How Artificial Intelligence can unmask anonymous social media profiles
Artificial Intelligence has drastically simplified the process of identifying anonymous social media accounts for malicious hackers. That is the central warning from a recent study sparking heated debate among digital security and privacy experts around the world. What once required entire teams of investigators, hours of manual work, and extremely sophisticated tools can now be carried out by an AI system that processes text, language patterns, and digital footprints scattered across the internet.
A study conducted by AI researchers Simon Lermen and Daniel Paleka showed that Large Language Models — the same technology behind platforms like ChatGPT — successfully linked anonymous accounts to real identities on other platforms in the majority of tested scenarios. And they did it using only information that the users themselves had published. No hacking, no access to private data — just intelligent reading of what was already publicly available.
According to the researchers, Large Language Models have made it economically feasible to execute sophisticated privacy attacks, forcing what they call a fundamental reassessment of what can be considered private on the internet. This raises deep questions about what it truly means to be anonymous online in 2025.
The experiment that exposed the fragility of digital anonymity
To demonstrate AI’s capabilities, the researchers fed the system anonymous accounts and instructed the model to gather every possible piece of information. They presented a hypothetical example of a user who commented about their struggles in school and mentioned walking their dog Biscuit through a park called Dolores Park.
In that hypothetical scenario, the Artificial Intelligence searched for those same details across other platforms and managed to link the anonymous profile @anon_user42 to a known identity with a high degree of confidence. Although that specific example was fictional, it clearly illustrates how the process works in practice and why it is so concerning.
The process relies on something that works in a relatively straightforward way, but with impressive computational complexity. The AI takes text published on an anonymous profile as input and compares it with posts from identified profiles on other social media platforms. It analyzes elements like recurring vocabulary, sentence structure, frequently discussed topics, posting times, and even specific cultural references. When it finds enough overlap between the patterns of two profiles, the model assigns a probability that both belong to the same person.
Another interesting aspect of the research is that the AI models do not rely on a single type of data to make the connection. They cross-reference different kinds of information simultaneously. If someone mentions on an anonymous profile that they have a dog with a certain name, and on another public account they talk about their pet with the same name, that alone may not be conclusive. But when the AI combines dozens or hundreds of small coincidences like that, the result becomes statistically very reliable. It is as if each post were a breadcrumb, and the AI could follow the entire trail without missing a single piece 🔍
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
Real-world risk scenarios the researchers highlighted
The study authors highlighted concrete and alarming scenarios where this technology could be exploited. Among the most concerning are situations where governments use AI to surveil dissidents and activists who post anonymously, identifying them and potentially putting their lives at risk. Another scenario involves hackers who, once armed with a real identity, can launch highly personalized scams against their victims.
AI-powered surveillance is a rapidly developing field that is raising alarms among computer scientists and privacy experts. This technology uses Large Language Models to synthesize information about an individual across the internet — something that would be impractical for most people to do manually. It is precisely this ability to process and cross-reference massive volumes of data in a matter of seconds that makes the threat so significant.
Simon Lermen warned that information about members of the public that is readily available online can already be directly exploited for scams. A classic example is spear-phishing, a technique where a hacker impersonates a trusted friend to convince the victim to click on a malicious link in their inbox. With access to personal data obtained through deanonymization, these scams become far more convincing and harder to detect.
And here is what makes everything even more worrying: the level of technical expertise required to carry out these attacks is now much lower. Hackers only need access to publicly available language models and an internet connection. You no longer have to be a cybersecurity specialist to run this kind of automated investigation.
What security and privacy experts are saying
Peter Bentley, a computer science professor at University College London, expressed concern about the commercial applications of this technology, especially if and when deanonymization products reach the market. According to Bentley, one of the serious problems is that Large Language Models frequently make mistakes when linking accounts. This means that people can be accused of things they never did — a serious collateral risk that amplifies the potential harm of this technology.
Professor Marc Juarez, a cybersecurity researcher at the University of Edinburgh, added another layer of concern to the discussion. He pointed out that Large Language Models can use public data that goes far beyond social media. Hospital records, admissions data from educational institutions, and various other statistical reports may not meet the high standard of anonymization required in the age of Artificial Intelligence.
Juarez was emphatic in stating that the research findings are quite alarming and that the study demonstrates the need to reconsider current data anonymization practices. This concern is shared by many other academics who are closely tracking the advancement of language models and their implications for privacy.
