While large language models struggle with tasks such as accurately counting digits or proposing meal ideas like pizza dishes, they excel in processing vast datasets to uncover subtle links. This capability positions them ideally for revealing the authors of anonymous online content, as outlined in a recent study.
A team from ETH Zurich and the MATS program linked to Berkeley developed a tool that gathered information from platforms featuring pseudonymous handles, such as Reddit. The approach involved assembling contributions from users in separate film-focused communities, then inputting this alongside details from a Netflix security breach into the model. This enabled the system to match those pseudonyms to actual individuals and their verified identities.
Sharing a single film suggestion on Reddit allowed the model to connect 3.1 percent of pseudonymous profiles to a particular Netflix user profile at 90 percent reliability. When five to nine suggestions were involved, identification rates rose to 23.2 percent. For more than 10 suggestions, the success reached 48.1 percent, including 17 percent matched with extremely high assurance.
In a separate test, the researchers linked pseudonymous profiles on Hacker News—a discussion platform without malicious intent—to verified professional details on LinkedIn. Over time, brief posts containing broad details like age, location, occupation, and similar elements enabled the model to confirm real identities with strong reliability. Although not applicable to all cases and achievable manually by investigators or enthusiasts, the speed and breadth of the automated method stand out.
A particularly revealing demonstration involved a brief 10-minute survey conducted anonymously by a researcher from Anthropic. Among 125 participants, 7 percent were uniquely identified through their written responses, revealing inferred details such as employment in fields like biological research, academic background, specialized software, and even regional language variations like British spellings for words such as 'analysing.'
The findings do not suggest that every anonymous user on any platform faces inevitable exposure. However, disclosing even seemingly innocuous personal facts heightens the risk of identification—a concern long present in online spaces. Peer-to-peer unmasking, official probes, and other forms of surveillance have existed since the internet's inception.
Yet, the shift toward automated tools that scan the internet for solid ties between pseudonymous and identified content introduces fresh threats to online anonymity. Although social platforms have moved beyond simple aliases, pseudonymous forums like those on Reddit remain vital, particularly for marginalized or at-risk communities. The study notes that 'deanonymization represents one of several mechanisms through which large language models aid illicit actors and authorities alike.'
According to coverage by Ars Technica, the study proposes countermeasures to reduce exposure. Sites such as Reddit could impose tighter restrictions on artificial intelligence access to user data via application programming interfaces, while developers of these models might track usage patterns to spot large-scale identification efforts.
Ultimately, the most effective strategy to avoid linking personal information to a pseudonymous profile is to refrain from sharing such details publicly in the first place.