Can OpenAI contractors read your ChatGPT conversations — and how do you stop it?

TL;DR. OpenAI pays hundreds of outside contractors to read real ChatGPT conversations under an internal program called Project Lily. Contractors see full conversation transcripts plus a memory summary that can reveal your location and usage patterns — without seeing your username. The setting to stop this is on by default for Free, Plus, and Pro accounts. To opt out: Settings → Data Controls → turn off "Improve the model for everyone." Enterprise and Business accounts have human review off by default.

---

An investigative report by 404 Media published in September 2026 revealed that OpenAI pays hundreds of contractors to read and rate real ChatGPT conversations as part of a program internally called Project Lily. The goal is to make ChatGPT's responses sound less sycophantic and less robotically flat. The method is human judgment at scale.

Here's what that actually means for professionals who use ChatGPT for work.

What is Project Lily?

Project Lily is OpenAI's program for using human reviewers — outside contractors paid through third-party staffing platforms including Crossing Hurdles and Mercor — to improve ChatGPT's response quality. Contractors are paid over $50 an hour to read a user's prompt, summarize the user's intent, and then rate four potential AI responses on a 1-to-7 scale.

This kind of human feedback loop is standard in AI development: it's how models learn to sound more natural and less robotic over time. What the 404 Media investigation made visible is the scale (hundreds of contractors), the scope (full conversation transcripts), and the gap between what OpenAI discloses and what most users understand is happening when they type into a chat window.

What can contractors actually see?

According to 404 Media's reporting, contractors can see:

• The full conversation transcript — your prompt and the AI's response, in context
• A user memory summary displayed above the prompt — this can include what topics you've discussed before and, sometimes, where you appear to be located
• Your intent inferred from the conversation — contractors explicitly summarize what you were trying to do

What they cannot see: your username. But the memory summary and conversation content can together contain enough context to make conversations identifiable in practice — which is exactly the gap that OpenAI's privacy filter is supposed to close.

How does OpenAI's privacy filter work — and where does it fall short?

OpenAI uses a "Privacy Filter" model to strip personal information from conversations before they reach reviewers. The company's own documentation for the filter acknowledges it "can make mistakes, miss uncommon identifiers and under-redact when context is limited."

Explicit personal details — names, email addresses, phone numbers — are more likely to be caught than implicit ones: your location inferred from references to a neighborhood or local landmark, sensitive topics identified from conversation history, or personal circumstances revealed through what you asked and how you framed it. The filter is a best-effort pass, not a guarantee.

Which ChatGPT plans are affected?

Human review is on by default for:

• ChatGPT Free
• ChatGPT Plus
• ChatGPT Pro

Human review is off by default for:

• ChatGPT Enterprise
• ChatGPT Business
• ChatGPT Edu

If you're on a personal subscription and haven't adjusted this setting, your conversations are eligible for human review.

How to turn it off — step by step

1. Open ChatGPT (web or mobile)
2. Click your profile icon → Settings
3. Go to Data Controls
4. Toggle off "Improve the model for everyone"

That's it. Two important caveats that apply regardless of the setting:

This only applies to new conversations. Past conversations already submitted for review cannot be recalled. Turning off the setting stops future submissions — it doesn't undo past ones.

Temporary Chat is a separate protection. For sensitive one-off conversations, ChatGPT's Temporary Chat mode is excluded from training and human review entirely. You won't have memory or conversation history in Temporary Chat, but the conversation won't be submitted to contractors. OpenAI retains it for 30 days for safety purposes before deleting it.

What this means for professionals

The existence of human review in AI development isn't new or unique to OpenAI. Anthropic uses human reviewers; Google's Gemini documentation openly states that a sample of conversations is reviewed by humans. The principle — that humans sometimes read AI conversations to improve model quality — has been disclosed somewhere in the terms of every major AI provider.

What the Project Lily reporting adds is specificity: a named program, a headcount, and a concrete account of what reviewers actually see. "AI companies may review your data" is different from "hundreds of outside contractors can see your full conversation plus a memory summary of your past use patterns."

For most professional use, this doesn't require changing how you work — it requires checking a setting you may not have looked at before. For anyone sharing sensitive client information, proprietary business data, or anything they'd be uncomfortable having a contractor read: the consumer setting is the floor, not the ceiling. The right tier for regulated professional work is Enterprise or Business, which has different defaults and available zero-data-retention agreements.

If your work involves protected health information, client communications, or other regulated data, see also: Does Claude train on your data? (Plus ChatGPT and Gemini) — how to opt out for the full picture on training vs. human review across all three major tools, and which AI vendors will sign a Business Associate Agreement for compliance-grade options.

---

Sources

• 404 Media — "Inside 'Project Lily': The Humans Reading Your ChatGPT Chats," September 2026: https://www.404media.co/inside-project-lily-the-humans-reading-your-chatgpt-chats/
• The Next Web — "Hundreds of contractors are reportedly reading real ChatGPT conversations": https://thenextweb.com/news/chatgpt-human-reviewers-gdpr
• The Decoder — "OpenAI has hundreds of contract workers reading your ChatGPT conversations," September 14, 2026: https://the-decoder.com/openai-has-hundreds-of-contract-workers-reading-your-chatgpt-conversations/
• OpenAI — Enterprise Privacy: https://openai.com/enterprise-privacy/