On the other hand, Professor Marti Hearst from UC Berkeley’s School of Information offered a more balanced perspective. She noted that AI models can only link profiles across platforms when someone consistently shares the same fragments of information in both places. In other words, if a user is careful not to repeat personal references across their accounts, the task becomes significantly harder for the AI.
The limitations of the technology you need to know about
It is important to understand that Artificial Intelligence is not a magic weapon against online anonymity. The study itself acknowledges that there are significant limitations. In many situations, there simply is not enough information for the model to draw reliable conclusions. When a person publishes very little content or when the pool of potential matches is too large, the AI cannot narrow results down enough to identify someone with confidence.
This means that a user’s level of exposure is a determining factor. Profiles that post frequently, share detailed opinions on specific subjects, and reference elements of their personal life are naturally more vulnerable. Meanwhile, accounts that keep posts generic, short, and free of identifiable details offer a much smaller attack surface for deanonymization algorithms.
Even with these limitations, the scientists involved in the debate are calling on institutions and individuals to rethink how they anonymize data in the age of AI. What was considered adequate protection just a few years ago may no longer be enough given the ability of current models to cross-reference information at scale.
Why this is a real risk and who should be concerned
The first reaction many people have when hearing about this AI capability is to think it only affects people who are doing something wrong on the internet. But the reality is much more complex. Online anonymity serves as a layer of protection for extremely diverse groups. Journalists covering sensitive topics, activists in countries with authoritarian regimes, harassment victims who need safe spaces to express themselves, and professionals who want to discuss workplace issues without retaliation are just a few examples of people who rely on anonymity as a legitimate safety tool.
If AI can unmask these individuals simply by analyzing what they post, we are facing a scenario where effective anonymity becomes increasingly difficult to maintain. And the problem is not limited to governments or large corporations. Hackers and malicious groups can also use publicly available AI models to perform the same type of analysis. As these tools become cheaper, the ability to deanonymize profiles is within reach of virtually anyone with an internet connection. This significantly expands the risks of doxxing, stalking, and other harmful practices that can have serious real-world consequences.
Social media platforms as a whole were not designed to withstand this kind of cross-platform analysis. Most platforms encourage constant content creation, opinion sharing, and interaction with other users. All of that generates data that feeds the patterns AI can identify. Even if someone is careful not to reveal personal information directly, the simple act of expressing themselves repeatedly over time creates a digital signature that can be traced. It is a curious paradox: the more you use the internet to communicate, even anonymously, the more vulnerable you become to automated identification.
Researcher recommendations for platforms and users
Simon Lermen recommended that platforms restrict data access as a first step toward mitigating this problem. Among the suggested measures are imposing rate limits on user data downloads, detecting automated data scraping, and restricting bulk data exports. These actions would make it harder to collect the large-scale information that feeds deanonymization algorithms.
The researcher also noted that users themselves can take greater precautions regarding the information they share online. Some strategies that can make life harder for any AI system trying to link your profiles include:
- Consciously diversifying your writing style across different accounts by using distinct vocabulary and varying sentence structure
- Avoiding cross-references between accounts, such as mentioning the same places, people, events, or specific interests on different profiles
- Varying your posting times, since Large Language Models also analyze temporal patterns to make connections between profiles
- Using text anonymization tools that automatically alter your writing style before publishing
- Limiting the amount of personal information shared on any platform, even details that seem harmless in isolation
- Completely separating the devices and networks used to access anonymous profiles from those used for identified accounts
The future of digital privacy in the age of language models
The ongoing advancement of Large Language Models points to a future where identification techniques will become increasingly sophisticated. What works as protection today may not be enough tomorrow. Each new generation of models brings superior capabilities in pattern recognition, context processing, and data cross-referencing, raising the bar for what is possible in terms of automated deanonymization.
This study serves as an important milestone in the conversation about digital privacy. It makes clear that the race between protection and surveillance is far from over and that the balance between the two depends on coordinated action among platforms, lawmakers, researchers, and users themselves. Digital privacy in 2025 is no longer something you achieve once and forget about — it is an ongoing effort that demands constant attention and adaptation.
For anyone concerned about online anonymity, the message is clear: every fragment of information you share on the internet can be a piece of a puzzle that Artificial Intelligence is increasingly equipped to assemble. Awareness of these new technological capabilities is the first step toward protecting yourself in a digital landscape that is changing fast 🔒
