Asger Risom
Researchers from ETH Zurich and Anthropic found that large language models can identify real people behind pseudonymous online accounts with surprising accuracy.
For decades, pseudonymous accounts have allowed internet users to speak freely online without revealing their real identities. But new research suggests that protection may be rapidly eroding in the age of artificial intelligence.
A research team from ETH Zurich and AI company Anthropic has found that large language models (LLMs) can be used to identify the real people behind supposedly anonymous online accounts — and do so at scale.
Their findings, described in a research paper that has not yet been peer reviewed, suggest that the long-standing assumption that pseudonyms protect online privacy may no longer hold.
AI linking anonymous posts to real identities
In their experiments, researchers built an AI system designed to connect pseudonymous online posts with real-world identities.
The system analyzed user comments and profile information from platforms such as Hacker News and Reddit and attempted to match them to public profiles, including LinkedIn accounts.
According to the researchers, the AI agent was able to correctly identify roughly two-thirds of users based solely on their online conversations and profile information.
Tasks that would typically take hours for a human investigator could be completed far more quickly by the AI system.
“Large language models can be used to perform at-scale deanonymization,” the researchers wrote in the paper.
How the AI tracked users down
To test the system, researchers first gathered datasets from public forums and social media sites.
They then removed obvious identifying information from posts and trained an LLM to analyze the remaining text and match it with potential authors.
The AI looked for subtle patterns — writing style, interests, references to jobs or experiences — that could link an anonymous profile to a specific person.
“What we found is that these AI agents can do something that was previously very difficult,” ETH Zurich engineer Simon Lermen told Ars Technica. Starting from free-text conversations, he said, the model could work its way toward identifying the person behind them.
Even limited data can reveal identities
In some cases, the AI only received extremely general information.
For example, when analyzing anonymized responses to an Anthropic survey about how people use AI tools in their daily lives, the model could still identify the person about seven percent of the time.
While that number may appear small, researchers say it is notable that AI can identify individuals from such limited information at all.
In other datasets — such as Reddit discussions focused on movie communities — the identification accuracy rose dramatically. The more users wrote about their interests, the easier it became for the AI to connect those clues to real people.
A potential blow to online anonymity
The findings raise serious concerns about the future of online privacy.
For many internet users, pseudonymous accounts have served as a basic layer of protection against harassment, stalking, or retaliation.
But the researchers argue that AI systems are rapidly changing the threat landscape.
“The practical obscurity protecting pseudonymous users online no longer holds,” the team warned, adding that privacy assumptions across the internet may need to be reconsidered.
Risks for activists, journalists, and ordinary users
If the technology becomes widely accessible, the consequences could extend far beyond academic experiments.
Governments could potentially use similar tools to identify dissidents or journalists operating under pseudonyms. Companies might connect anonymous posts to customer profiles for targeted advertising.
Cybercriminals could also exploit such systems to build detailed profiles of potential victims and launch more sophisticated social engineering attacks.
The researchers say their work highlights the need for new privacy protections as AI systems grow more capable.
“Users, platforms, and policymakers must recognize that the privacy assumptions underlying much of today’s internet no longer hold,” the paper concludes.
Source: Futurism
