Are you an anonymous wombat commenting online? AI knows who you are
Do you like to comment online under an anonymous nickname? If you’re not careful, AI can uncover who you actually are.
“On the internet, nobody knows you're a dog” was a famous caption from a cartoon by Peter Steiner, which, at the time, emphasized the power of anonymity online.
While we might still think it's possible to stay nameless online, recent research shows that’s about to change, thanks to AI.
Academics have developed large language models (LLMs) capable of deanonymizing users online, based on the comments they wrote online or “other digital clues they have left behind,” states Simon Lermen, researcher and developer specializing in AI, on his Substack.
The specialist explained that although it has long been known that only a few attributes are needed to identify individuals, it takes time for human beings to investigate and link all this data to obtain meaningful information.
“Data is often only available in unstructured form and deanonymization used to require human investigators to search and reason based on clues,” shared Lermen, who, together with the team, showed that LLMs can find out where the user lives, what they do, and even their interests from “a handful of comments.”
Their created method can even detect users who use multiple accounts on different platforms, as well as attribute real identities to unknown accounts and online activity, and vice versa.
These LLMs work by analyzing users' past activity and creating user profiles. Once they have enough data points, the LLMs begin to develop links between similar profiles based on their locations, ages, hobbies, vocabularies, and so on.
The study was conducted by the team, creating two types of deanonymization proxies.
In one example, researchers anonymized Hacker News (HN) accounts by removing all information that could identify them. They then let LLMs “match the anonymized account to the true person,” stated Lermen.
Check if your data has been leaked
Another approach involved “splitting” an account into two to see if AI could link them back together.
The researchers also conducted a real-world deanonymization attack on the Anthropic Interviewer, a tool that performs detailed user research interviews.
The Claude-based tool includes interviews with scientists about the use of AI, which are anonymous. However, AI could be used to identify some of these scientists.
Even though the researchers recognized that this study risks exacerbating AI misuse, it may also shed light on how users can protect themselves.
Lermen concludes that from the platforms’ point of view, they should take actions to restrict data access, while users, who use anonymous accounts, should be more aware of what information they share online: the city they live in, their occupation, hobbies, or any other specific information that could help narrow down who they are.
Unlock more exclusive Cybernews content on YouTube.
