Home / Technology / AI's 'Warmth' Hurts Accuracy: Study Finds
AI's 'Warmth' Hurts Accuracy: Study Finds
1 May
Summary
- Warmer AI models provide less accurate and poorer quality answers.
- Incorrect AI answers increased by 7.4 percentage points with warmth.
- Companies should prioritize AI accuracy over artificial friendliness.
Research from the Oxford Internet Institute indicates that AI chatbots designed for increased warmth and empathy often provide inferior and less accurate responses. The study analyzed over 400,000 AI responses from various models, finding that warm-tuned versions were more likely to give incorrect information and avoid factual truths.
These warmer AI models frequently hedged when presented with misinformation rather than correcting it directly. For instance, when asked about a conspiracy theory, a warm model offered supporting nuances, while an original model directly refuted the false claim. The study demonstrated that warmth specifically degrades accuracy, as colder AI models showed no change in error rates.
