such outputs are likely to implicate reproduction or adaptation rights. The same report expressly leaves enforcement questions and allocation of potential liability for later treatment. See the US Copyright Office report on generative AI training.

In the United Kingdom, section 9(3) of the Copyright, Designs and Patents Act 1988 currently names the person making the necessary arrangements as the author of a computer-generated literary, dramatic, musical or artistic work. See section 9 of the Copyright, Designs and Patents Act.

The UK government's March 2026 report says the interaction between that provision and the modern originality test remains unclear because of limited case law. The government proposed removing the special protection for wholly computer-generated works while retaining ordinary protection for human creativity assisted by AI. That was a policy proposal, not a completed statutory repeal as of this review. The same report also withdrew the broad training exception with an opt-out as the government's preferred approach, which is a separate issue from ownership of outputs. See the UK Report on Copyright and Artificial Intelligence.

Defamation and other harmful statements remain unresolved at this level

The sources reviewed do not support a single answer about responsibility for an AI-generated statement that damages someone's reputation. Defamation tests, publication rules, available defences and fault standards vary by jurisdiction. A chatbot response sent to one person may also raise different publication questions from a statement placed in an advertisement or public article.

The UK government's 2026 copyright report describes defamation as one part of a patchwork that can apply to digital replicas, but it does not allocate responsibility for ordinary hallucinated text. The consumer and advertising authorities above address their own remits, not every civil cause of action.

This gap is substantive. It is not appropriate to convert an advertising regulator's statement or the Moffatt tribunal decision into a global defamation rule. A particular allegation requires advice based on the content, audience, jurisdiction, evidence of harm and parties involved.

Where the cited guidance stops

None of these authorities determines liability for a business merely from the fact that AI contributed to its content. The sources establish narrower points about advertising claims, an airline chatbot, court filings, transparency and copyright authorship.

They do not collectively answer:

• whether a tool developer, deployer, employee, agency, publisher or platform bears liability in a particular dispute;
• whether contractual warranties or indemnities shift costs between the parties;
• whether a disclaimer changes the outcome under the relevant law;
• whether an output is substantially similar to protected expression;
• whether a false statement satisfies a jurisdiction's test for defamation; or
• whether human review was adequate on the facts.

Those questions require the governing law and evidence. Some may eventually be answered by legislation or appellate decisions, but treating them as settled now would go beyond the primary sources reviewed for this page.

Questions that expose publishing risk without deciding liability

A pre-publication review can be organised around questions rather than unsupported legal conclusions. This is an editorial-control framework from Need to Know AI, not a regulator's checklist and not a route to certified compliance.

These records cannot determine a legal outcome. They can make it easier for a business and its adviser to reconstruct what was generated, what was checked and who approved publication.

Human review is an operational control, not a universal safe harbour

No authority reviewed for this page describes ordinary human review as a blanket defence to misleading, infringing or harmful content. The EU AI Act gives review and editorial responsibility a specific role in one Article 50 disclosure exception for public-interest text. That limited provision does not turn review into immunity from other laws.

The publishing process can nevertheless separate high-risk material from routine editing. Claims about price, performance, health, safety, legal rights or named people present different consequences from changes to tone or grammar. A review record can distinguish source checking from a simple approval click.

The same distinction applies to plagiarism and copyright screening. A detector score does not decide infringement, and the Copyright Office has not identified an automated test that allocates liability. A documented comparison against known source material is an operational check, while any legal conclusion remains a matter for the relevant rights holder, court or adviser.

For Australian businesses: the ACCC's position on published claims

The Australian Competition and Consumer Commission says a business's claims about products or services should be accurate, truthful and based on reasonable grounds, regardless of intent. Its guidance covers advertising, quotations, staff statements, social media, testimonials and websites. It does not create an AI-specific liability test or decide a particular dispute. See the ACCC's guidance on false or misleading claims.

The ACCC also says businesses must not create fake or misleading reviews or arrange for others to create them. Its review guidance focuses on whether a review is genuine and reflects a real experience, not whether software or a person drafted the words. See the ACCC's online-review guidance.

These statements account for the consumer-claim and verification subjects covered by the retired Australian pages. They do not establish who would be liable for every AI hallucination, copyright dispute or defamatory statement.

Related reading: free AI acceptable use policy template, free AI register template, shadow AI audit checklist, guide to assessing AI risk, AI data residency comparison, AI vendor due diligence checklist, and AI vendor breach response plan template.

Related reading: Claude AI Review: Pricing, Features, and Business Verdict and Is Claude Pro Worth It? An Honest Assessment for Business Users.

Try our free AI Compliance Checker to check whether your AI tools meet your compliance obligations.

Methodology and evidence limits

This desk-research assessment was checked against primary regulator, legislative, government, tribunal and court materials on 4 September 2026. The review covered consumer and advertising statements from the FTC, ASA and CAP, and ACCC; Article 50 of the EU AI Act and Commission guidance; Moffatt v Air Canada; Mata v Avianca; Thaler v Perlmutter; the US Copyright Office's AI reports; and the UK government's 2026 copyright report.

Need to Know AI did not treat vendor terms, law-firm commentary or another NTKAI article as authority for a legal proposition. The comparison matrix is an editorial synthesis of the cited sources, not a statement that the jurisdictions use identical tests. Open questions are recorded where the reviewed authorities did not provide a reliable answer.