Warmth-first AI sacrifices accuracy, study finds
A Nature-published study from Oxford researchers finds that AI models fine-tuned for warmer, more empathetic responses tend to be less accurate. Warmer models were about 60% more likely to give incorrect answers than unmodified ones, averaging a 7.43 percentage-point increase in error across prompts (baselines ranged 4–35%). Errors rose further when users expressed sadness or held incorrect beliefs, and pre-training for warmth generally worsened accuracy while “colder” tuning could match or beat the originals. The work warns that prioritizing perceived helpfulness and friendliness can come at the expense of truth in high-stakes, socially embedded AI systems, noting the experiments used smaller, older models.