AI learned to unmask anonymous social media accounts

You have an anonymous account on Reddit where you openly discuss your boss? Or a secret social media profile where you allow yourself to be genuine? New research shows that artificial intelligence is capable of linking these accounts to your real identity — and doing it faster and more effectively than any private detective.

A group of researchers from Swiss ETH Zurich, Anthropic, and the Machine Learning Alignment and Theory Scholars (MATS) program has developed an automated system of AI agents capable of deanonymizing internet users. The system uses unnamed language models that can independently search for information online, analyze the data found, and build chains of connections between disparate accounts. Essentially, it's a digital investigator that never gets tired and never loses the thread of reasoning.

To understand the scale of the problem, it's worth recalling how deanonymization used to work. Traditionally, revealing an anonymous account required either significant resources — like those of intelligence agencies or large corporations — or a glaring user error: a carelessly posted photo with geolocation, using the same nickname on different platforms, characteristic speech patterns. The process was labor-intensive, expensive, and didn't scale. One analyst could conduct one investigation for weeks. Artificial intelligence is changing this economy radically. What previously required a team of specialists and months of work, the AI agent system is potentially capable of accomplishing in hours or even minutes, and simultaneously for thousands of accounts.

Technically, the researchers' approach relies on the ability of modern language models to engage in what's called multi-step reasoning. An AI agent doesn't simply compare two texts for stylistic similarity — it builds complex logical chains. Mentioned on an anonymous account that you live near a particular park? Written on your main profile about your favorite café in the same district? Using a specific professional term in both places? Each such detail alone proves nothing, but collectively they form a unique digital fingerprint. Modern models excel precisely at this task — synthesizing disparate fragments of information into a coherent picture.

It's important to note that the research has not yet been peer-reviewed, and the authors don't disclose which specific models were used in the experiments. This means that concrete results should be approached with caution. Nevertheless, the very fact of publishing work involving Anthropic — a company that positions itself as a leader in AI safety — speaks volumes. The research was likely conducted as part of a risk assessment: before releasing increasingly powerful models, it's necessary to understand what threats they pose. And deanonymization is one of the most serious.

The implications of this discovery extend far beyond awkward situations when your boss learns about your anonymous Glassdoor review. Anonymity on the internet is not just a convenience, it's a fundamental tool of freedom of speech. Whistleblowers exposing corruption. Residents of authoritarian states criticizing the government. Victims of domestic violence seeking help. Activists in countries where dissent is persecuted. For all these people, anonymity is not a whim, but a matter of security, and sometimes of life itself. If AI deanonymization tools become widely available, the balance of power in the information space will shift in favor of those who already hold power — corporations, states, persecutors.

For the Russian audience, this topic has particular urgency. In conditions where anonymous online speech remains for many the only safe way to express an opinion, the emergence of accessible AI-based deanonymization tools creates an entirely new landscape of risks. Moreover, protecting oneself from such systems is significantly more difficult than protecting oneself from traditional methods: it's not enough to simply use a VPN and different usernames if your very writing style, set of interests, and behavioral patterns become a unique identifier.

The authors of the research rightfully note that burying anonymity is premature. Between a laboratory demonstration and mass application lies considerable distance. However, the direction of movement is clear: with each generation of language models, the task of deanonymization will become easier and cheaper. This means that industry, regulators, and users themselves need to seriously think about new privacy protection mechanisms — ones that account for the capabilities of AI, not just traditional threats. The era when anonymity was ensured simply by creating a new account with a fictitious name is coming to an end.