OpenAI employs hundreds of contract workers who read real ChatGPT user conversations to improve the chatbot's responses, according to a 404 Media investigation. The reviewers rate answers on a scale of one to seven and are tasked with reducing excessive flattery and human-like behavior in the model's output. For companies that treat ChatGPT as a working tool, this means part of their correspondence may pass through human hands, not only through servers.
How the review pipeline is arranged
The outlet obtained leaked internal documents and spoke with additional sources. The reviewed prompts are anonymized, but can still contain sensitive personal data. OpenAI uses a privacy filter and admits it can make mistakes. One reviewer told 404 Media he did not think users knew that humans were reading their chats, and the reviewed prompts support that: some conversations included users explicitly asking ChatGPT to keep the contents private.
Recruitment runs through Crossing Hurdles, which describes itself as connecting skilled professionals with the future of AI work, and payment goes through the AI training company Mercor. One North America-based reviewer said they earn more than $50 an hour. The task itself is not free-form commentary: workers score responses on a one-to-seven scale and push the model away from flattery and overly human-like phrasing, which shapes how the system answers paying users next.
Users can turn off the setting called Improve the model for everyone, which is enabled by default, to stop their chats from being used for model training and potentially read by humans. The switch only applies to new conversations. OpenAI also offers a temporary chat mode that, according to the company, does not use input data for model training. The combination means the protection depends on the user's actions before the conversation starts, not after it.
What this means for business users
For a company that pastes contracts, client correspondence or internal reports into ChatGPT, the practical consequence is a change in the data-processing perimeter: the vendor's staff and its contractors become part of the chain. A small team usually has no policy for this and relies on the default settings, while a larger company with a security function is more likely to require an enterprise agreement, a documented list of subprocessors and a rule on which data may not be entered at all.
What remains unclear is how many conversations are actually reviewed, how long they are stored and how the privacy filter is tested. The disclosure itself sits on an FAQ page, where OpenAI states that authorized personnel and service providers may view user data to improve model performance and advises against entering sensitive information; that notice has been online since at least 2023 and has changed only slightly. 404 Media argues the wording is too hidden and too vague, and that the label Improve the model for everyone reads more like a friendly community contribution than consent to have humans read conversations. When asked where users are explicitly told about human review, OpenAI initially did not respond.
Human feedback of this kind remains a key part of improving models, and the practice is widespread. Anthropic confirmed to 404 Media that it uses human reviewers to check Claude responses and improve future models, for users who have Help improve our AI models turned on in their privacy settings, and it strips account details such as email addresses before review. Google also uses human reviewers and notes in Gemini that humans may review saved chats. The sign to watch is whether OpenAI moves the human-review notice out of the FAQ into the consent flow itself: if the opt-out rate then turns out to be high, the industry will have to price that feedback differently.
