A new study reveals AI can identify anonymous online accounts, raising concerns about online privacy. Researchers from ETH Zurich and Anthropic achieved 90% accuracy.
Summary
- A team from ETH Zurich, MATS, and Anthropic has created an AI that can link anonymous accounts on platforms like Hacker News to real LinkedIn profiles with 90% accuracy.
- This process utilizes web searches and advanced reasoning models, costing about $1 to $4 per identity—without hacking.
- The study's authors have kept their methods confidential and confirmed the research was ethically reviewed by ETH Zurich before publication.
A recent study has raised concerns about online anonymity, showing that AI can effectively reveal the identities behind anonymous accounts. The research has gained traction on platforms such as X and Reddit, suggesting that the days of true online anonymity may be numbered.
The paper, titled "Large-scale online deanonymization with LLMs," was authored by researchers from ETH Zurich, the AI safety group MATS, and Nicholas Carlini from Anthropic, known for developing the AI system Claude.
The study demonstrates that large language models, which power AI tools like Claude and ChatGPT, can analyze anonymous posts to identify users based on publicly available information. This method does not involve hacking; rather, it relies on conventional web browsing and search techniques.
Researchers have developed an AI that can identify any "anonymous" Reddit user in under a minute for just $2.
ETH Zurich and Anthropic's paper suggests that the concept of "practical anonymity" online is becoming obsolete.
They created a fully autonomous AI pipeline to process user data efficiently.
— Superman (@thesupermannx) September 23, 2026
The findings indicate that the protective obscurity of pseudonymous users is diminishing, necessitating a reevaluation of privacy threat models.
How the AI Works
The AI operates through a four-step process: Extract, Search, Reason, and Calibrate. In the first step, “Extract,” the AI summarizes a user's posts to gather insights about their identity, including location and interests.
Next, in “Search,” the AI creates a numerical representation of this summary to find similar profiles among a large database. The “Reason” phase involves analyzing the best matches to determine the most likely identity, mimicking a human investigator's approach but at a much faster rate. Finally, “Calibrate” assesses the AI’s confidence in its guesses before finalizing any identity.
In an experiment, the researchers evaluated 338 users from Hacker News who had linked their LinkedIn profiles. The AI successfully identified 226 of them, yielding a 67% accuracy rate. In another test, it identified nine individuals from transcripts of discussions with 125 scientists based solely on their work descriptions.
Implications of the Research
The study highlights that the cost of conducting such searches is minimal, ranging from $1 to $4 in subscription fees. This method does not require data breaches or hacking; it simply uses existing web search capabilities, making it difficult to combat.
Historically, minimal information has led to user identifications, as seen in a 2008 Netflix study. However, the current AI can analyze unstructured data, such as casual comments, much more efficiently.
Addressing Privacy Concerns
While the findings are concerning, the study involved known identities, indicating that the results may not universally apply to all anonymous accounts. The challenges of scaling also mean that identifying someone from a larger group is more complex.
Moreover, the researchers did not disclose their methodologies or identities uncovered, emphasizing that they are documenting existing AI capabilities rather than providing tools for doxxing.
The implications extend beyond casual users; those involved in sensitive areas like cryptocurrency are already aware of the risks posed by AI-enabled deanonymization. Past incidents, such as the 2025 Coinbase breach, have demonstrated the real-world consequences of identity exposure.
For individuals who use pseudonyms for safety, such as activists or those with privacy concerns, the accumulation of seemingly trivial information can create a detailed profile. While this risk has always existed, AI makes the process faster and more cost-effective.
To mitigate potential risks, users are encouraged to avoid sharing identifiable details and to adopt privacy-focused tools rather than succumb to panic.
