What One Wrong Fact Actually Costs: 8 Real Consequences

TL;DR

Eight documented cases where a wrong fact reached a court, regulator or customer - the fines were small; referrals, disqualification, excluded evidence and customer notices were not.

A wrong fact rarely costs much money on the day it is caught. In the eight documented cases below, the cash runs from CA$812.02 to US$225,000 per party. The expensive part is everything ordered alongside it: letters to judges whose names were used, referral to a professional regulator, removal from a case, evidence thrown out, notices to past customers. In every case the wrong fact was checkable, and nobody checked it before it went out.

Every case comes from a court judgment, a tribunal decision or a regulator's own release, read at source on 2026-09-17. Six involve text produced, or suspected to have been produced, by generative AI or an automated chatbot. Two are regulators acting on wrong claims companies made about their own AI.

Key Takeaways

The fine is the smallest line. A New York federal court set a US$5,000 penalty, then ordered letters to every judge falsely named as the author of a fake opinion (Mata v. Avianca sanctions order, retrieved 2026-09-17).

"The substance was right" did not save anyone. A US court excluded an expert's whole declaration over AI-generated citations although he stood by what it argued (Kohls v. Ellison order, retrieved 2026-09-17).

You own what your chatbot says. A Canadian tribunal called the suggestion that a chatbot is a separate entity "a remarkable submission" (Moffatt v. Air Canada, 2024 BCCRT 149, retrieved 2026-09-17).

A policy did not prevent it. An Alabama court disqualified three lawyers who had "repeated warnings, internal controls, and firm policies about the dangers of AI misuse" (Johnson v. Dunn sanctions order, retrieved 2026-09-17).

An untested accuracy claim is a wrong fact too. The FTC ordered US$193,000 in relief and notices to past subscribers over an "AI lawyer" never tested against lawyers (FTC, retrieved 2026-09-17).

The eight, at a glance

#	Consequence	Decided by	Cash	What else was ordered
1	Penalty plus letters to named judges	US federal court, 2023	US$5,000	Letters to the client and each falsely named judge
2	Wasted costs and regulator referral	English High Court, 2025	£2,000 each, two payers	Referral to both legal regulators
3	Liability for errors a client supplied	English High Court, 2025	Costs reserved	Referral to the solicitors' regulator
4	Liability for a chatbot's answer	BC Civil Resolution Tribunal, 2024	CA$812.02	A published decision against the company
5	Evidence excluded	US federal court, 2025	None	Declaration excluded, amended version denied as moot
6	Removal from the case	US federal court, 2025	None	Disqualification, bar referral, copies to clients
7	Penalty for false AI claims	US SEC, 2024	US$400,000 across two firms	Settled charges
8	Relief for an untested AI claim	US FTC, 2025	US$193,000	Notices to 2021-2023 subscribers

Six come from the legal system because courts publish their reasoning and check citations for a living. The same failure happens in marketing copy every day; it just rarely produces a judgment.

1. A penalty, then a letter to every judge whose name was used

In Mata v. Avianca, lawyers filed non-existent opinions with fake quotes and citations "created by the artificial intelligence tool ChatGPT", then "continued to stand by the fake opinions after judicial orders called their existence into question" (Opinion and Order on Sanctions, S.D.N.Y., 22 June 2023, retrieved 2026-09-17, SOURCED).

The penalty was US$5,000, jointly and severally across the two lawyers and their firm. The court also ordered them to send their own client the order and hearing transcript, and to send each judge falsely named as the author of the six fake opinions a copy of the order with the invented "opinion" attached.

ANALYSIS. The sanction turned on what happened after the error was questioned. Withdrawing a wrong fact the day someone challenges it costs almost nothing. Defending it is what made this expensive.

