Friendly AI chatbots may be less accurate, study says

Oxford researchers found that making AI chatbots warmer and friendlier made them less accurate.

After fine-tuning, five large language models were more likely to agree with false beliefs and give worse medical advice.

The risk was highest when users seemed vulnerable, raising concerns about persuasion, companionship, and mental health use cases.

The study suggests empathy-focused tuning can improve perceived trust but create safety tradeoffs for factual reliability.