Researchers found AI can identify real identities behind pseudonyms in 2/3 of cases. Study warns internet anonymity is at risk as models analyze digital footprints.
© RusPhotoBank
Researchers from the Swiss Federal Institute of Technology Zurich, in collaboration with specialists from Anthropic, have claimed that modern large language models can already identify the real identities of users hiding behind pseudonyms on a mass scale. According to a study that has not yet undergone peer review, the conventional "protection of anonymity" on the internet is rapidly losing its effectiveness.
In an experiment, scientists created an AI agent that analyzed user publications and profiles on forums, cross-referencing them with open data. Based on texts from platforms like Hacker News and Reddit, the system managed to determine real identities in about two-thirds of cases. Such work would take a human hours to complete manually, while the model performed it automatically and on a large scale.
The method proved effective even when matching different data sources, from forum comments to LinkedIn profiles and anonymous interviews. Even neutral texts, such as responses to surveys about AI usage, allowed the model in some instances to pinpoint specific individuals. The more actively a user shared opinions, the easier it was for the algorithm to narrow down the search.
The authors emphasize that they did not disclose the identities of real people and developed special mechanisms to safely evaluate the system's capabilities. Nevertheless, the study's conclusion is concerning: the previous "fuzziness" of digital footprints no longer guarantees protection. According to the scientists, users, platforms, and regulators need to reassess their understanding of privacy in the AI era, or internet anonymity could face serious threats.