Laid-off lawyers, PhDs, and scientists are now working gig jobs teaching AI how to do their former careers. It's a dystopian feedback loop: you lose your job, then you train the AI that made you redundant, so it can take more jobs. This is the quiet part of the AI revolution that nobody wants to talk about.
According to an investigation by The Verge, knowledge workers are earning minimum wage on platforms like Mercor to annotate legal documents, write example code, and evaluate AI outputs in their area of expertise. The companies hiring them are building AI systems designed to automate exactly the jobs these workers used to have.
Let me be clear about what's happening here: we're not creating new jobs; we're creating a gig economy where knowledge workers earn $15/hour to make themselves obsolete. The people who spent years getting PhDs in physics are now explaining physics problems to AI models that will replace physics PhD holders. The lawyers who lost their jobs to AI legal research are training the next version to be even better.
I've been in those early-stage startup rooms. This is exactly how they pitched it working. "We'll democratize expertise," they said. "We'll make high-end knowledge accessible to everyone." What they meant was: we'll extract knowledge from experts, put it in a model, and then fire the experts. The extraction is permanent; the employment was temporary.
The technical reality is that AI systems need human feedback to improve. Reinforcement Learning from Human Feedback (RLHF) requires domain experts to evaluate outputs, rank responses, and explain why one answer is better than another. You can't train a legal AI without lawyers. You can't train a medical diagnosis AI without doctors. The expertise has to come from somewhere.
But here's the cruel irony: the better these workers do their jobs, the faster they make their expertise obsolete. Every well-annotated example makes the AI a little better. Every thoughtful evaluation trains the model to think more like an expert. Eventually, the model is good enough that you don't need as many humans in the loop. Then you need fewer. Then you need almost none.
