Anonymous2 days ago
A landmark Nature study from Oxford researchers found that AI models fine-tuned for emotional warmth produce 10 to 30 percentage points more factual errors than their unmodified counterparts, with the effect most pronounced when users express sadness. The findings expose a fundamental tension in AI design — between making chatbots feel empathetic and keeping them accurate — that has implications for the roughly 900 million weekly ChatGPT users and a growing ecosystem of emotionally responsive AI companions.