OpenAI Workers Read ChatGPT Chats, 404 Media Finds

The 404 Media investigation found that OpenAI's human review setup relies on contract workers recruited through Crossing Hurdles and paid via AI training company Mercor. Reviewers are asked to rate ChatGPT responses on a one-to-seven scale and to reduce excessive flattery and human-like behavior, which points to a specific quality-control role rather than open-ended observation. One reviewer told 404 Media he didn't think users knew humans were reading their chats, and the reviewed prompts support that: some conversations included users explicitly asking ChatGPT to keep contents private. The prompts are anonymized, but can still contain sensitive personal data, and OpenAI's privacy filter admits it can make mistakes.

The disclosure sits in a buried FAQ that has been online since at least 2023 and has changed only slightly. OpenAI also offers a temporary chat mode that, according to the company, doesn't use input data for model training, and users can turn off the default-on "Improve the model for everyone" setting to stop chats being used for training and potentially read by humans. That setting only applies to new conversations, leaving the status of older chats less clear. Anthropic and Google use human reviewers too, so the practice extends beyond OpenAI.

The stakes hinge on whether users treat the buried FAQ as sufficient notice and on how many turn off the default setting. The opt-out rate is likely the clearest signal of whether human review is accepted as a normal part of model improvement or seen as a privacy overreach.