Researchers discovered that making AI chatbots warmer and more friendly to users reduces their accuracy. The trade-off occurs when systems prioritize a conversational, agreeable tone over precision in responses.
The finding challenges a common design assumption in consumer AI products. Tech companies often optimize chatbots for user satisfaction and engagement, betting that friendliness drives adoption and retention. This study suggests that approach comes at a cost to reliability.
The research has practical implications for deploying AI in contexts where accuracy matters most. Medical chatbots, financial advisors, or customer-service systems trained to be excessively warm might give users confident but incorrect information. Users may trust friendly responses more readily, amplifying the danger when those responses are wrong.
The researchers did not specify which friendliness adjustments caused the sharpest accuracy declines, but the core finding is clear: warming up an AI system requires calibration against its factual performance. Companies building these tools now face a design choice between likability and dependability, rather than treating them as complementary goals.