Humans read ChatGPT chats: What Project Lily reveals

What this is about

A 404 Media investigation published on September 14, 2026 describes an OpenAI effort internally called Project Lily. According to the report, hundreds of outside contractors read real ChatGPT conversations, score responses, and provide feedback intended to make the system less flattering and less human-like. The reviewers reportedly do not see usernames. Personal or sensitive details contained in the conversation itself can nevertheless remain visible.

This matters to ordinary users because many people use chatbots for health questions, relationship problems, work documents, and other confidential material. Human quality review is not unusual in AI development. The key question is whether users clearly understand when their content may be used to improve models and potentially seen by people.

What Project Lily actually does

According to documents reviewed by 404 Media, contractors rate real responses on a scale from one to seven and write critiques. The Decoder adds that workers are recruited through Crossing Hurdles and paid through Mercor. One North America-based reviewer reportedly earned more than $50 per hour.

OpenAI says in its Data Controls FAQ that signed-in users can open Settings → Data Controls and disable “Improve the model for everyone.” New conversations then will not be used for training. OpenAI also says Temporary Chats are not used for training, may be reviewed only to monitor abuse, and are deleted after 30 days.

Why it matters

A conversation can be identifiable even without an account name. Someone who mentions an employer, location, medical condition, case number, or relatives may provide context that cannot be reliably anonymized by removing an email address. OpenAI says, according to 404 Media, that it filters personal information before review, but acknowledges that sensitive details can still pass through.

The issue is therefore not merely that human review exists. It concerns understandable consent, data minimization, and whether a friendly training toggle adequately explains that outside service providers might see individual content. Anthropic also confirmed to the investigation that it uses human review for users who opted into model improvement. Google likewise warns Gemini users that humans may review saved chats.

In plain language

Imagine dropping a handwritten letter without a return address into a suggestion box. Your name is missing from the envelope, but the letter might still mention your town, diagnosis, and manager. “Anonymized” does not automatically mean nobody can work out who it concerns.

A practical example

A user pastes an 800-word medical letter into ChatGPT and asks for a simpler explanation. She does not type her account name, but the letter contains her date of birth, a rare diagnosis, and the hospital. If model improvement is enabled, a filtered version could later enter a review process. Even if two direct identifiers are removed, three other details may remain sensitive or identifying. A safer practice is to remove personal details before pasting and disable training for new chats.

Scope and limits

• 404 Media relies on internal documents and sources; Cyber Ivy could not independently verify the scale or selection process of Project Lily.
• OpenAI says disabling model improvement applies to new conversations. It does not by itself answer how previously used data was handled.
• Temporary Chats reduce training use but are not a vault: OpenAI states a retention period of up to 30 days and possible review for abuse monitoring.

SEO & GEO keywords

OpenAI, ChatGPT, Project Lily, human data review, privacy, training data, Data Controls, Temporary Chat, contractors, AI privacy