Study Finds AI Can Link Anonymous Accounts to Real Identities

A research paper from Anthropic and ETH Zurich indicates that large language models can help link pseudonymous online accounts to real people by combining clues scattered across public posts. The work shows how tasks that once required substantial manual investigation may be partially automated and performed at a larger scale.

The paper, Large-scale online deanonymization with LLMs, was released as an arXiv preprint. It had not completed peer review at the time described, and the experiments used controlled datasets rather than every condition found on the open internet. The results should therefore be treated as evidence of a privacy risk, not a claim that any anonymous account can always be identified.

How AI-assisted deanonymization works

A person may avoid using a real name but still reveal identifying details over time. Posts can contain information about location, occupation, education, interests, routines, age, or events in the person's life. Writing patterns and topic preferences can add further signals.

The researchers tested systems that extracted these clues from public text, searched for possible matching profiles, and ranked candidates according to how well their details aligned. A language model can help connect a collection of individually weak clues that a human analyst might otherwise need hours to compare.

What the experiments tested

One experiment attempted to match Hacker News users with LinkedIn profiles after obvious identifiers such as names and matching usernames had been removed. Other tests involved pseudonymous Reddit accounts and a dataset in which one person's posting history was divided into two profiles.

The paper reports that its language-model-based approach outperformed the traditional comparison methods used as baselines. In some experimental settings, the system reached recall of up to 68% at roughly 90% precision.

Those measures describe different things. A 68% recall means the system found about 68% of the matches that existed in that test set. Roughly 90% precision means about nine out of ten matches it proposed were correct under that evaluation. It does not mean the system identified 90% of all anonymous users, and results from a curated dataset may not transfer directly to unrelated platforms.

Why this changes the privacy risk

Pseudonyms protect journalists, whistleblowers, activists, abuse survivors, and ordinary users who want to discuss personal subjects without attaching them to a searchable real-world identity. That protection has never been absolute, but automated analysis can reduce the time and cost required to compare a profile with many possible identities.

The researchers estimated that their experimental workflow could process a profile for a few dollars. Even if the system makes errors, low-cost candidate generation can still enable surveillance or targeted harassment at a scale that manual investigation would not support.

False matches are also dangerous. A confident but incorrect identification could direct attention or retaliation toward an unrelated person. Any legitimate investigative use would require corroborating evidence, access controls, documented methods, and a process for contesting errors.

How users can reduce identifying clues

No single measure guarantees anonymity, but users can reduce linkable signals:

• Do not reuse usernames, profile photos, biographies, or distinctive phrases across identities.
• Avoid posting exact workplaces, schools, neighborhoods, schedules, or uncommon life events.
• Review old posts together; a sequence of harmless details can become identifying when combined.
• Strip location and device metadata from media before sharing it.
• Use separate contact information and avoid linking pseudonymous accounts to personal profiles.
• Delay or generalize time-sensitive details that reveal where you are.

Changing writing style alone is not a dependable defense because biographical and behavioral details may still overlap.

What platforms and researchers should address

Platforms can make privacy risks clearer, limit automated harvesting, and provide tools that warn users when posts contain identifying details. Researchers and AI providers should restrict access to systems designed for mass identification and evaluate not only accuracy but also misuse, bias, and false accusations.

The study used public information, but “public” does not necessarily mean people expected it to be aggregated and used for identification. As AI lowers the cost of connecting digital traces, meaningful online anonymity will increasingly depend on data minimization, platform design, and limits on automated surveillance.