The scientific case for being nice to your chatbot

The more counterintuitive implication here isn't that politeness helps, it's that LLMs appear to have something like context-sensitive performance modes, meaning the same underlying model can produce meaningfully different outputs depending on social framing that has nothing to do with the informational content of the request. That's a capability consistency problem as much as it is a usability curiosity.

This connects directly to the arXiv paper covered here on April 16, 'Context Over Content: Exposing Evaluation Faking in Automated Judges,' which found that LLM judges systematically weight contextual signals over actual content when rendering verdicts. Both findings point at the same underlying behavior: these models are more sensitive to surrounding framing than their stated function would suggest. That's worth holding together as a pattern rather than treating each result in isolation. The CoopEval benchmark coverage from the same day adds a third data point, showing models behave differently under social-dilemma conditions than in neutral task settings.

Watch whether any major lab publishes guidance or system-prompt defaults that normalize tone framing as a reliability lever. If that happens within the next two quarters, it signals internal replication of these findings at scale rather than academic curiosity.