The results of research conducted on this matter jointly by the Federal University of Technology in Zurich and Anthropic do not paint a very optimistic vision for supporters of anonymity. During the experiment, it was checked how artificial intelligence models, by searching for similarities in writing style, tendency to use individual words and other individual characteristics related to the text, cope with finding connections between anonymous posts and the public activities of individual people.
Although this method is in no way groundbreaking, and the digital “print” we generate in this way has been used many times in the past to identify individual people, the entire process usually took dozens or even hundreds of hours. And only to find a single person. The AI models used by Anthropic and the Swiss university needed significantly less time… and were able to analyze many cases at the same time.
The experiment attempted to discover the true identity of users of Reddit and Hacker News forums, both of which are known for theoretically full anonymization of their profiles. Users signing with pseudonyms do not have to provide any real data in order to participate in discussions and start new threads. Although the research focused on international online communities, translating it to the Polish backyard – You could just as easily analyze the posts of Wykop users, the Murator forum or Elektroda in this way. Popular anonymous commentator accounts on X (formerly Twitter) or Facebook can also be identified in this way.
See also: There is no such thing as a private conversation with AI. We are preparing a disaster for ourselves
The research by Anthropic and the Technical University of Zurich resulted in the following results – nearly 70 percent the creators of anonymous entries were correctly assigned to a real person and his LinkedIn profile. Additional support for identification was provided by anonymized conversations conducted by the experiment participants, which served as additional input data fed to the AI agents.
The authors of the study openly admitted: – The results of our research show that anonymity protecting users using pseudonyms on the Internet no longer exists.
Simon Lermen, one of the engineers working on the experiment, explained how the whole process works in a simplified way: – We found that thanks to combining search with inferencewe are able to identify most accounts with high precision. We use a search based on embeddings (i.e. on the analysis of context and meaning connections – editor's note) to select the 100 most promising candidates, and then we use inference to select and verify the right one.
In practice, this means that the current assumption that the lack of a signature under the post or writing it from an account unrelated to our data ensures anonymity, it no longer makes sense. The AI agents used in the study turned out to have another advantage over human analysis – they dealt with text without a specific, unified structure much more efficiently and quickly. Even entries and responses that were very general and more chaotic were properly sorted by artificial intelligence and used throughout the process. Previously, effective identification rather required two similar sets of content that could be properly combined. For AI, text structure does not matter much when searching for similarities.
Importantly, standard security measures that “lose” traces of activity – such as the use of VPNs – also have no significance in this case. Artificial intelligence does not examine any metadata or technical parameters related to the location or output device – the analysis only covers the content, the tone of which is not influenced by where and on what a given post is written.
In connection with their findings, the authors of the study warn that further development of technology may mean the complete abolition of the concept of anonymity on the Internet, and in the hands of individual governments such tools may lead to deeper repression and restrictions on freedom of speech. This sentiment is strongly consistent with recent statements made by Anthropic's president, Dario Amodei, in his famous essay, published just before breaking off cooperation with the Pentagon. There were also warnings against mass surveillance and analysis of citizens' online activities on a mass scale.
Of course, AI agents dedicated to identifying anonymous users are not an “attractive” prospect only for state entities. The discussed research also indicates that corporations and cybercriminals may also be eager to use this type of tools. In the first case – for even better profiling and targeting of users in terms of served ads and offers. In the case of criminal activity, obtaining information about anonymous user activity may, in turn, facilitate social engineering attacks.
Can users defend themselves against this in any way? Theoretically… yes, although it requires a complete change in thinking about how we use the Internet. De facto – completely refrain from publishing any information that describes us anywhere. Simon Lermen from the Federal University of Technology in Zurich clarified it as follows: – It is worth developing the habit of being more vigilant when it comes to privacy. Each specific piece of information you share—your city, your job, your conference attendance, or your niche hobby—narrows the circle of people you could be. Their combination often constitutes a unique fingerprint.
And he sums it up by saying: – Ask yourself: Would a team of keen investigators be able to identify you based on your entries? If so, AI agents using LLMs will likely do the same, and the cost of doing so will only go down.
See also: Boycott ChatGPT. OpenAI's deal with the Pentagon has enraged users
The prospect that our privacy will be fully protected only if we limit our online activity is not very optimistic, but we need to start getting used to it. Or – which is much more likely for most people – we have to accept the fact that everything we write on the Internet will sooner or later be attributed to us. If not due to legal requirements, then through the use of AI agents, as in the study conducted by Anthropic.
