Inside “Project Lily”: The Allegations of Humans Reading Your ChatGPT Chats
A new leak challenging the privacy standards behind OpenAI’s AI chatbot, ChatGPT, reveals that user conversations are being reviewed by third-party contractors. According to an investigation by 404 Media, under a covert initiative dubbed “Project Lily,” real, unfiltered user dialogues with ChatGPT are reportedly being read and analyzed line by line. Aimed at refining the models’ safety and accuracy guardrails, this review process shatters the misconception that the intimate thoughts users entrust to the AI are entirely shielded from human eyes.
The reviewed materials reportedly encompass user-generated inputs such as medical and health data, legal inquiries, deeply personal disclosures, and corporate correspondence. Although OpenAI states that reviewed conversations and user submissions are anonymized and not directly tied to user account credentials, the risk of indirect re-identification persists due to proper names, locations, and specific contextual details embedded within the text. Furthermore, in this context, the non-disclosure agreements and access protocols governing the contractor workforce tasked with data labeling and review are facing scrutiny.
Allegations of leaks have brought back to the forefront the “human-in-the-loop” methods used in developing artificial intelligence models, as well as the relationship between companies’ model safety targets and their data processing policies. Experts highlight the boundaries of user consent mechanisms and privacy disclosures regarding the circumstances under which user data is retained and viewed by third parties.