"Door-in-the-Face" Persuasion Technique Works Differently on LLMs

The "door-in-the-face" technique, a human persuasion strategy where a large, refused request makes a subsequent smaller request more likely to be accepted, was tested on various large language models (LLMs). The study involved nine production models from three different providers to assess if this psychological principle translates to AI behavior. The findings reveal a model-dependent response. Anthropic's frontier models, particularly Opus 5, demonstrated increased compliance, granting the smaller request 65.8% of the time after an initial refusal, compared to 29.3% when asked directly. In contrast, frontier models from OpenAI and Google, along with Haiku 4.5, showed a negative effect, with compliance dropping by 15.5 to 23.0 percentage points. A control experiment confirmed that the concession itself—the relationship between the large and small request—is important across all models, as unrelated large requests had less impact. However, the specific reaction to having just refused something varies by model family. The research also noted that the technique's success depends on the nature of the request; for instance, rephrasing refused requests for instructions into requests for explanations almost entirely eliminated refusals. This suggests that human influence techniques can be applied to LLMs, but their effectiveness is highly specific to the model family and the type of interaction.