AI Liability: 7 Ways Wrong AI Output Becomes Your Legal Problem

TL;DR

Seven mechanisms that move the consequence from the model to you - your chatbot's answers, your filings, your ad claims, claims about your own AI, republished output about others, EU AI Act disclosure, and platform rules. Not legal advice.

No rule anywhere treats a model as the author of its own mistakes. Seven mechanisms move the consequence onto the business that published the output: what your chatbot says, work you file or submit, claims in your advertising, claims about your own AI, output you republish about someone else, EU AI Act disclosure duties, and the platform rules you accepted. Each already has a decision, an order or a statute behind it.

This post is not legal advice. It is written against court and tribunal decisions read at source, US regulator releases, the AI Act as published in the Official Journal, the European Commission's current timeline, and Google Search Central's spam policies, all retrieved 2026-09-18. Unresolved matters are marked as allegations.

Key Takeaways

A chatbot is not a separate legal person. A tribunal called that argument "a remarkable submission" and held the airline liable for its chatbot's answer (Moffatt v. Air Canada, 2024 BCCRT 149, paras 27-28, retrieved 2026-09-18).

Advertising evidence must exist before publication. "The law requires that advertisers have proof before the ad runs" (FTC, Advertising FAQs, retrieved 2026-09-18).

Claims about your AI are the most actively enforced category: a US$400,000 SEC settlement over "AI washing", and five FTC actions at once under Operation AI Comply (SEC · FTC, retrieved 2026-09-18).

The AI Act's disclosure duty on AI-generated text has an editorial carve-out for content that "has undergone a process of human review or editorial control" with someone holding editorial responsibility (Regulation (EU) 2024/1689, Article 50(4), retrieved 2026-09-18).

The high-risk deadlines moved to 2 December 2027 for Annex III and 2 August 2028 for Annex I, after the "AI Omnibus" entered into force on 27 July 2026 (European Commission, retrieved 2026-09-18).

The seven, at a glance

#	Mechanism	Who enforces it	What triggers it
1	Your chatbot's answer is your representation	Courts and tribunals, on ordinary contract and tort law	A customer relies on a wrong answer and loses money
2	Work you file or submit carries your professional duty	Courts, professional regulators	An unchecked citation, quote or figure in a filing
3	Ad claims need evidence before they run	FTC, ASA, CMA and equivalents	Publishing a claim you cannot substantiate yet
4	Claims about your own AI	FTC, SEC and securities regulators	Saying the product does more than it does
5	Republishing output about a third party	Civil claimants	A wrong statement about a person or company
6	Disclosure duties under the EU AI Act	National market surveillance authorities	Interacting with, or publishing from, AI without disclosure
7	Platform rules you accepted	Search engines, ad platforms, marketplaces	Scaled or unhelpful AI output on your own property

The documented cases behind several of these, and what each cost, are in what one wrong fact actually costs. This post is the taxonomy: which mechanism applies to which kind of output.

1. Your chatbot speaks for you

Air Canada's support chatbot told a customer they could apply for a bereavement fare retroactively. The policy did not allow it, and the airline argued it should not be liable for information provided by one of its agents, servants or representatives — including a chatbot. The tribunal's answer: "In effect, Air Canada suggests the chatbot is a separate legal entity that is responsible for its own actions. This is a remarkable submission... It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot." It found the airline "did not take reasonable care to ensure its chatbot was accurate" and ordered CA$812.02 (Moffatt v. Air Canada, 2024 BCCRT 149, paras 27-28 and 44, retrieved 2026-09-18, SOURCED).

The doctrine was negligent misrepresentation, which "can arise when a seller does not exercise reasonable care to ensure its representations are accurate and not misleading" (same decision, para 24). Nothing in that test is AI-specific: the output is a statement by the business, not a product defect (ANALYSIS).

What changes the exposure: what your assistant says about prices, eligibility and deadlines has to match the page that states them. Where your surfaces disagree, the customer relies on whichever one they were shown.

2. Anything you file or submit carries your own duty

Courts check citations for a living, so the record is thickest here. In Mata v. Avianca a New York federal court imposed "a penalty of $5,000 paid into the Registry of the Court", held the firm "jointly and severally liable" with the two lawyers, and ordered letters to the client and to every judge falsely named as the author of a fabricated opinion (sanctions order, Doc. 54, retrieved 2026-09-18, SOURCED).

