Meta AI Contractors See Personal Data in User Chats
Contract workers hired by Meta through third-party training platforms reviewed real user conversations with Meta's AI chatbot and encountered personally identifiable information such as names, phone numbers, email addresses, and selfies, according to interviews with four contractors reported by Business Insider. The exposure occurred during AI training work intended to improve the chatbot's responses, with insufficient data minimization to prevent contractor access to identifying user information.
What happened
Contract workers hired by Meta through training platforms Outlier (owned by Scale AI) and Alignerr reviewed real user conversations with Meta's AI chatbot as part of AI training and improvement work. Four contractors told Business Insider that while rating and reviewing the AI's responses, they routinely encountered chats containing full names, phone numbers, email addresses, gender information, hobbies, selfies, and other personally identifiable information that could identify users.
One contractor estimated that personally identifiable information appeared in more than half of the thousands of chats they typically reviewed each week. Some personal data was provided by Meta alongside chat histories to help contractors personalize the AI's response, while in other cases users voluntarily disclosed personal information to the chatbot despite Meta's privacy policy warning against doing so.
Reported impact
Users' private conversations with Meta's AI—which contractors described as therapy-like sessions, intimate exchanges, and personal discussions about fantasies and problems—were accessed by human reviewers who could identify individual users. In one example cited in the reporting, a sexually explicit conversation contained enough personal information for identification within minutes.
Index analysis
The exposure of personally identifiable information to contractors appears to result from Meta's practice of including unredacted user data in training materials to help AI personalization. Two contractors working on similar projects for other tech companies reported that the inclusion of unredacted personal data was more common for Meta projects than for comparable work elsewhere.
Governance controls
Project documents reviewed by Business Insider showed that contractors were told the interactions they assessed were real conversations between users and the chatbot. Meta acknowledged that it does review user interactions with its AI, though the reporting suggests insufficient data minimization or anonymization controls in the contractor review process.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.