Skip to main content

Your ChatGPT Chats Might Be Read by Humans, Investigation Finds

Your ChatGPT Chats Might Be Read by Humans, Investigation Finds

Ever poured your heart out to ChatGPT, assuming it's just you and the algorithm? Think again. A new investigation by 404Media reveals that OpenAI employs hundreds of contract workers to review real user conversations—all in the name of making the model better.

These reviewers, recruited by Crossing Hurdles and paid through AI training firm Mercor, score ChatGPT's responses on a 1-to-7 scale. Their mission? Help reduce annoying habits like excessive flattery and overly human-like replies. One North American reviewer told 404Media their hourly wage topped $50.

But here's the rub: most users have no clue their chats might be read by actual people. Even anonymized conversations can still contain sensitive personal details. OpenAI says it uses privacy filters, but admits they sometimes slip up.

Image

The Default Setting You Might Want to Check

OpenAI currently has an "Improve the model for everyone" option turned on by default. That means your new chats can be used for training and human review—unless you switch it off. Temporary chat mode, on the other hand, doesn't feed your input into model training.

What's more, OpenAI hasn't exactly been shouting about this human review system from the rooftops. Its FAQ, which has mentioned since at least 2023 that authorized personnel and service providers may view user data, advises against entering sensitive information. But 404Media calls the disclosure "somewhat obscure," and it's easy to see why users might not grasp the full scope of human review.

You're Not Alone: Competitors Do It Too

Human feedback has become a cornerstone of refining large language models. Anthropic has confirmed it conducts human reviews of Claude conversations when privacy settings allow—anonymizing them first. Google also lets human reviewers check some Gemini chat logs.

So, is this just the price we pay for smarter AI? The incident shines a light on the uneasy balance between training efficiency and user privacy. Next time you type into ChatGPT, remember: a human might be reading along.

Key Points:

  • OpenAI uses hundreds of contract workers to review real ChatGPT conversations.
  • Reviewers rate responses on a 1-7 scale to reduce issues like excessive flattery.
  • Users can opt out via the "Improve the model for everyone" setting.
  • Privacy filters exist, but OpenAI admits they can fail.
  • Anthropic and Google also use human reviewers for their AI models.