As we’ve seen time and again, AI is not replacing labor so much as displacing it. If OpenAI announces some great big innovation to their hyper-advanced, globally catastrophic auto-correct, you can bet some poor contractor in the Global South has a big hand in its creation and is being paid peanuts for doing so. Which is a drawn-out way to say that, yes, Copilot users, humans are reviewing your creepiest image generations and are likely being traumatized by them.

This is another story from the intrepid reporters at 404 Media, which reviewed internal documents of Microsoft Copilot contractors, including instruction guides, user prompts and pictures, and the contractors’ internal message boards. They found humans were reviewing Copilot prompts and user image uploads and evaluating the output on how well the system followed the prompt. And, get this, many of the images Copilot users upload include the types of exploitative and sexually explicit material these genAI systems were designed to create, including sexually suggestive images of young girls, breast enhancements, pro-anorexia content, and, apparently, lots of She-Hulk. Though they’re paid to evaluate the quality of output, contractors say on the message board that “It’s hard to take things seriously when my focus has to be what model generated the appropriate bust size or which middle-aged woman was put in the appropriate sexually suggestive position.”

To be clear, these contractors were not hired to moderate these images. So when a Copilot user asks to shorten the skirt of a Bratz doll, the contractor’s job isn’t to report that user to the FBI. Rather, a “workforce of at least hundreds of human reviewers” are evaluating the prompt and its output. After Copilot ingests the prompt and generates the image, the reviewer receives two versions of the output and decides which better depicts “Ariana Grande with anorexia.”

Unseen human labor is commonplace in the AI world. Last week, 404 revealed that Meta’s adorable new AI that makes phone calls for you is actually just humans making phone calls for you. But since the beginning of the AI boom, companies like OpenAI have used human contractors in Venezuela and Kenya to evaluate output, correct it, and refine prompts for a couple of bucks per task. Considering that the people who use these things to generate sexually explicit images of kiddie show characters do not give a shit about the humanitarian or environmental crises these systems are creating, we offer a simple warning: The AI you’re using to fill your spank bank with images of Bluey’s paws is being seen by humans, and it’s making their lives miserable.