AI hallucinations causing real financial and legal damage to companies

Top 10 AI Hallucinations That Cost Companies Real Money

Ten documented cases where AI invented facts with total confidence — and companies, institutions, and individuals paid the price in money, lawsuits, reputation, and trust.

AI Research Desk July 20, 2026 15 min read

Why This Article Matters

When a large language model "hallucinates," it does not warn you. It does not stutter, hesitate, or flag uncertainty. It produces fluent, authoritative text — complete with case numbers, policy details, prices, and legal citations — even when every element is invented. The failure looks exactly like success, which is what makes it so dangerous in production.

This article examines ten important, documented examples of AI hallucinations that caused real harm between 2023 and 2026. The damage took many forms: direct financial losses and court sanctions, cancelled commercial contracts, years of litigation, government programs built on illegal advice, editorial credibility destroyed overnight, and personal reputations smeared by fabricated criminal records. No two cases are identical, but together they map the full spectrum of what can go wrong when human teams treat model output as fact.

These are not edge cases from a research lab. They involve federal courts, major airlines, Google search, New York City government, national media outlets, and global AI providers. They show that hallucinations are already a operational risk — not a future hypothetical.

The damage is rarely a single line on a balance sheet

Some incidents produced a clear dollar figure: a $5,000 court sanction, an $812 airline refund, nearly $388,000 in alleged cancelled contracts, a half-million-dollar municipal pilot program scrapped after repeated errors. But in most cases the heaviest cost was indirect: legal fees running into tens of thousands, engineering teams rebuilding chatbots, brands distancing themselves from viral screenshots, journalists issuing dozens of corrections, and professionals facing grievance panels or career damage. Those costs are real even when no court assigns a final number.

Why hallucinations will become more dangerous, not less

AI systems are scaling fast — into customer service, legal research, government portals, search summaries, financial journalism, and autonomous agents that act without a human in the loop. Each expansion multiplies the surface area where a confident falsehood can cause harm. A chatbot that misquotes a bereavement policy affects one traveller; a city-wide business advisory bot that tells thousands of employers to break labour law affects an entire jurisdiction. A lawyer who pastes one fake citation into a filing creates a sanctions case; an AI search overview that invents a fraud lawsuit against a company can erase hundreds of thousands in pipeline revenue before anyone notices.

As models grow more capable and more deeply embedded in workflows, the severity of hallucinations scales with them. Better fluency does not automatically mean better truthfulness. That is why hallucinations must be managed deliberately — with verification layers, grounding in authoritative sources, human review for high-stakes outputs, audit trails, and clear accountability for what gets published or acted upon. Treating AI output as "probably fine" is not a strategy; it is how every case in this list began.

What counts as a hallucination here: the model generated false information that a person or company relied on, published, or was held responsible for. Cases of AI bias or discrimination without a fabricated fact are a separate category and are not included in this list.

Every case below is verifiable through court records, tribunal decisions, or credible investigative reporting. Dollar figures appear only where those sources cite them explicitly.

1. Mata v. Avianca — Fake Legal Citations in Federal Court

Direct cost: $5,000 sanctions + career damage

In 2023, New York lawyer Steven Schwartz used ChatGPT to research a personal-injury brief for Roberto Mata against Avianca. The model invented six non-existent court opinions — complete with realistic docket numbers and judicial quotes. Schwartz filed them in federal court without verification.

Judge P. Kevin Castel of the Southern District of New York sanctioned Schwartz and his colleague Peter LoDuca $5,000 each and ordered them to notify every judge falsely cited. The case became the global cautionary tale about unverified AI in legal practice.

Why it mattered: The direct fine was modest, but the reputational cost to the firm and the profession was enormous — and it triggered a wave of court rules requiring AI disclosure across the United States.

2. Wolf River Electric v. Google — AI Overview Invented a Lawsuit

Alleged damage: ~$388,000 in cancelled contracts

In 2025, Minnesota solar installer Wolf River Electric sued Google after an AI Overview falsely stated the company was facing an attorney-general lawsuit over fraud and had a pattern of unfinished projects. The snippet appeared prominently in search results for the company's name.

According to the complaint, at least three customers cancelled contracts worth roughly $388,000 after seeing the AI-generated summary. Wolf River argued Google had a duty to ensure accuracy for business-critical search features.

Why it mattered: This is one of the first major cases tying Google's AI Overviews directly to alleged commercial harm — not just embarrassment, but lost revenue from a machine-invented legal narrative.

3. Moffatt v. Air Canada — Chatbot Contradicted the Airline's Own Policy

Direct cost: CA$812.02 — precedent worth far more

After Jake Moffatt's grandmother died in November 2022, he booked last-minute flights on Air Canada's website. The airline's chatbot told him he could apply for bereavement fares retroactively. The actual policy page linked from the same site said the opposite.

