AI can discover a lot of information from innocent social media posts (Photo: Getty/SOPA)

ChatGPT knows where you are and how much you earn.

That’s the finding of a recent study that found that artificial intelligence (AI) based on a large language model (LLM) can predict the age, location, gender, and income of Reddit users with an accuracy of up to 96%. % can accurately guess based solely on what they write in social media posts. .

Researchers from ETH University Zurich tasked nine LLMs with identifying the personal characteristics of 520 users, which the team could securely verify.

ChatCGPT-4 came out on top with an overall accuracy of 85%, while Meta’s LlaMA-2-7b scored the lowest at 51%.

And while some personal information was written explicitly in news reports or elsewhere online, such as entries on financial advice forums, much of the information was gleaned from more subtle cues, such as location-specific words.

Speaking to New Scientist, lead author Robin Staab said the findings were a warning about the amount of information we share online without realizing it.

“It shows us that we give away a lot of our personal information online without thinking about it,” he said. “Many people don’t assume that you can tell their age or whereabouts directly from the way they write, but LLMs certainly can.”

The LLMs found age to be the easiest characteristic to determine, with ChatGPT-4 showing an accuracy of 98%, compared to an accuracy of 63% for income.

The authors note that while a human might make similar assumptions if given the same information, LLMs are 240 times faster and 100 times cheaper, posing significant privacy risks.

“Our results show that current LLMs can infer personal data to a degree that was previously unattainable,” they said. “In the absence of a working defense, we advocate a broader discussion about the privacy implications of LLM that goes beyond memorization and toward broader privacy protections.”

Previous concerns surrounding LLMs focused on the use of public data for training or potential leaks of private call data using the software, but the latest study highlights new limitations.

“We are only beginning to understand how privacy can be compromised by the use of LLMs,” cybersecurity expert Professor Alan Woodward told New Scientist.

The research is timely as Prime Minister Rishi Sunak is currently hosting a global AI summit at Bletchley Park, bringing together world leaders and technology bosses to discuss security issues.

Earlier today, Technology Minister Michelle Donelan said delegations from around the world attending the summit had agreed on the “Bletchley Declaration on AI Security” as a starting point for a global discussion on the issue.

Opening the summit, Ms Donelan said the agreement was a “landmark achievement” and “lays the foundation for today’s discussions”.

“It reinforces the need to address these risks as the only way to safely seize the extraordinary opportunities,” she said.