Reporting describes Searchable’s analysis of eight million voluntarily contributed AI conversations from September 2025 through September 2026. Conversations containing insults had a reported error-correction rate of 31%, compared with 41% for neutral conversations. This was not a causal experiment separating task difficulty and context, including the possibility that an initially difficult problem caused users to become rude. Rather than turning the association into a rule that politeness alone improves performance, the useful focus is providing reproducible inputs, expected results and specific descriptions of errors.
Back to Tech
Insults in AI conversations were associated with lower correction rates
This is not a causal experiment proving that rude language reduces performance.