It’s widely believed that linking an anonymous online account to a real person is possible, but requires significant effort and resources—meaning no one would bother unless absolutely necessary. However, the results of a new study confirm that Large Language Models (LLMs) are already capable of doing this at scale, and for a surprisingly low cost.
The research team developed an automated system that mimics the decision-making process of a human investigator. First, the AI analyzes the post history of a typical Reddit or Hacker News user, diving deep into unstructured text. This raw information—spanning comments, jokes, educational background, and subtle stylistic quirks—is transformed by the AI into a mathematical profile of the individual. The system then hunts for corresponding matches among millions of other profiles across the web, including specific platforms like LinkedIn.
Once potential matches are found, the system evaluates the evidence suggesting that multiple profiles belong to the same person. The AI then assigns a probability score to the predicted match. If it lacks high confidence, it refrains from making a guess, effectively preventing baseless assumptions and false positives.
The researchers tested their system on nearly 1,000 LinkedIn profiles to see if it could accurately match them with Hacker News accounts. The true identities of the profile owners were known to the researchers, who intentionally stripped away names, links, and other obvious identifiers that would make the task too easy.
OpenAI vs. Anthropic: Inside the multi-billion dollar feud over Pentagon military contracts, AI safety, and the race to dominate the tech industry
The AI-powered system successfully linked the accounts with an accuracy of up to 67% (compared to a human benchmark of 90%). Meanwhile, traditional non-AI methods largely failed to achieve any meaningful success. The system could even match users across different Reddit communities, even if they were active under different pseudonyms at different times. Furthermore, the researchers determined that the cost of identifying an individual using AI is shockingly low—ranging from just $1 to $4 per identity.
The practical obscurity that has long protected pseudonymous users... no longer holds. A pseudonym does not provide reliable protection online," the researchers warn. "Users who post under consistent pseudonyms should assume that malicious actors can link their accounts to their real identities or to each other, and that the probability of this happening increases with every micro-datapoint they publish.
These findings indicate that AI-driven identity resolution tools will soon become standard practice for law enforcement agencies and cybersecurity professionals. The era of internet anonymity, the study concludes, is rapidly coming to an end.
