OpenAI is facing scrutiny as it has been revealed that hundreds of contractors are reviewing user prompts and conversations with ChatGPT under a project codenamed Project Lily. This initiative aims to enhance the AI's responses, but raises significant privacy issues as confidential and personal data may inadvertently be exposed during the process. Investigative reports have highlighted that while user data is anonymized, the chats can still contain sensitive information about users, such as their previous interactions with the AI and personal details.
The findings stem from leaked internal documents, which included guidelines for reviewers and insights from communications platforms used within OpenAI. According to the reports, contractors do not have access to user account names, but the conversations often reveal personal information, including summaries of users' past questions and contexts that may disclose their identities.
In response to the concerns, OpenAI stated that a model called the Privacy Filter is employed to remove any personal data before chats are passed on to reviewers. However, the company acknowledged that the filter is not foolproof, occasionally missing rare or unusually expressed personal information, and sometimes obscuring context that could clarify nuances.
Contractors engaged in this work reportedly earn over $50 an hour and are tasked with evaluating user prompts and the corresponding responses from ChatGPT. They must identify successful and unsuccessful elements of the responses and provide ratings on a scale from one to seven. Although they are required to flag factual inaccuracies, they are not mandated to verify the information through external sources.
The aim of this process is to make AI-generated responses feel less mechanical, avoiding clichés and ensuring the model adapts to the user's tone more naturally. Notably, OpenAI maintains that user chats are not used to improve models unless users opt into this feature, which is enabled by default for free and paid subscriptions but disabled for enterprise accounts.
Experts have pointed out that while human review of chats may enhance model safety, it creates a misleading sense of privacy for users who may believe their conversations are confidential. This situation is not unique to OpenAI; other companies like Anthropic and Google also use human reviewers to improve their AI systems, albeit with similar privacy considerations.
This development could impact how the market approaches user data privacy in AI interactions, potentially leading competitors to adopt more transparent practices regarding data handling and user consent. As awareness grows, companies may need to reassess their privacy policies to maintain user trust.
Informational material. 18+.
