OpenAI uses hundreds of contractors to review actual ChatGPT conversations as part of efforts to improve the chatbot, according to an investigation by 404 Media.
The findings were reported on September 14, 2026.
Drawing on leaked internal documents and additional sources, the investigation describes reviewers scoring ChatGPT responses from one to seven. Their assignments include helping reduce excessive praise and behavior that makes the chatbot seem human.
One reviewer questioned whether users realized people could read their conversations. Some reviewed prompts contained explicit requests for ChatGPT to keep information private. Although the prompts are anonymized, sensitive personal information can remain. OpenAI applies a privacy filter but acknowledges that it can fail.
According to the investigation, Crossing Hurdles recruits the contractors, while AI training company Mercor handles payment. A reviewer based in North America reported earning more than $50 per hour.
Users can disable Improve the model for everyone, which is on by default, to prevent new conversations from being used for training and potentially undergoing human review for that purpose. Changing the setting does not apply retroactively to earlier chats. OpenAI also says its temporary chat mode excludes input data from model training.
Human review disclosure draws scrutiny
OpenAI initially did not respond when 404 Media asked where it explicitly informs users about human review. An OpenAI FAQ explains that authorized staff and service providers may access user data to improve model performance, and warns users not to submit sensitive information. That disclosure has existed since at least 2023, with only minor changes.
404 Media argues that the notice is difficult to find and insufficiently specific about the extent of human review. It also criticizes the training setting's wording for presenting participation as a community benefit rather than clearly explaining that people may read conversations.
The coverage suggests that this framing may be deliberate and that an explicit consent request could lead many users to opt out. Those suggestions are interpretations, rather than established findings about OpenAI's intentions or user behavior. Human feedback nevertheless remains an important part of model improvement for AI developers.
Anthropic and Google also use human reviewers
Anthropic confirmed to 404 Media that human reviewers assess Claude responses to help improve future models. This applies to users who enable Help improve our AI models in their privacy settings. Before review, Anthropic anonymizes conversations by removing account information such as email addresses.
Google also employs human reviewers. Gemini includes a notice that saved conversations may be reviewed by people.