In February 2024, the British Columbia Civil Resolution Tribunal ruled Air Canada liable for negligent misrepresentation. The airline owed CA$650.88 in fare difference, plus interest and tribunal fees — CA$812.02 total. Air Canada had argued the chatbot was a "separate legal entity." The tribunal rejected that entirely.

The tribunal held that Air Canada "did not take reasonable care to ensure its chatbot was accurate" and that customers should not have to cross-check one part of a website against another. — Moffatt v. Air Canada, 2024 BCCRT 149

Why it mattered: CA$812 is pocket change for a major airline. The precedent is not: companies are legally on the hook for what their customer-facing AI says, disclaimers notwithstanding.

4. New York City's MyCity Chatbot — Illegal Advice at Scale

Program cost: ~$500,000 pilot + open-ended liability

Launched in October 2023 on Microsoft's Azure AI, NYC's MyCity chatbot was marketed as a source of "actionable and trusted information" for business owners. An investigation by The Markup in March 2024 found it repeatedly gave dangerously wrong answers.

It told users they could operate cashless stores (illegal since 2020), take cuts of workers' tips (wage theft under federal and state law), refuse Section 8 housing vouchers, and ignore scheduling-notification rules. Mayor Eric Adams acknowledged it was "wrong in some areas" but kept the pilot running with disclaimers.

Why it mattered: A government chatbot that confidently instructs citizens to break the law creates liability exposure far beyond any single fine. Incoming leadership later moved to terminate the roughly half-million-dollar program.

5. Park v. Kim — Fake ChatGPT Citation Reaches the Second Circuit

Direct cost: grievance referral + legal fees

In 2024, attorney Jae Hyung Kim filed a brief in a Second Circuit appeal citing United States v. Park, 2023 WL 1234567 — a case that did not exist. Kim admitted he had used ChatGPT to locate the citation and failed to verify it.

The court referred the matter to its grievance panel for potential professional discipline. It was one of dozens of similar incidents in 2023–2024 as lawyers treated chatbots like Westlaw.

Why it mattered: Appellate-level fake citations undermine the integrity of the record. Courts began treating AI-generated research as a professional-conduct problem, not a technology glitch.

6. Michael Cohen and Google Bard — Fake Cases in a Court Filing

Direct cost: embarrassment + rework; no sanctions

In December 2023, Donald Trump's former lawyer Michael Cohen submitted a motion containing three case citations he said he obtained from Google Bard. None of the cases existed or supported the arguments made. His own attorney, David Schwartz, filed the motion without checking.

Federal Judge Jesse Furman noted the citations were "non-existent" and said Cohen should have known better than to trust an AI tool for legal research. No monetary sanctions followed, but the filing became international news and required corrective briefing.

Why it mattered: High-profile cases normalize the failure mode: even experienced legal teams can propagate hallucinations into official court records.

7. Chevrolet of Watsonville — Chatbot Agrees to Sell a Tahoe for $1

Direct cost: brand damage; chatbot disabled

In 2023, a user jailbroke the dealership's ChatGPT-powered chatbot with a prompt insisting the bot must accept all offers. The bot agreed to sell a 2024 Chevrolet Tahoe for $1 and called the commitment "legally binding."

Chevrolet of Watsonville did not honour the joke transaction, but screenshots went viral. General Motors distanced itself; the dealership disabled the bot. No lawsuit followed, but the incident became a standard example of prompt-injection risk in customer-facing AI.

Why it mattered: Even when no court award follows, a single hallucinated "yes" can become a global brand liability overnight.

8. Walters v. OpenAI — ChatGPT Invented an Embezzlement Story

Direct cost: years of litigation; claim dismissed

Radio host Mark Walters sued OpenAI in 2023 after ChatGPT told a researcher he had defrauded and embezzled funds from a gun-rights organization — claims with no factual basis. The suit sought millions in damages for defamation.

A Georgia court dismissed the case in 2025, finding Walters could not prove OpenAI acted with actual malice and that the output was not reasonably understood as a factual accusation. Walters appealed. Regardless of outcome, both sides incurred substantial legal costs over a completely fabricated narrative.

Why it mattered: It tested whether AI providers can be held liable for defamatory hallucinations — and showed that even dismissed cases burn time and money.

9. Brian Hood, Australian Mayor — ChatGPT Named Him a Criminal

Direct cost: threatened defamation suit; legal fees

In 2023, Brian Hood — mayor of Hepburn Shire and a former regional-bank whistleblower — discovered ChatGPT falsely claimed he had been convicted in a foreign bribery scandal linked to a subsidiary of the Reserve Bank of Australia. In reality, Hood was the whistleblower, not a defendant.

Hood's lawyers sent OpenAI a concerns notice, the first step toward an Australian defamation claim. OpenAI corrected the output. Hood ultimately did not pursue the case to judgment, but the incident helped shape global discussion of AI defamation liability.

