A freelance market has formed around cleaning up what AI produces, and it is now large enough to count. On Freelancer.com, which has more than 88 million registered users, listings containing the phrases "fix AI," "AI hallucination" and "AI error" rose 87% over ten months, reaching 10,760 worldwide. On Upwork, orders to repair AI content grew 70% in a year. On Fiverr, searches for "AI fixing" services are up more than twentyfold over three years. Graphic design is the biggest category, followed by video editing, proofreading and copywriting. The work exists because companies generate a first draft with ChatGPT or Claude and then discover they cannot ship it.
Lisa, a graphic designer in Spain, watched the shift arrive as inbound mail. Before ChatGPT she picked up occasional logo and packaging jobs; afterward she was buried in requests to tidy up AI versions — sharpening blurred images for print, fixing defects, turning failed layouts into files someone could actually work with. It grew until 90% of her incoming logo and packaging requests were AI repair, and that work accounted for 60–70% of her annual income. Clients offered too little for it, some images were bad enough that they had to be rebuilt from scratch, and she was uneasy about taking part in AI copyright violations. She started turning the jobs down and narrowed her offer to design made by humans.
The repair market is growing on top of a creative market that is contracting. A study in Management Science found that in the eight months after ChatGPT launched, listings for freelance work exposed to automation fell 21%, and jobs for creating images fell 17% once image generators appeared; a later study found the decline had continued. At the same time more people are moving into freelancing, because companies are cutting staff and hiring less. The average project on Freelancer.com drew 54 proposals, 8% more than the year before.
Matt Barry, chief executive of Freelancer.com, said the repair jobs come mostly from small companies and entrepreneurs: they use AI to produce a first version, then run into problems they cannot solve themselves. AI can produce a draft, but a person still has to turn it into something usable, and the time and money saved at the first step can be spent again getting the result ready for commercial use — which is sometimes extremely laborious.
The clearest expression of the mismatch is what clients expect to pay. Todd Van Linda, a full-time illustrator in Florida, began taking AI repair work when the ordinary flow of illustration, game art and architectural plans slowed. The job is removing visual noise, raising resolution and hunting defects like an extra toe, and he is direct about the motive: the work was there and it paid. Some clients assume repair should cost less than original work; he does not discount. One prospective client offered about $500 to fix 13–15 illustrations for a children's book on the assumption that each would take roughly 15 minutes. Depending on the file, it takes anywhere from a few hours to a few days. His rate is $65 an hour and it does not go lower.
Kim Dunbar, a writer and editor in Australia, now spends about 60% of her working time on AI repair as blog and website copy jobs dry up; her income from original writing has fallen to roughly a third of its level during the pandemic. Her first job of this kind was editing an article a client had generated, and she learned the recurring failures quickly — monotonous phrasing, awkward paragraph structure, missing emotional shading. She has since worked on around 40 AI-generated nonfiction books, long articles and technical documents. Good jobs for writers and editors, she said, have become noticeably scarcer.
Nathan McConnell, a multimedia editor who freelances alongside a full-time job, started editing AI video — stripping reflections out of real estate clips, fixing robotic voiceover and deformed 3D models. His volume went from two or three such jobs a year to eight or ten now. One client handed him roughly 100 images of humanoid animals for a tarot deck, riddled with extra fingers, two left feet and other errors. At two to three hours per image in Photoshop he expected to finish in nearly two weeks, and stopped after the first ten cards because of health problems and a fussy client who kept changing what he wanted. Some material cannot be saved at all: video with broken physics, shadows falling the wrong way, animation built on another artist's work used without permission. Sometimes he tells the client outright that the file is beyond repair.
Read together, the numbers describe a transfer rather than a new profession. The 87% growth is the sound of cost moving from the generation step, where it is visible and priced as a subscription, to the repair step, where it is billed by the hour to whoever has the least bargaining power in the deal. The $500 offer is the tell: that client had priced 13–15 illustrations by what it cost to produce the broken versions, not by what it takes to fix them. Fifteen minutes an image is not a cynical lowball. It is what the work feels like it should cost to someone who just watched a model generate the originals in seconds.
The thing nobody in this market is pricing is provenance. Lisa stepped back partly because she did not want to be party to AI copyright violations. McConnell has been handed animation built on someone else's work without permission. And clients lie: Lisa has had one insist no AI was involved while the Google Gemini logo sat at the bottom of the image. The freelancer is the last person to touch the file and the one who delivers it, on the strength of an origin story they have no way to check, and none of that exposure is in the hourly rate. It also means the platform counts are floors, not ceilings — 10,760 listings are the jobs where the client was willing to write the word AI.
Dunbar does not find the work dull. She treats it as a puzzle and enjoys turning a weak draft into a clean one. What bothers her is that people increasingly treat AI not as a tool but as the central and sufficient part of the creative process, and she said it is a shame this work has to be done at all. She doubts it lasts: in five to ten years, she expects most of the humanizing work to disappear as models improve and companies can produce usable material on their own. McConnell reads it the other way — AI will still need constant supervision to catch what an untrained eye misses, logos that break brand rules, objects that appear and vanish inside a shot. He thinks people will be fixing AI's problems for a long time.
Van Linda and Lisa are not waiting for that argument to resolve. He stopped taking repair work and found that clients disillusioned with generators began coming to him instead, looking for original drawings made by hand. She updated her profile to say she no longer works with AI-generated images, still gets the requests, and has gone back to canvas, watercolor and the physical art she did before graphic design. She expects that within a year or two AI will do logos and packaging as well as she does, that only the best will earn a living at it, and she does not know whether she is good enough to stay in digital design. Every number in this market points up; the people inside it are building exits anyway, back toward the part of the job that cannot be drafted.