AI Risks Anonymity: Hackers & Surveillance Using LLMs Like ChatGPT

Researchers warn large language models can deanonymize online users with alarming accuracy

Published on Mar. 8, 2026

Got story updates? Submit your updates here. ›

A new study by AI researchers Simon Lermen and Daniel Paleka has found that large language models (LLMs) like ChatGPT can effectively 'deanonymize' online users by correlating their posts with information available elsewhere on the internet. The researchers demonstrated that LLMs can match pseudonymous profiles to real-world identities based on subtle clues, raising concerns about increased risks of personalized scams and government surveillance of dissidents and activists.

Why it matters

The research highlights a fundamental shift in online privacy, as sophisticated deanonymization attacks that previously required significant manual effort and expertise can now be carried out by malicious actors with access to publicly available language models. This dramatically lowers the barrier to entry for those seeking to unmask anonymous users, with potentially serious consequences for individual privacy and security.

The details

The study, detailed in a pre-press paper titled 'Large-scale online deanonymization with LLMs,' reveals that LLMs can match pseudonymous profiles to real-world identities based on seemingly innocuous details. Researchers found that even subtle clues, such as a user mentioning struggles in school and a dog named Biscuit while referencing Dolores Park, were sufficient for the AI to identify the individual with a high degree of confidence. The technology isn't foolproof, however, as LLMs are prone to errors and can falsely link accounts, potentially leading to wrongful accusations.

  • The study was published in March 2026.

The players

Simon Lermen

An AI engineer at MATS Research and co-author of the study.

Daniel Paleka

Co-author of the study on deanonymizing online users using large language models.

Peter Bentley

A professor of computer science at UCL who warns that people may be falsely accused of things they haven't done due to the errors in the deanonymization technology.

Marti Hearst

A professor at UC Berkeley's school of information who explains that the effectiveness of deanonymization depends on the consistency of information shared across different platforms.

Marc Juárez

A cybersecurity lecturer at the University of Edinburgh who points to the vulnerability of seemingly anonymized datasets like hospital records and admissions data in the age of AI.

Got photos? Submit your photos here. ›

What they’re saying

“We show that LLM agents can figure out who you are from your anonymous online posts. Across Hacker News, Reddit, LinkedIn, and anonymized interview transcripts, our method identifies users with high precision – and scales to substantial populations.”

— Simon Lermen, AI engineer at MATS Research (archyde.com)

“People are going to be accused of things they haven't done.”

— Peter Bentley, Professor of computer science at UCL (archyde.com)

“It's quite alarming. I think this paper is showing that we should reconsider our practices.”

— Marc Juárez, Cybersecurity lecturer at the University of Edinburgh (archyde.com)

What’s next

Lermen recommends that platforms implement measures to restrict data access, including enforcing rate limits on user data downloads, detecting automated scraping, and limiting bulk data exports. He also emphasizes the importance of individual users being more cautious about the information they share online. As of today, no major social media platforms have publicly announced plans to alter their data access policies in response to the research.

The takeaway

The study's findings force a fundamental reassessment of what can be considered private online, as malicious actors can now leverage publicly available language models to deanonymize users with alarming accuracy. This raises serious concerns about the potential for increased personalized scams, government surveillance, and the need for platforms and individuals to reconsider their data sharing and privacy practices.