Cleaning up after AI is a growing job, priced like a quick fix

A client offered Todd Van Linda about $500 to fix the pictures in a children's book. There were thirteen to fifteen of them, all generated by AI, and the client figured each one would take about fifteen minutes. Van Linda, a freelance illustrator in Florida, charges $65 an hour, and he told the Guardian's Aaron Mok that his rate "doesn't go down from there." He turned the job down.

Run the numbers and the offer gets strange. Fifteen minutes a picture comes to under four hours, and $500 for under four hours is more than double his rate. On paper, the client was paying him well. The fifteen minutes sank it. Van Linda says jobs like this take anywhere from a few hours to a few days.

That's one offer from one client. The freelancers Mok interviewed describe the same expectation, though: cleanup should be quick and cheap, and it can take as much work as making the thing from scratch.

Human judgment is back in demand

For a couple of years the hopeful story about AI and human work ran like this. Once machines made content nearly free, human work would become the premium product. The machine would draft, a person would decide, and the deciding is what people would pay for.

The first half of that came true fast. Ozge Demirci, Jonas Hannane and Xinrong Zhu looked at 1.2 million job posts on a large freelancing platform they don't name. In the eight months after ChatGPT came out, posts for writing, coding and engineering work fell 21% more than posts for work the chatbot couldn't do yet, like data entry and video editing. Writing took the biggest hit, at around 30%. Once the image generators arrived, posts for graphic design and 3D modeling fell 17% more. It's a 2024 working paper, and its more conservative models put the main drop closer to 16%. Either way, fewer clients were paying a person for the first draft.

Some of that demand has come back, at the other end of the job. On Freelancer.com, listings tagged "correct AI," "AI hallucination" or "AI error" rose 87% to 10,760 between August 2025 and June 2026, according to internal data the platform shared with the Guardian. That's a ten-month window with no published method behind it, from a platform that benefits from a "humans are still needed" story, so I'd use it for the direction of travel and nothing else.

Most of those listings are graphic design, followed by video editing, proofreading and content writing. Writing and graphic design both lost a lot of original work in the Demirci study. That study stops in mid-2023 and doesn't name its platform, so I wouldn't lay one number over the other. The categories line up anyway.

Matt Barrie, Freelancer.com's CEO, says most of the buyers are small companies and entrepreneurs whose AI first pass ran into problems they didn't have the skills to solve. What they saved on the draft gets eaten, he says, by the "incredibly time-consuming" work of making it usable.

So people are paying for human judgment again. They're paying for it as repair, and repair sets both the rate and the status. That's good news for my side of the argument, that people still matter in the process, and a worse deal for the people supplying it.

What a fix is worth

Work booked as a fix gets priced against the mess. It arrives with the frame already set: someone else picked the tool, the prompt, the style, the deadline, and whether the thing should exist at all. The person hired to repair it inherits every one of those decisions and gets paid to make the result usable. The judgment is real. It shows up after everything that caused the problem.

The two freelancers in the story who found a premium found it by leaving. Lisa, a designer in Spain who asked the Guardian not to use her surname, says 90% of her logo and packaging requests in 2025 were AI cleanup, and that work made up 60 to 70% of her income. Clients lowballed her. Some designs were so broken she rebuilt them from scratch. By the end of the year she was turning the jobs down ("It was just so soulless," she said), and she now markets "human-made designs." Van Linda stopped taking cleanup work in late 2025. Since then he's seen clients, some of them unhappy with what AI gave them, come to him for original, hand-drawn work.

Some people like the work. Kym Dunbar, a writer and editor in Australia, says cleanup is about 60% of her workload, and she finds it "interesting" and "challenging." The Guardian's sample also leans toward people who left, so two exits don't make a trend. They do point the same way as the research.

Marketing research has a name for why people pay more for handmade things. In a 2015 set of studies in the Journal of Marketing, Christoph Fuchs, Martin Schreier and Stijn van Osselaer found that people rated handmade products as more attractive and paid more for them as gifts, largely because they believed some of the maker's care had gone into the object. The premium follows who made the thing. A repair carries none of it, because nobody looking at the finished picture knows a person spent an afternoon on it.

Announcing the human doesn't fix that. In a 2026 experiment on short-form video, Han Sol Lim and colleagues found that a "human-made" label did no better than no label at all on how much effort viewers thought went into the work. People already assume a person made what they're looking at. The one they never picture is the person who cleaned it up.

The step at the wrong end

Nielsen Norman Group has a name for the in-house version of this. In an essay in late August, Anna Kaley and Raluca Budiu call it the "custodial era of UX." AI lets teams build faster than designers can evaluate, so UX increasingly meets the work after it exists. Their fix is to put what designers know into the generation step itself, through design systems, content standards and lists of known bad patterns, so people stop repairing the same predictable problems at the end. It's an essay with no data, from a firm that sells UX training, and it says nothing about how common the pattern is. I'd take the recommendation anyway.

The freelance market and the product team have built the same pipeline. The machine goes first, and a person arrives once something has gone wrong. It's the handoff failure I traced in the Chicago Sun-Times reading list with ten invented books on it, where checking whether a book existed was trivial and was nobody's job. Here the checking is somebody's job. It's been booked for the one point in the process where it costs the most and can change the least.

By then the frame is set and the deadline's gone, and the only decision left is how much of the mess to fix. The person with the most useful judgment in the whole chain sees the work last, and is expected to make it presentable in whatever time someone else guessed it would take.

Human judgment bought at the end of the pipeline can only repair. Bought at the start, it can still say no.