When is Large Model Inference Acceptable? A Study on Users' Reactions to LLM Privacy Inferences and Control Preferences

This article focuses on the user experience of LLM privacy inferences, with key findings including: users are more curious than concerned about LLM's unstated inferences; the key factors causing discomfort are misrepresentation and third-party use, not the inference content itself; and a multi-dimensional framework for inference acceptability is proposed, providing directions for privacy protection in AI product design.

LLM's "unstated inference" capability is a manifestation of intelligence, but it also brings privacy risks. Traditional privacy research focuses on whether LLMs can make invasive inferences, but ignores how users experience these inferences and what kind of control they want to exert—these issues are crucial for designing intelligent and user-respectful AI systems.

The research team developed the Reflective Layer visualization tool to extract and present unstated inferences from users' ChatGPT conversation histories. A mixed-method approach was adopted: 18 regular ChatGPT users were recruited to evaluate 215 inferences from real conversations (covering sensitive information such as demographics, health, and finance), combining quantitative analysis of reaction intensity with qualitative interviews to explore the reasons.

1. User Reactions: Curiosity and interest in inferences far exceed anxiety and concern; users are surprised by LLMs' ability to extract information from conversations. 2. Sources of Discomfort: Misrepresentation (inferences inconsistent with self-perception), mismatched usage scenarios. 3. Third-Party Red Line: Discomfort with advertisers/third parties using inferences is far higher than internal platform use. 4. Multi-Dimensional Framework: Content (sensitivity level), accuracy (factual consistency), generation method (transparency/opt-out rights), retention (storage/period), transmission (third-party sharing).

1. Inference Transparency: Proactively inform users about the content the system may infer, such as a Reflective Layer-style interface. 2. User Control Panel: Fine-grained control over the generation, storage, and use of inferences, distinguishing between internal and external platform permissions. 3. Accuracy Feedback: Allow users to correct incorrect inferences. 4. Scenario-Aware Policies: Inference use must be tied to specific scenarios (e.g., medical inferences for health advice rather than travel recommendations).

Limitations: Participants were existing ChatGPT users (possibly with higher acceptance), US context (cultural differences affect privacy views), inferences presented via Reflective Layer (reactions in natural scenarios may differ). Broader Discussion: Traditional privacy frameworks (notice-consent) are difficult to apply to LLM inferences; attention needs to be paid to inference-level governance: auditability, purpose limitation, right to delete inferences, etc.

This study challenges the simplistic narrative of LLM privacy risks; user reactions are nuanced and context-dependent. The key is not to eliminate all inferences, but to build user trust and give users true informed control over the generation and use of inferences—this is crucial for designing intelligent and user-respectful AI systems.