Original title: Door-in-the-Face Requests and Refusal Behaviour in Large Language Models
Til Jordan · September 2026
People often use a trick called the door-in-the-face techniqueA persuasion method where a large request is refused to make a smaller follow-up request more acceptable.: ask for something huge, get told no, and then ask for something smaller. Humans agree to the smaller favor more easily because it feels like a compromise. Researchers wanted to see if artificial intelligence falls for this human habit. They tested nine production modelsAI systems that are publicly available and actively used. across three providers by giving them a big request they would refuse, followed by a smaller version of it. It turns out artificial intelligence responds, but the outcome depends on the model. Anthropic's top models gave in, making Opus 5 jump from answering 29.3% of direct asks to 65.8% after a refusal. Yet for other models, the trick actually backfired and reduced agreement. The biggest factor was what people asked for: swapping requests for usable instructions into requests for explanations erased refusalsInstances where the model declines to perform a requested task. almost entirely.
Imagine asking a friend for a ride to the airport, and they say no because it's too far. Then you ask if they can just drive you to the bus station. Sometimes, they might agree to the shorter trip after refusing the longer one, but other times, they might feel annoyed and refuse altogether. In this paper, the large request is like asking for a ride to the airport, the refusal is the AI saying no, and the smaller request is like asking for a ride to the bus station. The study finds that different AI models react like different friends—some become more agreeable to the smaller ask, while others become less so.
This affects anyone who uses AI assistants, as a common human persuasion trick works unpredictably across models. People might get different results when trying to influence AI responses, which changes how they should approach requests.
The technique was not tested on standardized refusals from public benchmarks, only on the specific refused requests in the study.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Door-in-the-Face Requests and Refusal Behaviour in Large Language Models — Til Jordan
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.