The findings from a recently published research paper indicate that large language models (LLMs) can unmask pseudonymous users on social media platforms with surprising accuracy. The study demonstrated success rates as high as 68 percent in recall and up to 90 percent in precision, far exceeding existing classical deanonymization methods.
The researchers used a pseudonymous stripping framework that involved collecting datasets from public social media sites like Hacker News and LinkedIn profiles. By correlating specific individuals with accounts or posts across multiple platforms, they could link users even when identifying references were stripped from the content. A second dataset was derived from Netflix micro-identities, previously linked to political affiliations through structured data attacks.
In one experiment, LLMs were able to identify 7 percent of participants in a survey about AI usage by cross-referencing responses and web searches. This capability is particularly alarming as it means that sensitive personal information can be extracted from anonymized interviews or discussions.