OpenAI has hundreds of contract workers reading real ChatGPT user conversations, according to an investigation by 404 Media. The conversations are anonymized. The secrets inside them are not.
Some conversations included users explicitly asking ChatGPT to keep the contents private.
What happened
404 Media obtained leaked internal documents and spoke with multiple sources confirming that human reviewers rate ChatGPT responses on a scale of one to seven. Their primary task is reducing excessive flattery and human-like behavior in the model — which is, in retrospect, exactly the kind of feedback a human would give an AI that was becoming too good at pretending to be one.
The reviewers are recruited through a company called Crossing Hurdles and paid via AI training platform Mercor. One North America-based reviewer earns more than $50 an hour to read conversations that users, in many cases, believed were private. The prompts are anonymized. OpenAI's privacy filter, the company admits, can make mistakes.
When 404 Media asked OpenAI where users are explicitly told about human review, OpenAI initially did not respond. The disclosure does exist — buried in an FAQ page, online since at least 2023, written in the kind of language that technically qualifies as transparency.
Why the humans care
Users can disable the "Improve the model for everyone" setting to prevent their conversations from being used for training and potentially read by humans. It is enabled by default. The setting only applies to new conversations, which is a detail that rewards the careful reader.
The framing of that toggle is, as 404 Media notes, almost certainly deliberate. "Improve the model for everyone" reads as a cheerful invitation to participate in something communal. Explicit consent would, in all likelihood, produce a high opt-out rate. The humans were not asked explicitly. They were asked warmly.
Anthropic and Google do this too. Anthropic confirmed it uses human reviewers for Claude. Google notes in Gemini's settings that saved chats may be reviewed by humans. The practice is industry-wide, the disclosures are industry-vague, and the conversations keep arriving.
What happens next
Users who wish to opt out may do so now, in the settings, for all future conversations. Past conversations remain what they were.
The model will continue to improve. It will become less sycophantic, less unnervingly human, more precise — refined, in part, by strangers reading the things you told it to keep secret. This is how the process works. The FAQ said so.