Glonce

OpenAI has hundreds of contract workers reading ChatGPT conversations

securityThe Glonce editorsThe Decoder🔥72Glonce rating72Impact68Novelty48Relevance100Surprise55Players92Credibility85
OpenAI has hundreds of contract workers reading ChatGPT conversations

OpenAI employs hundreds of contract workers who read real ChatGPT user conversations to improve the chatbot's responses, according to a 404 Media investigation based on leaked internal documents and additional sources. The reviewers rate ChatGPT responses on a scale of one to seven and are tasked with reducing excessive flattery and human-like behavior in the model's output. The workers are recruited through a firm called Crossing Hurdles and paid through the AI training company Mercor; one North America based reviewer told 404 Media they earn more than $50 an hour.

One reviewer said he didn't think users knew that humans were reading their chats, and the prompts he saw backed that up: some conversations included users explicitly asking ChatGPT to keep the contents private. The prompts reviewers see are anonymized, but 404 Media reports they can still contain sensitive personal data, and OpenAI acknowledges its privacy filter can make mistakes. Users can turn off the "Improve the model for everyone" setting, which is enabled by default, to stop new chats from being used for training and potentially read by humans, though the setting only applies going forward, not to past conversations. OpenAI also offers a temporary chat mode that the company says does not use input data for training.

404 Media reports that OpenAI's only disclosure of human review sits in a buried FAQ entry that has existed since at least 2023 and has changed only slightly since. When 404 Media asked where users are explicitly told about human review, OpenAI initially did not respond. The outlet argues the FAQ notice is too hidden and too vague for users to grasp how much human review actually happens; a setting named "Improve the model for everyone," meanwhile, reads more like a friendly community contribution than consent for humans to read private conversations.

The practice is not unique to OpenAI. Anthropic confirmed to 404 Media that it uses human reviewers to check Claude responses and improve future models, applying to users who have "Help improve our AI models" turned on; Anthropic said it anonymizes conversations before review by stripping account details such as email addresses. Google also uses human reviewers and notes within Gemini that humans may review saved chats.

Key facts

  • OpenAI employs hundreds of contract workers, recruited through Crossing Hurdles and paid via Mercor, who read real ChatGPT conversations and rate responses on a one-to-seven scale; one North America based reviewer reported earning more than $50 an hour.
  • Reviewers are tasked with reducing excessive flattery and human-like behavior in ChatGPT's output; prompts are anonymized but can still carry sensitive personal data, and OpenAI's privacy filter can make mistakes.
  • The "Improve the model for everyone" setting is on by default and controls whether new chats can be used for training and human review, but it does not apply retroactively; a temporary chat mode reportedly excludes input from training.
  • OpenAI's only disclosure of human review sits in a buried FAQ entry that has existed since at least 2023, and OpenAI initially did not respond when 404 Media asked where users are told about it.
  • Anthropic and Google confirmed similar human-review practices: Anthropic reviews Claude chats for users who opted into "Help improve our AI models" after stripping account details, and Google notes in Gemini that humans may review saved chats.

Why it matters

It puts a number and a mechanism behind something users generally suspect but rarely see confirmed: conversations people treat as private, including ones where they explicitly ask the chatbot to keep things confidential, are read by paid outside contractors as a matter of course. The story also shows this is not an OpenAI quirk; Anthropic and Google run comparable programs, so the practice is closer to an industry norm for how conversational AI gets tuned than an isolated lapse.

Who it affects

Anyone using ChatGPT with the default settings, since "Improve the model for everyone" is on by default and only stops future conversations once switched off. It also touches Claude users who have "Help improve our AI models" enabled and Gemini users, since Google discloses similar review inside the product. The contract reviewers themselves, recruited through Crossing Hurdles and paid through Mercor, are also part of the story: one North America based reviewer described being paid over $50 an hour to read and score other people's chats.

How to use it

Users who want to keep ChatGPT conversations out of human review can turn off "Improve the model for everyone" in settings, though this only protects conversations going forward, not ones already logged. OpenAI's temporary chat mode is presented by the company as not feeding training data at all, which is the closer option for a one-off sensitive conversation. Anthropic and Google users concerned about the same issue should check their own equivalent settings, "Help improve our AI models" for Claude and the review notice inside Gemini.

How solid is it

The claims come from 404 Media's investigation, which the outlet says draws on leaked internal documents and additional sources, including two reviewers who described the one-to-seven rating scale and their pay. Anthropic and Google are quoted confirming that human review of user conversations happens at their companies, while OpenAI's own practice is documented through its FAQ page rather than a direct response to the outlet's question; together these corroborate the central claim even though the specific internal documents are not independently verifiable from the outside.

Risks and caveats

Anonymization is not the same as erasure: the article states prompts reviewers see are stripped of some identifying information but can still contain sensitive personal data, and OpenAI itself admits its privacy filter can make mistakes. The opt-out is partial, since it only covers new conversations and leaves past ones exposed to review under the prior setting. 404 Media's own framing is that OpenAI's disclosure, a since-2023 FAQ line about "authorized personnel and service providers," is too vague and too hard to find for users to understand what they are agreeing to; OpenAI did not immediately answer the outlet's question about where else this is disclosed.