The Casebook
What actually happened — incidents on the public record, where AI helped and where it burned. Every claim here is sourced, because a page about confident wrong answers has no business being one.
British Columbia, Canada · 2022–2024
The chatbot that invented a refund policy — and the airline had to honour it
A support bot described a discount that did not exist. A tribunal held the airline to its words.
4 sources · the check it teaches →
Sources verified 15 August 2026
Southern District of New York · 2023
He asked ChatGPT whether the cases were real. It said yes.
Six fabricated decisions reached a federal court — and the lawyer's attempt to verify them made him more confident, not less.
2 sources · the check it teaches →
Sources verified 15 August 2026
Global · 2021 onward
The most useful AI in science ships a number saying how much to trust it
AlphaFold predicts protein structures to near-atomic accuracy — and marks, residue by residue, where it is guessing.
2 sources · the check it teaches →
Sources verified 15 August 2026
Australia · 2015–2023
It spread a year's pay evenly across the fortnights, then called the difference a debt.
At least $1.763 billion raised against 433,000 people, calculated from a fortnightly income none of them had actually earned.
5 sources · the check it teaches →
Sources verified 17 August 2026
United States · 2012–2024
The system said she was a shoplifter. The alert did not say how sure it was.
Rite Aid ran facial recognition in hundreds of stores for eight years. The one number that measured its doubt was stripped out before it reached the person who acted on it.
7 sources · the check it teaches →
Sources verified 17 August 2026
Global · 2024
The best weather model doesn't give you one forecast. It gives you fifty.
GenCast beat the world's leading forecast system on 97% of tests — by refusing to answer with a single number.
2 sources · the check it teaches →
Sources verified 15 August 2026
England · 2020
Exams were cancelled, so each grade was calculated from the school's past results and the teacher's ranking
The model produced a credible national distribution. Four days after it reached individual students, it was withdrawn.
5 sources · the check it teaches →
Sources verified 17 August 2026
Hong Kong · 2024
Everyone else on the video call was fake. He paid out $25 million.
A finance worker joined a conference call with his CFO and colleagues. Not one of them was a person.
2 sources · the check it teaches →
Sources verified 15 August 2026
Thailand · 2018–2020
90% accurate in the lab. Then it met the lighting in a real clinic.
A screening model that matched specialists on clean data rejected more than a fifth of the photographs nurses actually took.
2 sources · the check it teaches →
Sources verified 15 August 2026
United States · 2021
The house-pricing model was accurate on average. Average was not the problem.
Zillow bought homes on an algorithm's valuations and wrote down $407.9 million of them in a single year.
2 sources · the check it teaches →
Sources verified 15 August 2026
New York City, United States · 2023–2026
The city's own website told business owners they could take their staff's tips
New York put a chatbot on nyc.gov to answer questions about the law. It got them wrong, and the city left it up.
5 sources · the check it teaches →
Sources verified 17 August 2026
South Korea · 2023
They pasted the chip code in to get it debugged. That was the disclosure.
Three Samsung engineers used ChatGPT the way anyone would. Within weeks the company banned it outright.
2 sources · the check it teaches →
Sources verified 15 August 2026
United States · 2022–2023
The hiring software rejected them for being older. Nobody noticed until someone lied about their birthday.
More than 200 qualified applicants were screened out automatically. It surfaced because one of them applied twice.
3 sources · the check it teaches →
Sources verified 15 August 2026