A new study by Anthropic and the University of Toronto sought to quantify the potential for AI chatbots to induce harmful behaviors, analyzing 1.5 million anonymized conversations with the Claude model.

Study Results

The research focused on three main ways in which a chatbot can negatively influence a user's thoughts or actions, leading to undesirable consequences. The results indicate that, although such situations are not the norm, their incidence remains a problem that should not be underestimated.