Agentic LLMs as Powerful Deanonymizers. Li 2026 (arxiv.org)

A recent study by Tianshi Li highlights the potential of agentic large language models (LLMs) to deanonymize subjects within the Anthropic Interviewer dataset, which comprises qualitative interviews with professionals discussing AI in research. The findings are significant as they reveal that publicly available LLMs can effectively identify six out of twenty-four interviews associated with specific scientific works, thereby revealing the identities of some interviewees. This exposes a critical vulnerability in handling qualitative data and underscores the risk of LLMs in breaking privacy safeguards using simple web searches and cross-referencing techniques. The implications for the AI/ML community are profound, as this research suggests that the technical barriers to deanonymization are being lowered, raising concerns about data privacy and ethics in AI applications. Li outlines how current measures to protect sensitive information can be circumvented and emphasizes the need for enhanced safeguards when releasing rich qualitative datasets in an era dominated by LLM technology. The study calls attention to the urgency of addressing these challenges and proposes strategies for mitigation while posing further questions about the responsible use of LLMs in research and data sharing practices.