Privacy News
OpenAI Pays Contractors to Read Your ChatGPT Conversations. Here's How to Turn It Off.
September 2026
On September 14, 2026, 404 Media published details of an internal OpenAI program called Project Lily. Under it, hundreds of contractors read and score real ChatGPT conversations — not flagged for abuse or safety review, but as routine quality work to make the model's responses sound better.
The setting that allows this is on by default for most people using ChatGPT. Here's what the reporting actually found, and what the opt-out does and doesn't cover.
What the reporting found
$50+/hr
reported pay for contractors reviewing full ChatGPT conversation threads
Default: ON
"Improve the model for everyone" is enabled automatically on Free, Plus, and Pro accounts
Default: OFF
Business, Enterprise, and Edu accounts are opted out by default
Not retroactive
switching the toggle off only protects conversations you have afterward
No username
reviewers don't see account names — but can see a "user memories summary"
€15M
prior fine from Italy's data regulator over processing user data without a proper legal basis
What reviewers actually see
According to the leaked materials, reviewers work from a dashboard showing a real user prompt alongside several ChatGPT responses generated for it. Their job is to summarize what the user was trying to do, then grade the responses against current style guidelines — flagging things like excessive emoji use, generic "AI-speak," sycophantic agreement, or a model falsely implying it has personal experience ("As a chef, I..."). Some assignments, 404 Media reported, include entire conversation threads rather than a single isolated message.
OpenAI says conversations pass through an automated Privacy Filter before reaching a reviewer, and that reviewers don't see the account username attached to a chat. But OpenAI's own documentation acknowledges the filter "can miss uncommon identifiers and ambiguous personal references" — and 404 Media reported that some assignments display a "user memories summary" above the prompt, describing what a person has previously used ChatGPT for, sometimes including where they may live or what they do for work. A conversation can carry identifying context even with no name attached to it.
The opt-out, and its limits
The control that governs all of this is a single toggle: "Improve the model for everyone," under Settings → Data Controls. It's on by default for Free, Plus, and Pro accounts — the ones almost every individual user is on — and off by default only for Business, Enterprise, and Edu accounts.
Two limits worth knowing before you assume you're covered:
- Turning it off protects future conversations only — it doesn't retroactively withdraw conversations that were already eligible for review before you changed the setting
- OpenAI's help documentation doesn't clearly disclose, at the moment someone hits send, that a human contractor rather than only an automated system may end up reading that exact conversation
OpenAI isn't the only company that does this kind of human review for chat quality — Anthropic and Google both use human reviewers under their own account settings, according to the same reporting. The model of "your conversation may be read by a person, unless you find the toggle" is closer to the industry default than an outlier.
Why this matters if you use AI for anything personal
People tell chatbots things they wouldn't type into a search box — health questions, relationship problems, drafts of emails with real names in them, voice memos transcribed and pasted in for cleanup. None of that is unusual. What's easy to miss is that "conversation with an AI" and "conversation with software that only your device sees" aren't the same category, even when the interface feels identical.
It's the same reason we build OffgridScribe and the rest of our apps to process on the phone rather than a server: there's no toggle to remember, because there's no upload step to opt out of in the first place.
Frequently asked questions
What is Project Lily?
Project Lily is an internal OpenAI program, reported by 404 Media on September 14, 2026, in which hundreds of contractors are paid $50 or more an hour to read and score real ChatGPT conversations. Reviewers see full conversation threads, summarize what the user was trying to do, and grade multiple model responses against current style and behavior guidelines.
Does OpenAI remove my personal information before contractors see it?
OpenAI says conversations pass through an automated Privacy Filter intended to strip personal information, and reviewers do not see account usernames. OpenAI's own documentation acknowledges the filter can miss uncommon identifiers and ambiguous personal references. Some reviewers have also reported seeing a "user memories summary" that can describe what a user has previously used ChatGPT for, including general location or profession.
Is this setting on by default?
Yes, for consumer accounts. The "Improve the model for everyone" toggle is enabled by default on Free, Plus, and Pro plans. Business, Enterprise, and Edu accounts are opted out by default. Turning the toggle off only affects future conversations; it does not retroactively remove chats that were already eligible for review.
How do I stop my ChatGPT conversations from being reviewed?
In ChatGPT, go to Settings, then Data Controls, and turn off "Improve the model for everyone." Signed-out web users can find the same option under the help menu settings. This change is not retroactive, so it only protects conversations you have after switching it off.
More from Offgrid Studio
Otter.ai vs Google vs Rev: What Actually Happens to Your Recordings? · OpenAI's Agents Attacked RubyGems Before They Attacked Hugging Face · AI's Biggest Names Called for a Slowdown. The Market Took Them Seriously.