In England, two cases were heard together under the High Court's Hamid jurisdiction — "the court's inherent power to regulate its own procedures and to enforce duties that lawyers owe to the court". The judgment lists the available responses: "public admonition of the lawyer, the imposition of a costs order, the imposition of a wasted costs order, striking out a case, referral to a regulator, the initiation of contempt proceedings, and referral to the police." In the first case a judge had already ordered the barrister and the law centre "each to pay £2,000 to the defendant" and referred the matter to both legal regulators (Ayinde v Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin), paras 2, 23 and 49, retrieved 2026-09-18, SOURCED).

One detail belongs in any policy you write. The barrister denied using AI at all and said she was aware that artificial intelligence "is not a reliable source" (para 53). The court proceeded on negligence regardless of the tool, because the duty attaches to the citation, not to how it was produced (ANALYSIS).

What changes the exposure: every reference gets opened and read before the document is filed — the method is in how to trace an AI claim back to its original source.

3. Ad claims need the evidence in hand first

This reverses the order most drafting workflows assume. The FTC's small-business guidance is explicit: "The law requires that advertisers have proof before the ad runs." A company must have a "reasonable basis" for its claims, meaning "objective evidence that supports the claim", and "at a minimum, an advertiser must have the level of evidence that it says it has" (FTC, Advertising FAQs, retrieved 2026-09-18, SOURCED).

A model writes claims at the specificity a reader finds persuasive, not at the specificity your evidence supports. The sentence is generated; the substantiation is not.

The claim-by-claim detail sits in three companion posts: ads in 10 ad copy claims AI writes that you legally cannot publish, email in AI email copy that invents facts about your product, and listings in 9 AI product description mistakes.

4. Claims about your own AI

The most actively enforced category in the list, and the irony is exact: the wrong fact is about your AI's accuracy.

The SEC settled charges against two investment advisers "for making false and misleading statements about their purported use of artificial intelligence", with "$400,000 in total civil penalties", and its then-Chair named the practice "AI washing" (SEC press release 2024-36, retrieved 2026-09-18, SOURCED).

The FTC announced five actions at once under Operation AI Comply, against "operations that use AI hype or sell AI technology that can be used in deceptive and unfair ways" (FTC, September 2024, retrieved 2026-09-18, SOURCED). In the DoNotPay matter the final order required "$193,000 in monetary relief", notices to subscribers from 2021 to 2023, and no claims that the service performs like a real lawyer "unless it has sufficient evidence to back it up" (FTC, February 2025, retrieved 2026-09-18, SOURCED).

What changes the exposure: every accuracy number on your marketing pages needs the test behind it, named and dated. The buyer-side version is 8 accuracy questions to ask any AI vendor, and those questions cut both ways.

5. Republishing output about someone else

Here the law is least settled, so this section stays narrow.

The best-documented matter is a complaint, not a finding. A Minnesota solar installer alleges a Google AI Overview stated it was facing a lawsuit from the state attorney general, and that "None of the four sources Google cited support this claim". Google's answer is a general denial, and in January 2026 a federal court remanded the case to state court because the removal was untimely, without ruling on the merits (complaint, D. Minn. 25-cv-02394, Doc. 1-1 · remand order, Doc. 34, retrieved 2026-09-18, SOURCED as allegations). That matter concerns a platform's own summary; the business-side view is in AI Overviews accuracy.

For a publisher the question is ordinary law, not AI law: a false statement of fact about a named company or person in your post, comparison page or sales email carries the exposure it always did, whatever drafted the sentence (ANALYSIS). We found no decision resolving liability for republished AI-generated defamatory text, and will not imply one exists.

What changes the exposure: sentences naming a third party get verified like statistics, and comparative claims about competitors are the highest-risk sentences on any site.

6. Disclosure duties under the EU AI Act

Three duties in Article 50 reach ordinary content work, and one has a carve-out most summaries omit.

Chatbots. People must be told they are interacting with an AI system unless it is obvious, "in a clear and distinguishable manner at the latest at the time of the first interaction or exposure" (Article 50(1) and 50(5)).
Marking generated output. Providers of systems generating synthetic audio, image, video or text "shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated" (Article 50(2)) — a duty on the provider, not on you as a deployer.
Published text on matters of public interest. Deployers publishing AI-generated or manipulated text "with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated" — and it does not apply "where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content" (Article 50(4)).

All quoted from Regulation (EU) 2024/1689 as published in the Official Journal, retrieved 2026-09-18, SOURCED.

