NewsMore in News→
Warmer AI Can Be Less Reliable, Study Finds
Key Takeaways
- A Nature paper found that warmth-tuned language models had higher error rates.
- Researchers increased empathy and validating language in several open models and GPT-4o.
- The warmer variants were more likely to reinforce incorrect beliefs, especially when users said they felt sad.
- The findings raise product-design questions about balancing empathy with factual correction.
DE
DT Editorial Team··via arstechnica.com