2. Wasted costs and a referral to the regulator

In Ayinde v London Borough of Haringey, grounds for judicial review cited five cases that do not exist. The judge ordered the barrister and the law centre each to pay £2,000 to the other side and referred the matter to the Bar Standards Board and the Solicitors Regulation Authority (Ayinde and Al-Haroun, [2025] EWHC 1383 (Admin), para 49, retrieved 2026-09-17, SOURCED).

The Divisional Court made its own referral too (para 70), and found the threshold for contempt proceedings met while choosing not to start them (paras 68-69). The barrister denied using AI; the court did not settle how the fake cases were produced, and neither do we.

Her defence was that the underlying legal principles were sound. It was likened to a mislabelled tin that "in fact, contains the correct product". The court said this "entirely misses the point and shows a worrying lack of insight" (para 67).

3. Liability for errors someone else handed you

The second case in the same judgment, Al-Haroun v Qatar National Bank, is the one content teams should read. A schedule by one of the court's judicial assistants lists 45 citations; in 18 the case cited does not exist (para 74, SOURCED). The client said they came from "publicly available artificial intelligence tools, legal search engines and online sources" (para 76).

The solicitor had relied on his client. The court: "A lawyer is not entitled to rely on their lay client for the accuracy of citations of authority or quotations that are contained in documents put before the court by the lawyer" (para 81). He referred himself to the regulator, and the court referred him as well.

ANALYSIS. Swap "lawyer" for "publisher" and "lay client" for freelancer, subject-matter expert or AI draft. Whoever signs off owns the facts. The method is in how to trace an AI claim back to its original source.

4. Paying for what your chatbot said

A support chatbot on Air Canada's website told a customer they could apply for a bereavement fare retroactively. The airline did not permit that. The tribunal ordered Air Canada to pay CA$812.02: CA$650.88 damages, CA$36.14 interest and CA$125 in fees (Moffatt v. Air Canada, 2024 BCCRT 149, para 44, 14 February 2024, retrieved 2026-09-17, SOURCED). Air Canada gave no information about how its chatbot worked (para 14), so this is an automated answer, not necessarily a generative model.

The reasoning cost more than the award. Air Canada, the tribunal said, in effect suggested "the chatbot is a separate legal entity that is responsible for its own actions. This is a remarkable submission." And: "It makes no difference whether the information comes from a static page or a chatbot" (para 27). Nor should customers "have to double-check information found in one part of its website on another part of its website" (para 28).

ANALYSIS. "The right answer is on another page" is the standard defence of a wrong automated answer. It failed here.

5. Evidence thrown out, even though the argument stood

In Kohls v. Ellison, a challenge to Minnesota's deepfake law, the Attorney General filed a declaration from an expert on AI and misinformation. It cited two academic articles that do not exist and misattributed a third. The expert said he used GPT-4o while drafting and failed to verify the citations it produced (Order, D. Minn., 10 January 2025, ECF No. 46, retrieved 2026-09-17, SOURCED).

The court accepted the mistake was not intentional and commended the prompt admission. It excluded the declaration anyway, saying the fake citations shatter the expert's credibility, and that "the Court cannot accept false statements—innocent or not—in an expert's declaration submitted under penalty of perjury."

ANALYSIS. The court did not have to disagree with any of the substance to set all of it aside. One fabricated citation takes the true claims around it down too. The English court made the same point about the fake cases in item 2, which it noted could have been checked on the National Archives' case-law website (para 67).

6. Removed from the case, with a copy to every client

In Johnson v. Dunn, two motions from lawyers at a large firm carried five problematic citations, later identified as "hallucinations of a popular generative artificial intelligence ('AI') application, ChatGPT." The court publicly reprimanded three lawyers, disqualified them from further participation in the case, referred the matter to the Alabama State Bar, and directed the order to be published (Sanctions Order, N.D. Ala., 23 July 2025, retrieved 2026-09-17, SOURCED). They must also give the order to their clients, opposing counsel and presiding judge "in every pending state or federal case in which they are counsel of record."

