You might assume that when you type a question into ChatGPT, Claude, Gemini, or Copilot, it’s between you and your machine. That assumption is often incorrect. A new investigation from 404 Media reveals that OpenAI pays hundreds of contractors, some earning more than $50 an hour, to read real ChatGPT conversations and rate how the AI responded, an internal effort code-named “Project Lily”. Anthropic has confirmed it runs a similar human review process for Claude.

• Reviewers don’t see your username, but they can see a summary of what you’ve asked the AI in the past, which can reveal your profession, location, or personal circumstances without a name attached. Automated filters try to strip out personal details before a human sees the conversation, but every company that does this admits sensitive information still slips through.
• Opting out of data sharing is never a full guarantee. Across the major platforms, giving a thumbs-up or thumbs-down on a response, tripping an automated safety flag, or simply using a personal account instead of a business one can all put your conversation in front of a human reviewer no matter what your settings say.

Treat any AI chat window the way you’d treat an email to a coworker or a post to a social media platform: assume it could be read by someone you’ve never met. Never put in a client’s confidential information, health details, financial account numbers, or anything you wouldn’t want to see outside your own head. If your company provides a business or enterprise AI account, use that one for work tasks instead of a personal login, and if you’re on a personal plan, check your privacy settings rather than assuming the defaults protect you.