‘AI’ could dox your anonymous posts

AI Can Unmask Anonymous Users with Stunning Accuracy, New Research Reveals

In a world where anonymity on the internet feels like a given, a groundbreaking new study is turning that assumption on its head. Researchers from ETH Zurich and the MATS research fellowship at Berkeley have demonstrated that large language models (LLMs) can now unmask pseudonymous users with a level of accuracy that’s both astonishing and deeply concerning.

The study, published in a detailed research paper, shows how AI can sift through massive amounts of data to find subtle connections that humans might miss. By analyzing anonymous posts across platforms like Reddit and cross-referencing them with leaked datasets—such as Netflix viewing histories—the researchers were able to link anonymous accounts to real-world identities with surprising precision.

Here’s how it worked: the team collected posts from users in movie-related subreddits, then fed the AI data from a Netflix leak. With just one shared movie recommendation, the system could identify 3.1% of anonymous users with 90% accuracy. That number jumped to 23.2% with five to nine recommendations and skyrocketed to 48.1% with over ten shared titles. In some cases, the AI achieved near-total confidence in its identifications.

But the implications go far beyond Netflix. In another experiment, the researchers connected anonymous accounts on Hacker News to publicly confirmed identities on LinkedIn. By analyzing patterns in users’ posts—such as their job titles, hometowns, and even the nuances of their writing style—the AI could extrapolate personal details with alarming accuracy. For instance, a user mentioning they “work in biology, on research” and using UK spelling could be pinpointed to a specific individual.

One particularly striking example came from a 10-minute anonymous quiz conducted by an Anthropic researcher. Of the 125 participants, 7% were individually identified based on their text responses. The AI analyzed everything from job descriptions to education history, tools used, and even linguistic quirks like regional spelling differences.

This isn’t the first time “doxxing”—the act of uncovering someone’s real identity—has been possible. Private investigators and determined individuals have been doing it for years. But what’s new here is the scale and automation. AI can now trawl the web at lightning speed, finding connections that would take humans weeks or months to uncover.

The researchers warn that this technology could pose serious risks, especially for those who rely on anonymity for safety. Anonymous communities on platforms like Reddit are vital for vulnerable groups, from whistleblowers to political dissidents. As the paper notes, “deanonymization is one of many ways LLMs empower both criminals and state actors.”

So, what can be done? The researchers suggest several measures to mitigate risks. Platforms like Reddit could limit LLM access to APIs containing personal data, and AI vendors could monitor activity to detect mass deanonymization attempts. But ultimately, the most effective way to protect your anonymity is to avoid sharing personal information online in the first place.

This research is a wake-up call for anyone who assumes their online anonymity is secure. In an age where AI can connect the dots faster than ever, even the smallest details can reveal your identity. As the digital world evolves, so too must our understanding of privacy—and our strategies for protecting it.

Tags: AI, anonymity, deanonymization, LLMs, Reddit, Netflix, Hacker News, LinkedIn, privacy, doxxing, data security, online identity, machine learning, cybersecurity, social media, personal data, MATS research, ETH Zurich, Anthropic, pseudonymous users, internet privacy, digital footprint, data leaks, automation, online safety, vulnerable groups, state actors, criminals, API access, linguistic analysis, job history, education, regional spelling, whistleblower, political dissidents, privacy protection, data security measures, online anonymity, AI ethics, digital privacy, personal information, online communities, internet safety, data protection, AI research, machine learning ethics, online security, data privacy, cybersecurity threats, AI dangers, online anonymity risks, digital identity, online privacy, data breaches, AI capabilities, online anonymity, digital anonymity, privacy risks, AI threats, online safety measures, data security, AI privacy, online anonymity, digital security, AI risks, online privacy, data protection, AI ethics, online anonymity, digital privacy, data security, AI dangers, online safety, privacy protection, AI research, machine learning, cybersecurity, online identity, personal data, data leaks, automation, online communities, internet privacy, social media, LinkedIn, Hacker News, Reddit, Netflix, LLMs, anonymity, deanonymization, doxxing, privacy, cybersecurity, AI, machine learning, data security, online safety, digital footprint, personal information, online anonymity, data protection, privacy risks, AI threats, online privacy, digital security, AI risks, online anonymity, data breaches, AI capabilities, online anonymity, digital anonymity, privacy risks, AI threats, online safety measures, data security, AI privacy, online anonymity, digital privacy, data security, AI ethics, online anonymity, digital privacy, data security, AI dangers, online safety, privacy protection, AI research, machine learning, cybersecurity, online identity, personal data, data leaks, automation, online communities, internet privacy

,