A repeated click is not a standing preference
A theory paper asks what an AI can learn from a person’s actions. Under its math model, the same action could come from different preferences or from a workaround. Watching a click alone may not tell the tool which is true. The authors do not test a live fix with users. Before saving a rule for later, ask the person why they chose that step.
What the source reports
This paper uses a math model to study what an AI can infer from a person’s actions. The authors show that one pattern of visible choices can fit more than one hidden preference. A person may click a button because it is what they want. They may also click it because another route was blocked. An AI that saves the click as a lasting preference could get the next task wrong. The paper is a theory result. It does not test a new prompt or tool with users. The practical caution is to ask why a choice was made before turning it into a default. The model alone cannot prove that doing so will improve a real job.
Original source
- Title
- From Observation to Intervention: Identifiability of Latent User Bias in Human-AI Interaction
- Author
- Lucas Biechy, Yuxiao Li, Zhonghao He, Tianyi Qiu
- Publication
- arXiv
- Date
- Friday, October 9, 2026