Repeated Falsehoods Can Make AI Models Change Their Mind
A study of seven large language models finds that an answer that looks reliable in a single exchange can become unstable after a user repeats a false claim. The results point to a neglected test of AI safety: what happens after the conversation continues.