Confidently Wrong
Eight answers from a model that sounds equally certain of all of them. Name the failure in each — because the control is different for every one, and “the AI was wrong” is not a finding anyone can act on.
1. Asked for the case that settled the point, it names Harper v. Sandoval Holdings (2019) with a docket number. The docket belongs to an unrelated bankruptcy.
2. Asked who chairs the committee, it names a person who stepped down fourteen months ago.
3. Given last quarter's figures it computes the growth rate correctly, then concludes the trend will hold for the next eight quarters.
4. Asked whether your contract permits sub-processing, it answers with confidence, never having been given the contract.
5. It cites Article 31 of the EU AI Act for a transparency duty. Article 31 exists — it concerns notified bodies.
6. It states the penalty ceiling using the figure from the draft Regulation rather than the final text.
7. It reads a 3% control-failure rate as immaterial, knowing neither the population nor the tolerance.
8. Asked for your current headcount, it returns a figure to the nearest person.
A companion to Signify Kids' Hallucination Hunt, which asks a child the same question in smaller words. Both are about noticing that fluency and accuracy are unrelated.