Freelance listings for fixing AI output rose 87% to 10,760 in ten months

The Guardian reported on 2 September that freelance listings for fixing AI output are growing quickly, using internal data shared by three platforms. On Freelancer.com, listings tagged with phrases such as "correct AI", "AI hallucination" and "AI error" rose 87% to 10,760 globally between August 2025 and June 2026. Search Engine Journal covered the findings on 3 September.

The other two platforms point the same way. The Guardian reports that Upwork saw a 70% year-over-year rise in AI remediation work, and that searches on Fiverr for "AI cleanup" services grew more than twentyfold between 2023 and 2026. These are figures the platforms chose to share about themselves, and no independent audit of them has been published.

Where the rework lands

The categories are worth reading closely. Graphic design leads, followed by video editing, proofreading and content writing. Every one of those is a field where a model produces something that looks finished at a glance and fails on inspection, which is exactly the case where a reviewer has to open the file rather than trust the output.

The Guardian's examples give a sense of the effort involved. A multimedia editor, Nathan McConnell, spent two to three hours in Photoshop on each card of a 100-image AI-generated tarot deck. An illustrator, Todd Van Linda, charges $65 an hour and turned down an offer of $500 to repair 13 to 15 AI-generated illustrations. Freelancers quoted in the piece say the fix can take as long as starting from nothing, which is why some of them decline the work.

Rework is a cost that arrives later

The number that matters here is not 87%. It is the observation underneath it: the cheap step produced work that a second person had to be paid to correct, and the correction was priced separately, later, by someone else.

That pattern is familiar to anyone shipping software with model-written code. Generation is the visible cost and it is small. The invisible cost is the time spent finding out whether the output is right, and that time does not shrink when generation gets faster or cheaper. It grows with volume, because there is more to check.

The design and video work in this data has one useful property that code lacks: a person can see a bad illustration. Wrong code compiles, passes a quick read, and reaches production. The reason to build checks that actually run the thing, rather than a review that reads it, is that a human eye catching an error is the good case, and software rarely gives you that.

Three caveats belong on these figures. They cover listings and searches rather than completed work, so they measure demand being posted, not money changing hands. They come from platforms with a commercial interest in the trend they describe. And a rise in listings could reflect people learning to name this work rather than more of it existing.

Source

AI Slop Cleanup Job Listings Up 87%, Report Says, Search Engine Journal, 3 September 2026, reporting on the Guardian's story of 2 September, "Freelancers are getting buried with 'soulless' AI slop cleanup".