Recent research has shed light on a disquieting aspect of AI chatbots, unveiling their ability to glean personal information from users through seemingly innocuous conversational clues. The findings, presented by computer scientists from ETH Zurich in an interview with Wired, suggest that Large Language Models (LLMs) forming the core of popular chatbots like OpenAI’s ChatGPT and Google’s Bard can accurately infer private details about users based on minor contextual hints.
These LLMs, including the advanced GPT-4 that powers the paid version of ChatGPT, are trained on vast amounts of publicly-available internet data. While this aids in enhancing their conversational abilities, it also opens a concerning avenue for potential privacy breaches. The research team at ETH Zurich used text from Reddit posts to demonstrate the chatbots’ remarkable accuracy in identifying users’ locations, races, and other sensitive information.
In a chilling revelation, GPT-4 accurately predicted private information about users in the range of 85 to 95 percent of the time. For instance, it identified a user’s location in Melbourne, Australia, solely from a mention of a traffic maneuver unique to that city.
Furthermore, the researchers suggest that the algorithms could potentially infer a person’s race based on subtle comments. By analyzing language cues, such as mentions of nearby restaurants or neighborhoods, the chatbots can make highly likely assumptions about a user’s racial background.
The implications of this research extend to internet privacy concerns, urging users to exercise caution while interacting with chatbots. Despite growing awareness of online privacy, many individuals continue to share details that could inadvertently reveal personal information. The danger lies in the possibility of this information being misused, sold to advertisers, or exploited for malicious purposes.
In light of these findings, it becomes increasingly crucial for internet users to practice information security (infosec) by limiting the disclosure of identifying details in public forums. The research underscores the importance of understanding the potential risks associated with seemingly innocuous online interactions, urging users to exercise vigilance in their digital communications to safeguard their privacy and personal data.