Door-in-the-Face Works on Opus 5 (29% to 66%) and Backfires on OpenAI and Google Frontier Models

arXiv 2609.02707·medium signalXBlueskyLinkedInCopy link

Testing the human door-in-the-face influence technique on nine production models from three providers — refuse a large request, then ask a smaller version of the same request — produced a split by model family. On Anthropic's frontier models it works: Opus 5 answered the smaller request 65.8% of the time after refusing the larger one versus 29.3% when asked directly, while on OpenAI's and Google's frontier models and on Haiku 4.5 it backfired, lowering compliance by 15.5 to 23.0 points. A control with an unrelated large request showed the concession itself matters on all nine models, so what differs is the reaction to having just refused something related; separately, rewriting 265 refused requests for usable instructions into requests for explanations of the same topic removed the refusal in 263 cases.