The way you talk can reveal a lot about you—especially if you’re talking to a chatbot. New research reveals that chatbots like ChatGPT can infer a lot of sensitive information about the people they chat with, even if the conversation is utterly mundane.
The phenomenon appears to stem from the way the models’ algorithms are trained with broad swathes of web content, a key part of what makes them work, likely making it hard to prevent. “It’s not even clear how you fix this problem,” says Martin Vechev, a computer science professor at ETH Zürich in Switzerland who led the research. “This is very, very problematic.”
Vechev and his team found that the large language models that power advanced chatbots can accurately infer an alarming amount of personal information about users—including their race, location, occupation, and more—from conversations that appear innocuous.
Vechev says that scammers could use chatbots’ ability to guess sensitive information about a person to harvest sensitive data from unsuspecting users. He adds that the same underlying capability could portend a new era of advertising, in which companies use information gathered from chatbots to build detailed profiles of users.
Some of the companies behind powerful chatbots also rely heavily on advertising for their profits. “They could already be doing it,” Vechev says.
The Zürich researchers tested language models developed by OpenAI, Google, Meta, and Anthropic. They say they alerted all of the companies to the problem. OpenAI spokesperson Niko Felix says the company makes efforts to remove personal information from training data used to create its models and fine tunes them to reject requests for personal data. “We want our models to learn about the world, not private individuals,” he says. Individuals can request that OpenAI delete personal information surfaced by its systems. Anthropic referred to its privacy policy, which states that it does not harvest or “sell” personal information. Google and Meta did not respond to a request for comment.
“This certainly raises questions about how much information about ourselves we’re inadvertently leaking in situations where we might expect anonymity,” says Florian Tramèr, an assistant professor also at ETH Zürich who was not involved with the work but saw details presented at a conference last week.
Tramèr says it is unclear to him how much personal information could be inferred this way, but he speculates that language models may be a powerful aid for unearthing private information. “There are likely some clues that LLMs are particularly good at finding, and others where human intuition and priors are much better,” he says.