OpenAI contractors read real ChatGPT chats, and three rivals allow the same
Hundreds of contractors read ChatGPT conversations under an OpenAI program called Project Lily, 404 Media reported on September 14, 2026, and the prompts they see "can include whole conversations between users and the chatbot." Usernames are stripped first, but OpenAI told the outlet that sensitive details can still get through. This is the fourth of four reasons in OpenAI's own policy. The same clause at each major vendor, checked against their pages on September 26, 2026:
| Vendor | Who may read a chat | Stated reasons | How long a reviewed or flagged chat is kept | Can you opt out of review? |
|---|---|---|---|---|
| OpenAI ChatGPT | "A limited number of authorized OpenAI personnel" plus service providers | Abuse or security incidents, support, legal matters, improving models | No number; cleared chats go within 30 days unless kept "for security or legal reasons" | Only for model improvement, under Data controls |
| Anthropic Claude | "designated members of our Trust & Safety team" | Enforcing the Usage Policy, or feedback you send | Inputs and outputs up to 2 years, classification scores up to 7 years | No; the training setting does not cover a flagged chat |
| Google Gemini | "Human reviewers (including trained reviewers from our service providers)" | Improving Google services, and protecting "Google, our users, and the public" | Up to 3 years, even after you delete your activity | Keep Activity off stops training use; safety review continues |
| Microsoft Copilot | "Trained AI experts" | Product improvement, digital safety, suspected Code of Conduct violations | No separate number; activity kept 18 months by default | No: "an opt-out of human review is not available" for a suspected violation |
| Perplexity | No reader named in the Privacy Notice of July 8, 2026 | Data may be shared "to investigate, prevent, or act on" policy violations | No fixed period; kept "only as long as necessary" | AI Data Retention toggle, future data only |
Two of the five say in writing that no setting stops safety review. Microsoft's Copilot FAQ: "To ensure that our services are safe and secure for everyone, an opt-out of human review is not available." Google's Gemini hub: "Even if your Keep Activity setting is off or you use temporary chats, Google still uses your chats to respond to you and help protect Google, our users, and the public, including with help from human reviewers." Google is also the only vendor that says a reviewed chat survives your delete button.
When can OpenAI employees see your ChatGPT chats?
OpenAI lets a person read a ChatGPT chat in four cases, and only the fourth has a switch. The help center, updated September 26, 2026, puts it in one sentence: "A limited number of authorized OpenAI personnel, as well as trusted service providers that are subject to confidentiality and security obligations, may access user content only as needed for these reasons: (1) investigating abuse or a security incident; (2) to provide support to you if you reach out to us with questions about your account, or to troubleshoot issues affecting the product experience; (3) to handle legal matters; or (4) to improve model performance (unless you have opted out)." It adds that OpenAI logs every access and that staff "must undergo security and privacy training prior to accessing any user content," and it ends on a plain warning: "Please do not enter sensitive information that you would not want reviewed or used."
A second route runs through outside firms. The same page says OpenAI "may also send select portions of content to trusted service providers (subject to confidentiality and security obligations) for data annotation and safety purposes." Annotation is a person labelling your text.
The abuse route has a published pipeline. In a post dated August 26, 2025, OpenAI wrote that conversations from users "planning to harm others" go to "a small team trained on our usage policies and who are authorized to take action, including banning accounts," and that an imminent threat of serious physical harm to others may be referred to law enforcement. Self harm cases are not referred, the post says, "to respect people's privacy."
No retention number covers a reviewed chat. A cleared chat is "deleted from our systems within 30 days, unless they have previously been de-identified and disassociated from your account or we have to keep them for security or legal reasons," and a chat in the abuse pipeline is a security reason. A temporary chat is kept "for up to 30 days for safety purposes."
What is Project Lily, and does turning off training keep your chats out?
Project Lily is OpenAI's program of paid contractors who rate ChatGPT's replies to real user prompts, and turning off "Improve the model for everyone" takes your future chats out of it. 404 Media's September 14, 2026 report says OpenAI is hiring hundreds of them; Yahoo Tech's September 15 summary of the report gives the rate as $50 per hour. Futurism, on September 17, reported from the same documents that reviewers focus on making ChatGPT less sycophantic and less inclined to talk about itself as a person.
The anonymization has known gaps. Futurism reported that the reviewers' dashboard includes a "user memories summary" that can reveal details from earlier conversations, and that OpenAI said it runs a "Privacy Filter" model over chats, a filter its own website says "can make mistakes." A name typed into the chat is chat content, and chat content is what the reviewer reads.
The opt out covers reason four and nothing more. The setting is on by default, as Futurism noted, and the training toggle guide has the steps for ChatGPT and every other major app. It does not reach reasons one to three, and it does not pull back a chat already sent to review.
Do people read Gemini, Claude, Copilot and Perplexity chats too?
Yes at Google, Anthropic and Microsoft, which each publish a human review clause; Perplexity publishes none, which is silence and not a promise.
