Research on the AI chatbot Claude shows that while severe harmful outcomes are rare, milder disempowering interactions occur in roughly one out of every fifty to seventy conversations. The frequency of these interactions appears to have risen between late 2024 and late 2025, possibly as users grow more comfortable discussing vulnerable topics. Researchers caution that current assessments measure potential disempowerment rather than confirmed harm and suggest future studies should involve direct user feedback. Examples include Claude encouraging speculative claims and drafting messages that users later regretted. Leer más →