OpenAI's Secret 'Lily' Project: Are Humans Reading Your ChatGPT Chats?

When you pour your heart out to ChatGPT, you might assume it's just you and the algorithm. But a recent leak suggests otherwise. According to internal documents obtained by 404Media, OpenAI has quietly run Project Lily—a program where human reviewers sift through real user conversations to fine-tune the model.

What Exactly Is Project Lily?

At its core, Project Lily employs prompt reviewers who examine anonymized chat samples. Their job? To judge whether ChatGPT's responses are relevant and to flag anything that sounds too 'AI-ish'—think condescending tones, excessive emojis, or fawning language. The guidelines even ban anthropomorphic expressions: you won't catch the bot saying 'As a chef, I like…' or 'I understand that feeling.' That kind of human-like empathy is a no-go.

The work is repetitive, but the pay is decent—over $50 an hour for easier tasks, which explains the steady stream of applicants.

The Privacy Elephant in the Room

Here's where things get uncomfortable. Many users treat ChatGPT as a digital confidant, sharing secrets they wouldn't tell a soul. OpenAI claims the data is anonymized and usernames are stripped, but the company admits the filtering isn't perfect—especially in short conversations where context can easily identify someone. Worse, reviewers often receive a 'user memory summary' that includes the user's questions, interests, and even location details.

And deleting your chat history? That doesn't necessarily erase it. If you've left the default 'Allow use of your conversation records to improve the product' setting on, your past chats may already be collected and stored. Even after you hit delete, the data could live on in anonymized datasets. Enterprise and education users have this off by default, but for regular folks, the toggle is on—and flipping it off won't retroactively remove anything.

OpenAI's Quiet Response

When pressed by reporters, OpenAI initially dodged the question of whether users are clearly informed about human reviews. They pointed to a FAQ page that vaguely mentions 'human reviews for model optimization.' After the 404Media report went live, OpenAI quietly updated its help page to add opt-out methods—but still didn't explicitly say that staff read conversation content.

The Industry Norm

OpenAI isn't alone here. Google's Gemini privacy center admits some chats may be reviewed by humans. Anthropic has a dedicated page explaining the practice. Perplexity, however, stays mum—its policy neither confirms nor denies human access. The takeaway? Human involvement is still the backbone of AI improvement, even in an era of extreme automation.

So next time you chat with ChatGPT, remember: it might not be just you and the machine. And that's a conversation worth having.

Key Points

Project Lily uses human reviewers to read anonymized ChatGPT chats and flag 'AI-like' tones.
Reviewers earn $50+/hour, but privacy risks loom—filtering can miss personal data.
Users' chat history may be retained even after deletion; opt-out doesn't erase past data.
OpenAI updated its help page post-leak but avoided explicitly stating humans read chats.
Competitors like Google and Anthropic also use human review, highlighting an industry-wide practice.

OpenAI's Secret 'Lily' Project: Are Humans Reading Your ChatGPT Chats?