AI Chat Privacy Risks: Ruben Hassid Warns Business Users About Data Exposure

In a recent LinkedIn post, Ruben Hassid explores critical privacy concerns surrounding the use of AI chatbots like ChatGPT and Claude within a business context. Hassid, a researcher into prompt lifecycles and AI data privacy, cautions business professionals against assuming their conversations on these platforms are truly private, especially when using team or business plans.

The core of Hassid’s concern lies in how data is handled by AI providers, particularly when utilizing corporate accounts. He highlights that even features like incognito mode may not guarantee privacy on a team plan.

“Both ChatGPT and Claude have incognito mode. But on a Team plan, the workspace owns the data.”

Hassid emphasizes that this data ownership by the workspace means that conversations, even those intended to be private, can be accessed by account administrators. He directly quotes Anthropic’s policy to underscore this point:

“Incognito chats are included in organizational data exports available to account Owners.”

The Implications of AI Data Sharing

Ruben Hassid illustrates the potential for data exposure with a compelling anecdote involving mathematician Tristan Buckmaster and OpenAI. Buckmaster reportedly developed a solution to a Millennium Prize problem using Codex, an AI model. OpenAI later revealed they also arrived at the solution shortly after, raising questions about whether they had trained on Buckmaster’s private sessions.

While OpenAI stated that no human or agent directly accessed his sessions, Hassid points out the broader implications. According to Hassid, OpenAI confirmed that they, like other LLM companies, do train on de-identified user data.

“Yes. And so does every LLM company.”

This practice, Hassid argues, means that virtually all user input, regardless of its sensitive nature, could potentially be used for training future AI models.

Navigating AI Privacy for Personal and Professional Use

Hassid acknowledges that many individuals, himself included, use these AI tools for personal matters, such as discussing health, finances, or personal relationships. However, he strongly advises caution.

Risks for Personal Conversations

The researcher warns that personal conversations shared on business accounts could be exposed to employers through data exports. He cites examples of data accessibility:

• 100,000 shared chats indexed by Google.
• 20 million ChatGPT chats accessed by lawyers for The New York Times.

To mitigate these risks, Hassid recommends a clear separation of AI usage:

• Personal Use: Utilize a personal account, disable data training, use incognito mode, and avoid giving a ‘thumbs up’ to responses.
• Client Work: Use a dedicated business account or refrain from using AI tools for such sensitive information if no secure business solution is available.

In conclusion, Ruben Hassid’s post serves as a stark reminder for professionals to be acutely aware of the data privacy policies governing the AI tools they use. As he puts it, “Your boss shouldn’t learn about your divorce from just an export.” His insights underscore the need for careful consideration of which AI platform and account type to use for different types of conversations to protect both personal and professional information.