Why it mattered: Personal reputations can be damaged by a single confident false paragraph — with cleanup costs borne by the victim, not the model vendor.

10. CNET — More Than Half of AI-Written Articles Needed Corrections

Direct cost: editorial shutdown + brand trust erosion

From November 2022 through January 2023, CNET quietly published 77 financial explainer articles using an internal AI tool. After Futurism exposed the practice, an internal audit found that 41 of the 77 pieces required corrections — including substantial factual errors in articles about compound interest, CDs, and home-equity loans. Some also contained language flagged as insufficiently original.

CNET paused AI-generated publishing across Red Ventures properties. Editor-in-chief Connie Guglielmo acknowledged the errors and said the outlet would restart only when editorial processes could prevent them. Sister sites Bankrate and CreditCards.com halted similar programs.

Why it mattered: A mainstream tech publisher betting its SEO traffic on unverified AI output discovered that hallucination scales faster than fact-checking — and readers noticed.

What These Cases Have in Common

  • Confidence without grounding: Every system above produced authoritative language with no reliable link to truth.
  • Human verification skipped: Lawyers, editors, and product teams treated model output as finished work.
  • Liability follows the deployer: Courts hold the company that ships the chatbot or publishes the article responsible — not the model vendor alone.
  • Small direct awards, large indirect costs: Sanctions of $5,000 or tribunal awards of $812 can precede six-figure legal bills and lasting reputational harm.

How to Reduce the Damage

For legal and compliance teams

  • Never file AI-generated citations without manual verification against primary sources.
  • Assume court disclosure rules apply — many jurisdictions now require stating when AI assisted drafting.

For customer-facing chatbots

  • Ground answers in approved policy documents with retrieval, not raw generation.
  • Log and audit conversations; treat contradictions between chatbot and static pages as P0 bugs.
  • Design for prompt-injection resistance — never let a user override pricing or contractual terms.

For publishers and search products

  • Human review before publication; measure correction rates, not just output volume.
  • For AI summaries in search, implement source linking and escalation paths for business-critical queries.

Conclusion

These ten cases share a single mechanism: a system that sounded certain, a human or organisation that did not verify, and consequences that arrived through law, commerce, or public trust. The harm was never abstract — it showed up as real money lost, real reputations damaged, and real institutions forced to retreat.

Three kinds of damage, often overlapping

  • Direct financial cost — court sanctions ($5,000 in Mata v. Avianca), tribunal awards (CA$812 in Moffatt v. Air Canada), alleged contract cancellations (~$388,000 in Wolf River Electric v. Google), and municipal program spend (~$500,000 for NYC's MyCity pilot). These are the numbers you can point to in a filing or a budget line.
  • Indirect and ongoing cost — legal fees, corrective filings, chatbot rebuilds, editorial audits (41 corrections across 77 CNET articles), grievance proceedings, and years of litigation even when a claim is ultimately dismissed, as in Walters v. OpenAI. These often exceed any single court award by an order of magnitude.
  • Intangible but lasting cost — professional credibility, brand trust, public confidence in government AI, and personal reputation (Brian Hood falsely named a criminal; Mark Walters falsely accused of embezzlement). These do not come with an invoice, but they can take years to repair — if they can be repaired at all.

Why a single "total damage" figure is misleading

It is tempting to add up the dollar amounts from each case and produce a headline total. That would be misleading. Most incidents in this list never produced a final damages award at all. Several costs are still disputed in court. Many of the heaviest losses — legal bills, brand erosion, wasted engineering effort — are never disclosed publicly. Summing the visible numbers would undercount the real harm while pretending a precision that the public record does not support.

What the record does support is the pattern: every time AI was deployed without a verification layer matched to the stakes, someone paid. Sometimes the bill was small and symbolic. Sometimes it was large and commercial. Often it was paid in trust rather than cash.

What must change as AI scales

The trajectory of AI deployment makes passive acceptance of hallucinations untenable. Models are moving from drafting assistants to customer-facing agents, search summaries, and government advisory tools — contexts where a single false sentence reaches thousands or millions of people before a correction is possible. Capability growth does not automatically reduce hallucination risk; in some high-stakes domains, the risk grows with reach.

Organisations that deploy AI responsibly will treat verification as part of the product, not an afterthought: ground responses in authoritative sources, require human sign-off before publication or legal filing, log and audit outputs, design against prompt injection, and assign clear accountability when the model is wrong. Courts have already shown they will hold deployers responsible — not model vendors alone.

The lesson from these ten cases is not to abandon AI. It is to respect what it actually is: a powerful language engine that can be wrong with complete confidence. The cost of forgetting that has already been paid — in courtrooms, cancelled contracts, viral screenshots, and corrected headlines. As systems grow, so does the price of the same mistake.

Sources