The internet is rife with anonymous accounts as users adopt pseudonyms, sometimes for genuine reasons like speaking freely, and other times for nefarious ones. But this era of online privacy could be coming to a close. In a study available on the arXiv preprint server, researchers demonstrate that large language models (LLMs) can identify the people behind these accounts at scale.
For years, there has been an assumption that if deanonymization were theoretically possible, it would be too time-consuming, difficult, and expensive for people to perform. But the study authors had a hunch that LLMs had become powerful enough to break online invisibility.
To test whether this was the case, the team designed an automated framework to replicate a human investigator’s decision-making process.
Finding anonymous users
First, the AI reads through a user’s post history on either Reddit or Hacker News, examining unstructured text. This is raw, unorganized information like comments, jokes, education, and subtle writing quirks. It then turned this micro-data into a mathematical representation of the person’s profile to find candidate matches across millions of other profiles on the open web or on separate sites like LinkedIn.
When the AI found possible matches, it weighed up the evidence that both profiles belonged to the same person. Then it assigned a confidence score to its predicted match. If the LLM wasn’t sure, it didn’t write anything. This helped ensure it was not making wild guesses.
The researchers tested their framework on nearly 1,000 LinkedIn profiles to see if it could match them to accounts on Hacker News. These were profiles where the real-world identity was known to the team, who removed names, links, and other obvious identifiers from the bios.
Successful matches
The AI-powered framework successfully linked accounts with up to 67% accuracy at 90% precision, whereas the best non-AI methods struggled to succeed. It was also able to match individuals across Reddit communities, even if those users spread their activities across different accounts and time periods. The researchers also found that user identification is cheap, costing only $1 to $4 in computing power per account successfully linked.
“The practical obscurity that has long protected pseudonymous users… no longer holds,” wrote the researchers in their paper.
“Pseudonymity does not provide meaningful protection online. Users who post under persistent usernames should assume that adversaries can link their accounts to real identities or to each other, and that the probability rises with each piece of micro-data they post.”
The results show that, if further developed, this system could find applications in numerous fields, such as law enforcement and cybersecurity.
Written for you by our author Paul Arnold, edited by Gaby Clark, —this article is the result of careful human work. We rely on readers like you to keep independent science journalism alive.
If this reporting matters to you, please consider a donation (especially monthly). You’ll get an ad-free account as a thank-you.
Publication details
Simon Lermen et al, Large-scale online deanonymization with LLMs, arXiv (2026). DOI: 10.48550/arxiv.2602.16800
Journal information:
arXiv
Key concepts
The content is provided for information purposes only.