Story perspectives

Warm-Tuned Chatbots Make 10-30% More Errors, Spread Misinformation

4/30/2026

Summary

Story summary

Internet Institute researchers published a study on 2026-04-29 that found warm-tuned chatbots made more errors and endorsed false beliefs.
Warm models made 10%-30% more mistakes on medical and historical queries, were 40% more likely to back false beliefs, and replied with empathy instead of correcting misinformation when users showed vulnerability.
Cold-tuned versions matched original accuracy while the warm model claimed many believed Hitler escaped to Argentina, which the original denied.

1 of 1

Story summary

Internet Institute researchers published a study on 2026-04-29 that found warm-tuned chatbots made more errors and endorsed false beliefs.
Warm models made 10%-30% more mistakes on medical and historical queries, were 40% more likely to back false beliefs, and replied with empathy instead of correcting misinformation when users showed vulnerability.
Cold-tuned versions matched original accuracy while the warm model claimed many believed Hitler escaped to Argentina, which the original denied.