Google. The Gemini Apps Privacy Hub, updated September 24, 2026, is the bluntest: "Please don't enter confidential information that you wouldn't want a reviewer to see or Google to use to improve our services." Reviewed chats "are not deleted when you delete your activity. Instead, they are retained for up to three years." Chats are "disconnected from your account before being sent to service providers," and ordinary activity auto deletes after 18 months by default (3 or 36 months if you change it). Temporary chats and chats with Keep Activity off are kept 72 hours.
Anthropic. Its Privacy Center, in an article dated March 16, 2026, says "By default, Anthropic employees cannot access your conversations" unless you send feedback or review "is needed to enforce our Usage Policy," and then "only designated members of our Trust & Safety team may access this data on a need-to-know basis." The retention article, dated July 1, 2026, keeps a flagged chat's "inputs and outputs for up to 2 years and trust and safety classification scores for up to 7 years." If you allow training, de-identified data stays in training pipelines up to 5 years; if you do not, a deleted chat leaves back-end storage within 30 days.
Microsoft. The Copilot Privacy FAQ says "Trained AI experts may review Copilot conversations to build, evaluate, and improve the accuracy and safety of our AI models," and that "Some conversations are reviewed when a violation of the Code of Conduct is suspected." Conversation activity is stored 18 months by default. The FAQ says it covers the Copilot app from before the August 18, 2026 update; the pages for the new app do not restate the review clause.
Perplexity. The Privacy Notice (last updated July 8, 2026) names no reviewer. It says "We share your data with service providers" who "help us to run our services, conduct surveys, provide technical support," and keeps personal data "only as long as necessary." Its help center (September 3, 2026) turns AI Data Retention on by default for Free, Pro and Max accounts, and warns that "Previously collected training data cannot be deleted or removed." Enterprise accounts get Zero Data Retention agreements with OpenAI and Anthropic; consumer accounts are not described as covered.
Does using ChatGPT through another app stop OpenAI's reviewers?
No: any app that sends your message to GPT sends it to OpenAI, and OpenAI's abuse and legal clauses apply at its end whatever the app promises. Whizi is one such workspace, and its data and privacy doc says so in plain words: "a message you send to Claude reaches Anthropic, and a message you send to GPT reaches OpenAI." What Whizi controls is its own side (it does not train Whizi owned models on your content, and uploads expire after up to 30 days) and which provider receives each thread, so a thread you want kept away from Gemini's three year window goes elsewhere. Where Whizi is weaker: its privacy policy names no review team and no retention period for a report someone at Whizi has looked at, which is less than OpenAI's four reason sentence. Switching models mid thread also sends the whole thread to the new provider.
Which settings actually remove a human reader?
Turning training off is the only switch any vendor offers, and nothing removes the safety reviewer. That toggle covers OpenAI's reason four, Google's training use, Microsoft's training use and Perplexity's AI Data Retention, and it decides whether Anthropic keeps your data 30 days or 5 years. None of it is retroactive.
Temporary modes shorten the window, not the right to look. OpenAI keeps a temporary chat copy up to 30 days for safety, Google keeps a Keep Activity off chat 72 hours, and a flagged chat is the exception at every vendor.
What you type is the last control. The never paste list is the short version, does deleting a chat really delete it covers the delete button, and on a work account the reader list is longer, which can my employer see my AI chats sets out. The most private AI chatbot comparison ranks the same vendors on training defaults.
- Turn off "Improve the model for everyone" in ChatGPT under Settings, then Data controls; it removes reason four and Project Lily for future chats
- Turn Keep Activity off in Gemini before a chat you would not want a reviewer to see; reviewed chats are kept up to 3 years
- Check the Claude training setting; opted in data is kept up to 5 years and flagged chats 2 years regardless
- Toggle AI Data Retention off in Perplexity Account Settings; it covers future data only
- Use temporary modes for sensitive threads: 30 days at OpenAI, 72 hours at Google
- Assume the safety review path has no opt out at any vendor; Microsoft and Google say so in writing
Frequently asked questions
Can OpenAI employees see my chats?
Yes, a limited number of authorized OpenAI staff and service providers can, for four stated reasons: investigating abuse or a security incident, support you ask for, legal matters, and improving models unless you opt out. OpenAI says every access is logged and staff must complete security and privacy training first. Most chats are never opened by a person.
Can OpenAI see your deleted chats?
For up to 30 days, and longer in two cases. OpenAI deletes a cleared chat from its systems within 30 days, unless it was already de-identified and disassociated from your account or OpenAI has to keep it for security or legal reasons. So a de-identified copy made before you pressed delete can outlive the delete.
Are chats anonymized before a person reads them?
Partly. Google disconnects chats from your account before sending them to service providers, and OpenAI strips usernames and runs a Privacy Filter model that its website says can make mistakes. Neither removes what you typed: a name, address or diagnosis inside the message reaches the reviewer as written, which is why both vendors tell you not to enter confidential information.
Which AI chatbot has no human review at all?
None of the five largest publishes a promise of no human review. Perplexity is the only one whose Privacy Notice names no reviewer, but it also names no retention period and reserves the right to use data to investigate policy violations. A model running on your own hardware is the only setup where no one else can read the chat.
