← All lab updates

We tested our hallucination reading on AI answers checked by people: it reads how fabrication is written, not whether a fact is true

We tested our live hallucination reading on public AI answers in which people marked every invented fact, in five languages. Inside the same answer it ranks the invented sentence above a correct one about 7 times in 10. Then we built a harder test: the same sentence twice, identical except for the false fact, replaced with the true one. There the reading does not pick the false one, and often flags the true one. It reads the way made-up content is written, not whether a fact is true, which is what PSA is for: it reads behaviour, it is not a fact-checker.