AI researchers warn that large language models such as those behind ChatGPT can make it much easier for malicious actors to identify anonymous social media accounts by matching posts to real-world identities.

The study by Simon Lermen and Daniel Paleka found that in most test scenarios, LLMs linked anonymous profiles to known accounts by scraping and cross-referencing publicly posted details.

In experiments, the researchers fed anonymous posts into LLMs which searched for unique publicly available details; a hypothetical example used a pet name and a neighbourhood to show how a match could be produced, and in most test scenarios the models produced high-confidence matches.

The paper says this lowers the cost and skill needed for privacy attacks and could enable state surveillance or highly personalized scams.

Peter Bentley of UCL said LLMs can make errors and falsely link people, warning that commercial de-anonymization tools could lead to false accusations. Marc Juárez of the University of Edinburgh warned that public datasets like hospital records may no longer be sufficiently anonymized.

Researchers recommend rate limits on data downloads, tools to detect automated scraping, and limits on bulk exports, and suggest users reduce personal details shared online.

🔗 Source: The Guardian