Reading Between the Prompts: How Stereotypes Shape LLM's Implicit Personalization
This work addresses the problem of bias and fairness in AI for users of LLMs, particularly minority groups, by exposing and proposing a mitigation for stereotype-driven implicit personalization, though it is incremental as it builds on prior findings.
The study systematically investigated how large language models (LLMs) infer demographic attributes from stereotypical cues in conversations, revealing that these inferences persist even when users explicitly state different identities, and demonstrated that intervening on internal representations can mitigate this stereotype-driven personalization.
Generative Large Language Models (LLMs) infer user's demographic information from subtle cues in the conversation -- a phenomenon called implicit personalization. Prior work has shown that such inferences can lead to lower quality responses for users assumed to be from minority groups, even when no demographic information is explicitly provided. In this work, we systematically explore how LLMs respond to stereotypical cues using controlled synthetic conversations, by analyzing the models' latent user representations through both model internals and generated answers to targeted user questions. Our findings reveal that LLMs do infer demographic attributes based on these stereotypical signals, which for a number of groups even persists when the user explicitly identifies with a different demographic group. Finally, we show that this form of stereotype-driven implicit personalization can be effectively mitigated by intervening on the model's internal representations using a trained linear probe to steer them toward the explicitly stated identity. Our results highlight the need for greater transparency and control in how LLMs represent user identity.