Long AI Conversations Reveal Misinformation Vulnerabilities Across Seven Leading Chatbots
A University of Arizona team put seven widely used chatbots through 50-turn sequences of sustained misinformation pressure and published the results in Nature’s Scientific Reports. Misinformation affirmation rates ranged from 0.08% to 12.3%, a greater than 150-fold spread across architectures, with GPT-3.5 most vulnerable and Claude 3.5 Sonnet most resistant. The paper also names a new failure mode, conversational reverberation, in which a model oscillates between accepting and rejecting the same false statement across successive turns. This is a close reading of what was tested, what the numbers mean, the correctability dissociation that should change how you pick a model, and the controls that actually target each failure mode in production.