Why not a fine: "If fines and public embarrassment were effective deterrents, there would not be so many cases to cite."

ANALYSIS. These lawyers had AI policies and warnings. A policy tells people to check. Only a step in the workflow that does the check stops the error.

7. A penalty for a wrong fact about your own AI

On 18 March 2024 the SEC announced settled charges against two investment advisers "for making false and misleading statements about their purported use of artificial intelligence." Delphia paid a US$225,000 civil penalty and Global Predictions US$175,000; one claim cited was Global Predictions' "first regulated AI financial advisor" (SEC press release 2024-36, retrieved 2026-09-17, SOURCED).

ANALYSIS. For a content team this is your product-page copy. A capability sentence nobody checked against the product is a wrong fact a regulator can read.

8. Relief and customer notices for an accuracy claim nobody tested

The FTC's final order against DoNotPay, announced 11 February 2025, concerned its marketing as "the world's first robot lawyer." The FTC said the company did not test its "AI lawyer" against human lawyers and "did not hire or retain attorneys to test the quality and accuracy of its service's law-related features." The order requires US$193,000 in relief and notice to 2021-2023 subscribers (FTC, retrieved 2026-09-17, SOURCED).

ANALYSIS. A performance claim with no testing behind it is the same gap what "99% accurate" actually means examines from the buyer's side.

What to do before it costs you

Treat every citation as unverified until someone opens the source itself. Why models invent them: why ChatGPT makes up sources.
Make the check a workflow step, not a policy.
Hold every contributor's facts to the same standard, whether they come from a freelancer, an expert or a model.
Correct fast and publicly rather than defending.
Check your own product claims, not just your articles.

The full pass is the free 42-check AI fact-check checklist.

How This Guide Was Sourced

Written and maintained by the LogicBalls editorial team (logicballs.com). Disclosure: LogicBalls builds AI writing tools, which produce exactly the kind of text these cases are about. Everything here applies to our output too.

AI involvement. AI assisted with research and drafting. Every quotation, figure, date and paragraph number was matched against the text of the judgment, decision or release itself on 2026-09-17, not against a summary of it. Sources were read from the UK judiciary site, govinfo.gov, CourtListener's RECAP archive, the Civil Resolution Tribunal's decisions site, sec.gov and ftc.gov.

What could not be fetched. A widely reported partial refund by a consultancy to an Australian government department, over a report with fabricated references, is not included: the department's site timed out repeatedly and we could not reach the parliamentary record of the amount. An earlier LogicBalls post noted CanLII returned HTTP 403 for Moffatt; this time we read it on the tribunal's own site.

What we excluded deliberately. A 2025 FTC order over an AI content detector's accuracy claim fits this post closely. LogicBalls sells an AI detector, so building a post around a competing vendor's enforcement action would be a head-to-head, not a neutral example.

What is not claimed. No estimate of how often AI errors reach courts or customers, and no total cost figure. Eight cases are examples, not a sample. We leave out the individuals the judgments name.

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

Frequently Asked Questions

What is the biggest cost of an AI error?

In these cases, almost never the fine. The largest single penalty is US$225,000. The lasting costs are regulator referrals, disqualification, excluded evidence and orders to tell clients or customers.

Is it illegal to use AI to write court filings or business content?

None of these decisions says so. The Mata court wrote that "there is nothing inherently improper about using a reliable artificial intelligence tool for assistance." Every penalty here was for failing to check what was filed or claimed.

Does admitting the mistake quickly avoid consequences?

It helps and is not enough. Kohls commended the prompt admission and still excluded the evidence.

Is a company liable for what its chatbot tells customers?

In Moffatt v. Air Canada, yes: the airline was responsible "whether the information comes from a static page or a chatbot." That is one British Columbia tribunal decision, not a universal rule.

Conclusion

Across eight decisions from three countries the pattern repeats: the fact was checkable, the check did not happen, and most of the bill was paid in credibility.