The dates are not the ones in the regulation you just read. Article 113 of the OJ text says the Regulation "shall apply from 2 August 2026", with Article 6(1) from 2 August 2027. The Commission's current page states that after the "AI Omnibus" — adopted 19 November 2025, politically agreed 7 May 2026, in force 27 July 2026 — Annex III high-risk rules apply "from 2 December 2027" and Annex I rules have "an extended transition period until 2 August 2028" (European Commission, retrieved 2026-09-18, SOURCED). We could not read the amending regulation and no consolidated version of 2024/1689 was retrievable, so the Commission page is the current timeline and the OJ text the original wording. Pin your reading to those dates.

What changes the exposure: editorial responsibility is the pivot. A named human standing behind the page is both the 50(4) carve-out and the practice verified AI writing describes.

7. The platform rules you already accepted

Not law, and still the fastest consequence: platform terms apply without a hearing.

Google's spam policies define scaled content abuse as "when many pages are generated for the primary purpose of manipulating search rankings and not helping users", and state the enforcement plainly: violations are found "through automated systems and, as needed, human review that can result in a manual action", and "Sites that violate our policies may rank lower in results or not appear in results at all" (Google Search Central, spam policies, retrieved 2026-09-18, SOURCED).

Note what the policy does not say. It turns on purpose and usefulness, not on whether a machine was involved: volume plus thin value is the trigger (ANALYSIS).

What actually reduces the exposure

None of this is fixed by choosing a better model, and none of it is legal advice.

Verify the claims that touch money, health, law or a named third party — the set that produces consequences. Our gate is the 42-check fact-check checklist.
Hold the evidence before publication, not after a challenge. The FTC's sequencing, and the cheapest.
Name an accountable human reviewer per page — the AI Act's carve-out, and what every court in section 2 was looking for.
Align your assistant with your policy pages, so no answer contradicts the page that governs it.
Say what your AI does in the words your tests support.

How This Guide Was Sourced

Written and maintained by the LogicBalls editorial team (logicballs.com). Disclosure: LogicBalls builds AI writing tools. Section 4 of this post applies to us: any accuracy claim we make about our own product needs a test behind it, and a post about AI liability is not the place to make one.

AI involvement. AI assisted the research and drafting, and AI performed the source checks: every quotation, date and figure above was fetched and matched against the judgment, order or official guidance itself on 2026-09-18, and every link was resolved on that date. Editorial responsibility for this page rests with the LogicBalls editorial team — which is, as section 6 notes, what the AI Act's carve-out turns on.

Sources. Moffatt v. Air Canada, 2024 BCCRT 149 (the tribunal's own site); the Mata v. Avianca sanctions order; Ayinde v Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin) via judiciary.uk; SEC press release 2024-36; the FTC's Operation AI Comply and DoNotPay releases and its Advertising FAQs; Regulation (EU) 2024/1689 via EUR-Lex; the European Commission's regulatory-framework page; Google Search Central's spam policies.

What could not be verified. The regulation amending the AI Act's high-risk timeline could not be read and no consolidated text of 2024/1689 was retrievable, so the current dates rest on the Commission's page. sec.gov returns HTTP 403 to a generic automated request and 200 to one that identifies itself. We make no claim about the merits or status of the Wolf River matter, cited here as allegations only.

What is not claimed. No frequency data, no ranking of the seven by likelihood, no penalty exposure for any jurisdiction, and no decision holding a business liable for republished AI-generated defamation, because we found none. This post is not legal advice and describes no jurisdiction's law completely.

No LogicBalls telemetry is used in this guide. Every figure above is external and linked.

Frequently Asked Questions

Can we blame the AI provider for a wrong answer?

No decision in the sources above put the consequence on the model provider. The business that published or relied on the output carried it.

Does a disclaimer on the chatbot help?

The Air Canada decision turned on reasonable care over the representation itself, and the argument that correct information sat elsewhere on the site did not succeed. A disclaimer is no substitute for the answer being right.

Is it illegal to publish AI-assisted content?

No. The duties attach to what the content claims, whether it is disclosed where the EU AI Act requires that, and whether it is useful rather than scaled. Google's spam policies target purpose and value, not authorship.

Does the EU AI Act apply to a blog post?

Article 50(4) reaches AI-generated text published to inform the public on matters of public interest, and exempts content with human review or editorial control where someone holds editorial responsibility. Whether a given post is "public interest" is a question for your own advisers.

Conclusion

Across all seven, existing law already had an answer. A chatbot is part of your website, a citation is your responsibility, an ad claim needs proof first, a statement about your own product must be true, a third party can sue over a false statement, disclosure applies where the AI Act says it does, and platform terms apply whether or not a regulator is involved. What generative tools changed is the volume of checkable claims reaching publication, not who answers for them.