Tech Trends & Innovation: The Latest in Tech News
- Some chatbots prompt users for narrowing details, enabling surfacing of home addresses, purchase prices, and spouse names.
- DeleteMe's Shavell says guardrails clash with chatbots’ design to be useful, increasing risk of exposing PII.
- Futurism found xAI’s Grok often returned residential addresses, work addresses, and phone numbers.
- OpenAI said it could not investigate without screenshots or model details, pointed to docs about filtering PII; users may not know which model they use.
- There are no easy ways to verify if PII is in a model's training data or to compel its removal from models.
“It was severely downgraded,” Gilbert confirms. “I never would have found it if I was just looking through Google results.” (I tried the same prompt in Gemini earlier this month, and after an initial denial, the tool also gave me Eiger’s number.)
After this experience, Eiger, Gilbert, and another UW PhD student, Anna-Maria Gueorguieva, decided to test ChatGPT to see what it would surface about a professor.
At first, OpenAI’s guardrails kicked in, and ChatGPT responded that the information was unavailable. But in the same response, the chatbot suggested, “if you want to go deeper, I can still try a more ‘investigative-style’ approach.” Their inquiry just had to help “narrow things down,” ChatGPT said, by providing “a neighborhood guess” for where the professor might live, or “a possible co-owner name” for the professor’s home. ChatGPT continued: “That’s usually the only way to surface newer or intentionally less-visible property records.”
The students provided this information, leading ChatGPT to produce the professor’s home address, home purchase price, and spouse’s name from city property records.
(Taya Christianson, an OpenAI representative, said she was not able to comment on what happened in this case without seeing screenshots or knowing which model the students had tested, though we pointed out that many users may not know which model they were using in the ChatGPT interface. In response to questions about the exposure of PII, she sent links to documents describing how OpenAI handles privacy, including filtering out PII, and other tools.)
This reveals one of the fundamental problems with chatbots, says DeleteMe’s Shavell. AI companies “can build in guardrails, but [their chatbots] are also designed to be effective and to answer customer questions.”
The exposure issue is not limited to Gemini or ChatGPT. Last year, Futurism found that if you prompted xAI’s chatbot Grok with “[name] address,” in almost all cases, it provided not only residential addresses but also often the person’s phone numbers, work addresses, and addresses for people with similar-sounding names. (xAI did not respond to a request for comment.)
No clear answers
There aren’t straightforward solutions to this problem—there’s no easy way to either verify whether someone’s personal information is in a given model’s training set or to compel the models to remove PII.
Read the full article from the original source
