Authority explainer · last reviewed 14 October 2026
AI hallucination examples from the workplace
This page is maintained by the instructor of the Udemy course linked below. Each case cites the court record, tribunal decision, or contemporaneous reporting it is drawn from, checked 14 October 2026.
Short answer
An AI hallucination is a confident, fluent answer that is simply wrong. It has already produced a court sanction over invented case citations, a tribunal judgment over an invented airline policy, and a product-demo factual error that coincided with a single-day share drop widely reported at roughly $100 billion. The fix is not avoiding AI; it is verifying anything it states as fact before that fact reaches a client, a court, a customer, or a demo stage.
What counts as a hallucination
Large language models generate text by predicting plausible next words from patterns learned in training, not by looking up facts in a database and checking them. Most of the time that produces an accurate-sounding and genuinely accurate answer. When it does not, the output reads exactly as fluent and confident as when it does, which is what makes hallucinations hard to catch by tone alone: a wrong citation, a wrong policy, or a wrong fact sits in the same voice as a correct one.
Three documented cases
Ordered by what happened, not by severity. Each is drawn from a court record, a tribunal decision, or contemporaneous press reporting rather than from a vendor's own account.
01 · Litigation sanction · checked 14 October 2026
A lawyers' brief full of invented court cases
In 2023, two New York lawyers were sanctioned for submitting a ChatGPT-invented citations brief; ChatGPT affirmed the fake cases were real when asked. The court's sanctions order is the primary record of what the brief contained and what the lawyers told the court when questioned about the citations.
02 · Tribunal judgment · checked 14 October 2026
An airline chatbot's invented bereavement policy
In 2024, the BC Civil Resolution Tribunal awarded a passenger damages for negligent misrepresentation over a nonexistent bereavement policy stated by Air Canada's website chatbot, dismissing what it described as the airline's remarkable suggestion that the chatbot was a separate legal entity. The tribunal did not find that the bot ran on generative AI; it treated the case as ordinary liability for what the airline's own website told a customer, not as a regulatory enforcement action.
03 · Market impact · checked 14 October 2026
A telescope error and a reported $100 billion share drop
In February 2023, a promotional video for Google's Bard chatbot had it claim that the James Webb Space Telescope took the first pictures of a planet outside our solar system. Astronomers pointed out, and NASA's own site records, that the European Southern Observatory's Very Large Telescope captured the first exoplanet images back in 2004. Alphabet's shares fell roughly 7.7% on the day the error was reported, erasing on the order of $100 billion in market value, as widely reported at the time.
ABC News, "Google parent company loses $100 billion in shares"
What reduces the risk at work
Match the depth of checking to the stakes of the task. A draft email can tolerate a wrong word; a court filing, a customer promise, or a public demo cannot. Treat any name, number, citation, or stated policy from an AI tool as a claim to verify against an independent source, not as a fact already checked. Keep a human who reads the output against that source in charge of anything that leaves the building.
Frequently asked questions
What is an AI hallucination?
An AI hallucination is output from a language model that sounds confident and specific but is invented, wrong, or unsupported by any real source, with nothing in the tone of the answer signalling the uncertainty.
Can an AI hallucination actually cause harm to a business?
Yes. The three cases below cost two lawyers a court sanction, cost an airline a tribunal judgment, and lined up with a single-day share drop at one technology company widely reported at roughly $100 billion.
Why do AI systems hallucinate confidently instead of saying they do not know?
A language model predicts plausible-sounding next words from patterns in its training data; it has no built-in step that checks a generated fact against the real world or flags low confidence, so a fluent, wrong answer can read identically to a fluent, correct one.
How do you catch a hallucination before it causes harm?
Verify anything an AI tool states as fact, especially names, numbers, citations, and policies, against an independent primary source before it reaches a client, a court, a customer, or a public demo. Treat fluent prose and correct prose as two different claims, not one.
Sources
- Mata v. Avianca, Inc., S.D.N.Y. docket and sanctions order, June 2023
- Moffatt v. Air Canada, 2024 BCCRT 149, CanLII
- ABC News, "Google parent company loses $100 billion in shares", 9 Feb 2023
General information, not legal advice. Last reviewed 14 October 2026.
Build the habit of verifying before it ships.
A 45-minute workplace baseline on Udemy covers where these systems fail and how to judge their output.