About
Confidently Wrong records cases where an AI tool gave a wrong number, figure, policy or citation with no sign of doubt, and someone acted on it at work. It's for people who get handed AI output and have to put their name next to it.
Rules
- Sourced or it's out. Every field comes from the linked reporting, court record, regulator decision or company statement. We open every link before publishing and drop dead ones.
- No embellishing. If the sources don't say who caught it or what it cost, the entry says unknown.
- Our opinion is labeled. The "check that would have caught it" field is our analysis. It's marked on every page.
- Corrections are public. If we get something wrong, we fix the entry and note the date.
The four checks
- source citation
- Every claim links to a primary record a person can open, and the quote or number is actually in it.
- recompute in code
- Numbers come from a calculation or a table lookup, not from text generation. The model can phrase the result; it doesn't produce it.
- refuse when data missing
- If there's no document or record behind the answer, the system says so and hands off instead of guessing.
- human review
- A named person checks the output against the source before it ships.
We pick the plainest single check that would have stopped the error before it reached anyone. Several cases could take more than one; the case page explains our choice.
What we leave out
Prompt-injection pranks, demos nobody relied on, and claims we can't trace to a primary or reputable source. We considered and left out three widely shared stories because the AI's role or the details couldn't be confirmed from accessible sources.
We don't rank or score AI vendors.
Other trackers
For court filings specifically, Damien Charlotin's AI Hallucination Cases database is far more complete. The AI Incident Database covers AI harms broadly. BankingNewsAI tracks banking incidents. We focus on numbers and citations in everyday work.
Who
Confidently Wrong is part of DON'T TRUST ME AI, a field journal of independent checks on AI claims, run by Leland Jourdan II. The view behind it: keep generative AI out of the decision path, and let a refusal be the right answer when the numbers can't be verified.