How to Tell When an AI Is Guessing
AI hallucinations read exactly like real answers. Six tells that give a guess away, a 60-second check that catches most of them, and what my own test found.
Earlier this month a lawyer was struck off in the UK for citing cases that didn't exist. His AI had made them up, in the same fluent, footnoted voice it uses for cases that do. A database that tracks court decisions involving fabricated AI material passed 2,000 entries this year, and 815 of them involve lawyers, people whose job is checking sources.
The reason they got caught out is the reason you will: an AI's guess reads exactly like its knowledge. There's no change of tone, no hesitation, no smaller font. OpenAI's own researchers explained why last September. Models are trained and tested like students sitting an exam with no penalty for wrong answers, so they learn to bluff.
You can't fix the training. You can learn the tells, and there are about six of them. I also ran a small test of my own for this guide, which turned out to be more reassuring than I expected, and more instructive.