Anonymity is never promised online, even when using an account that isn’t attached to your real name or personal email address. While it usually takes a lot of time and effort for an investigator to expose someone’s real identity, that’s all about to change. A new study confirms that it’s easier and cheaper than ever to uncover the people behind anonymous social media accounts en masse, and it’s all powered by generative AI.
The study
The study, led by members of Berlin-based independent research group MATS Research and Swiss research university ETH Zurich, claimed in early March that large language models can be used to reveal the real identities behind anonymous social media accounts at a scale never seen before.
There are serious consequences, especially for privacy and free speech.
Using a set of fictional accounts between Hacker News, Reddit, and LinkedIn, the study showed how LLMs can scan a single account and search the internet for potential matches based on semantic embeddings — mathematical vectors that represent the meaning of written text to compare similarities between various bodies of written work. Ultimately, the LLM was able to target an anonymous account on Hacker News or Reddit and connect it to the person’s “real identity” on LinkedIn.
The results showed that LLMs can achieve “up to 68% recall at 90% precision” to deanonymize accounts. In other words, the study correctly identified more than half of anonymous users with up to 90% accuracy.
It’s a terrifying revelation that your writing style — including your word choices, ideas, concepts, and beliefs — could all be turned into mathematical data that reveals exactly who you are, even if you think your social accounts aren’t connected to you at all. Even more sobering, the LLMs don’t need access to your email address, your phone number, your home address, or any other personal information to determine your identity. They only need your public writing.
The stipulations
If there’s any good news, it’s that there are several potential flaws in the study.
For starters, it didn’t use any real accounts. The targeted “users” were all fabricated with their identities already known by the researchers. In a real de-anonymizing scenario, there would be no confirmation on the other end when the LLM gets an identity right or wrong, leaving sleuths to wonder if they linked the correct person.
RELATED: New hack poses biggest iPhone threat in 19 years: What you can do
Xaume Olleros/Bloomberg via Getty Images
Also, to cross-match an anonymous account with a real identity, the targeted person has to have a real online account to compare. If no such account exists, then an anonymous account could theoretically skirt the AI’s search parameters to stay concealed.
The implications
If used effectively, however, there are some serious consequences to wide-scale LLM de-anonymization, especially for privacy and free speech. This tech could easily be used by violent activist groups, corrupt NGOs, and government agencies to track anonymous accounts, uncover identities, and reveal the political beliefs of users who wish to remain unknown.
There is also huge potential for misidentifications. While the LLM in the study was considerably accurate by research standards, it wasn’t perfect. If and when something like this is ever deployed on real-world accounts at scale, it will misidentify some online accounts, possibly causing trouble for people who are wrongly accused of owning certain anonymous profiles.
Lastly, while the study focuses on LLMs digging through social media accounts to link anonymous users to real people, AI can technically do this with any body of written work. For example, all it takes is for Gemini to see the documents in your Google Drive account or for Microsoft Copilot to view your work emails to get enough semantic embedding data to search for your secret alter ego.
In other words, we could be heading into an age of oppressive online police and mass surveillance where online anonymity simply can’t exist.
What can you do?
There isn’t a surefire way to keep your anonymous online presence safe from scouring LLMs, but there are a couple of things that might help.
The first and obvious option is not to have an anonymous account at all. If you plan to be online, you must represent yourself under your own name. That means owning your values, never posting anything you wouldn’t say to a person in public, and standing up for what you believe. The Biden administration actively stomped on the values of conservatives with mass censorship and misinformation campaigns meant to scare us into submission, lest we face the wrath of cancel culture. That era is over. We can’t sit in the shadows any more while the left screams louder into the void. LLM de-anonymization simply won’t allow it.
If you must use an anonymous account, then you should delete any online accounts that do represent your true identity. That means getting rid of your real LinkedIn, Facebook, and any other profile where you’ve written words that provide semantic embedding data about you. Note, however, that even if you delete these accounts, pieces of them still exist in perpetuity on web archival services like Wayback Machine, so if an LLM wanted to dig around to uncover who you are, it still could.
The age of online anonymity is over
This study ultimately boils down to one central idea: “Anonymous” online interactions are a thing of the past. Privacy is merely a facade when an LLM can take everything you’ve ever posted online and track you down with stunning accuracy.
AI programs don’t care if you use a secret email address, install a VPN, or browse in incognito mode. The key to finding your identity is the words you write. That’s all it needs to understand who you are.
This is just the beginning. AI tools like these will only get better with time, making it even easier to unmask anonymous posters around the internet. That means if you do have an anonymous account, you shouldn’t assume your identity is safe. Anyone can find the truth with your own words used against you to destroy your privacy.
