Hundreds of contract workers are reading your real ChatGPT conversations. That’s not a conspiracy theory; it’s the current operational reality for OpenAI, according to a new investigation by 404 Media. The outlet secured leaked internal documents and spoke with sources who confirmed that these reviewers rate AI responses on a scale of one to seven. Their job isn’t just to check for accuracy. They specifically target excessive flattery and overly human-like mannerisms in the model’s output. This human-in-the-loop system is designed to fine-tune the bot, but it comes with a significant privacy trade-off that most users never explicitly consented to.
The reviewers are recruited through Crossing Hurdles and paid via Mercor, with some North America-based workers earning more than $50 an hour. One reviewer admitted they didn’t think users knew humans were scanning their logs. The data backs up this concern. Many reviewed prompts included users explicitly asking ChatGPT to keep the conversation private. While OpenAI anonymizes the data and uses privacy filters, the company admits these safeguards can fail. Sensitive personal information can slip through. The disclosure about this practice is buried deep in an FAQ page, stating that authorized personnel may view data to improve performance. It has been there since 2023, but few read the fine print.
This isn’t an OpenAI-only problem. Anthropic confirmed to 404 Media that it also uses human reviewers for Claude, provided users have the improvement setting enabled. Google does the same for Gemini. The industry standard relies on this human feedback loop to refine models, but the framing of consent is questionable. The default setting, labeled “Improve the model for everyone,” sounds like a benign community contribution. It doesn’t scream “human eyes will read your secrets.” If companies asked for explicit, informed consent, opt-out rates would likely skyrocket. That’s probably why they don’t.
You can stop this today. Go into your settings and turn off “Improve the model for everyone.” Note that this only applies to new conversations. For immediate privacy, use the temporary chat mode, which OpenAI says doesn’t use input data for training. But ask yourself: if the best way to keep your data private is to dig through multiple menus to disable a default feature, how many users actually do it? The convenience of a smarter bot is being subsidized by your privacy, and the cost is hidden in plain sight.
