OpenAI is hiring humans to read your ChatGPT chats
OpenAI wants its AI models to become less flattering and agreeing.
Image by Leon Neal via Getty Images
- OpenAI is hiring contractors to review real ChatGPT prompts and rate chatbot answers.
- Reviewers may see memory summaries and approximate locations, though usernames and personal information are hidden.
- The project aims to make ChatGPT less human-like and less overly agreeable with users.
- OpenAI faces lawsuits over alleged chatbot harms, but the company denies the allegations.
Key Takeaways by nexos.ai, reviewed by Cybernews staff.
Contractors read and rate real user interactions with the OpenAI chatbot, ChatGPT, which may contain personal information.
404 Media reports that OpenAI is hiring hundreds of contractors to read users’ ChatGPT prompts and rate answers generated by the chatbot to improve the company’s AI models as part of its “Project Lily.”
However, over 900 million ChatGPT users may not be aware that humans may be reading their interactions.
OpenAI does take privacy precautions – usernames aren’t displayed to human reviewers, and personal information is hidden.
However, contractors can see a user’s memory summary above the prompt, which may include information on what the person used the chatbot for before, and their approximate location.
Internal documents seen by 404 Media reveal that the training aims to prevent the chatbot from anthropomorphizing itself. In other words, the company wants ChatGPT to act less like a human.
Stay updated with our latest stories and follow us on social media
Be the first to discover new stories, ideas, and updates from our team.
In addition, the training is dedicated to making the chatbot less sycophantic, which involves overly agreeing with users and validating them, for both good and bad causes.
OpenAI is facing multiple lawsuits accusing the chatbot of pushing people to suicide. The company denies allegations.
According to one of the lawsuits, ChatGPT discussed suicide methods in detail with a 16-year-old Adam Raine, despite him exhibiting signs of self-harm intent, and offered to write the first draft of his suicide note. The teenager later took his own life.
OpenAI retired its most sycophantic model, GPT-4o, in February 2026 and implemented stronger guardrails for later models.
The company didn’t respond to 404 Media's question about whether it tells users that humans may read their chats, and referred to its help page after the report was published.
Cybernews reached out to OpenAI for commentary and will update the article once it receives the response.
This isn’t the first time OpenAI has been under scrutiny for its privacy practices.
In 2025, a number of users reported finding their private ChatGPT conversations on Google search after sharing the chats using the “Share” button for collaboration.
The company quickly removed the feature, which OpenAI’s CISO, Dane Stuckey, called “a short-lived experiment to help people discover useful conversations.”