Companies spent two years telling investors that AI would replace expensive human labor. Now some of those same companies quietly rehire humans, at lower rates, to fix what the AI produced. Nobody calls this a reversal. They call it “quality assurance,” or “editing,” or just cleanup.
New reporting this week lays out exactly who’s doing that work, and what it costs them.
The Job Title Nobody Applied For
Lisa Carstens, an illustrator in Spain, didn’t set out to fix broken logos for a living. But that’s what fills her inbox now: brand marks with fuzzy lines, garbled text baked into the image itself, layouts that fall apart the moment you look past the thumbnail. Sometimes she can patch it. Often the job is a full redesign, wearing a client’s hope that “just clean it up a bit” would be enough.
In Georgia, writer Kiesha Richardson has watched half her gigs quietly change shape. She isn’t drafting anymore. She’s untangling AI copy full of the same tells anyone who talks to a chatbot recognizes on sight: the em-dash habit, the reflexive “delve,” the “deep dive” that promises depth and delivers a paragraph. The work takes real time. The pay doesn’t reflect that. “I am a bit concerned,” she says, “because people are using AI to cut costs, and one of those costs is my pay.”
In India, developer Harsh Kumar inherits a messier problem: codebases built through what people now call vibe coding, shipped without anyone who fully understood them signing off. Some of it leaks system details it shouldn’t. Kumar isn’t especially worried about his own future in it. “Humans will be required for long-term projects,” he says. “Humans developed AI.”
Why the Pricing Is the Real Story
The easy headline is “AI creates new jobs, actually.” That’s not quite it.
New job categories aren’t remarkable by themselves. Automation has always displaced some labor and created other labor. What’s notable here is the price attached to this particular kind of work. Clients pay cleanup rates for something that often requires a redo from scratch. The pricing assumes fixing is easier than making. The people doing the fixing say that assumption is wrong often enough to matter.
A logo with corrupted typography isn’t a ten-minute polish. It’s a new design with worse starting material. Rewriting AI copy that’s structurally hollow isn’t editing. It’s writing, done under the fiction that it wasn’t. That gap between the labor and its price tag is wage compression wearing a productivity costume.
Slop Now Has a Supply Chain
What’s forming underneath this is closer to an industrial process than a labor trend: generate at scale, cheap and fast, then route the output through a human pass before anyone notices the seams. AI didn’t eliminate the editorial layer. It relocated it downstream and gave it a smaller paycheck.
That has a name circling the discourse already: AI slop, and it now has its own remediation market attached to it. Marketplaces report a real surge in requests for error correction and human oversight, across design, copy, and code. This isn’t a niche complaint from workers who feel threatened. It’s how a lot of AI output currently reaches customers: ship first, patch with people later.
The pattern isn’t limited to freelance marketplaces, either. In China, some companies now require staff to document their own workflows in detail so a system can learn the job before the person doing it gets phased out, a version of the same trade playing out one rung up the ladder.
Better Models Won’t Fix This on Their Own
The obvious rebuttal: this is a transitional phase, and model quality keeps improving, so cleanup demand should shrink. Maybe. But that bet assumes the bottleneck is capability. What these freelancers describe sounds more like a verification problem. Nobody upstream checks the output before a client sees it. Better models cut the rate of obvious errors. They don’t remove the incentive to skip a check when the check costs money and shipping fast doesn’t.
There’s a quieter cost too, one that shows up less in productivity numbers and more in how people talk about the work. Carstens and Richardson both describe the job as creatively empty, polishing something that was never really made. That reads less like a productivity story and more like an early symptom of something researchers have started giving a name: a chronic anxiety about professional irrelevance that sets in whether or not anyone actually loses their job. The freelancers absorbing cleanup work today are, by definition, skilled enough to catch what AI misses. If the pay never matches the labor, the obvious move is for the most capable people to walk away from cleanup work entirely, leaving it to whoever will do it cheapest. That’s the same dynamic that produced the slop in the first place.
It’s also why a lot of younger freelancers already treat single-client stability as a bad bet on principle, stacking three or four income streams instead of betting on one relationship holding steady. Cleanup work becomes one line item among several, not the whole plan.
Companies didn’t remove humans from the loop. They moved them to the end of it, where mistakes cost more to catch, and nobody gets much credit for catching them.
Related: Is AI Close to Human Intelligence? What 2026 Reveals
