A new category of freelance work is growing around something AI was supposed to reduce: cleanup. Businesses are increasingly hiring developers, designers and other specialists to repair, refine and finish work that AI already started.
The first wave of generative AI was largely about creation. Businesses wanted to know how quickly AI could write an article, generate an image, build a presentation or produce working code.
The next phase is starting to look different.
Businesses are discovering that generating something quickly and getting something ready to use are not the same job.
In August 2026, Upwork reported that job posts asking people to fix, refine or finish AI-generated work had increased 70% year over year. In software and web development, this type of work has grown 8.3 times since 2023. In design and creative work, it has grown 7.9 times, according to its marketplace analysis of AI remediation work.
These are not simply jobs asking someone to use AI. The analysis specifically looked for projects where AI had already produced something and the client needed a professional to correct, improve or complete it.
That makes the trend more interesting than another measure of AI adoption. Businesses are beginning to spend money not only on generating work with AI, but also on dealing with what happens after that work is generated.
AI has made the first draft much easier to produce
The cost of producing a first version of something has fallen quickly.
A small business can generate a basic website before hiring a developer. A marketing team can create dozens of images before speaking to a designer. Someone with limited programming experience can ask an AI coding tool to build a feature that previously would have required considerably more technical knowledge.
In many cases, that is genuinely useful. The business gets further before it needs specialist help.
What changes is the point at which that help becomes necessary.
Instead of hiring a developer to start with an empty file, a client may now arrive with an application that mostly works but has architectural problems, security issues or bugs that are difficult to trace. A design client may arrive with several generated concepts and ask a professional to turn one of them into something distinctive enough to represent the brand.
The specialist has not disappeared from the process. The specialist is entering later.
That may sound like a small shift, but economically it matters. If AI continues to reduce the cost of producing rough work, more of the value in professional services may move toward reviewing, correcting and taking responsibility for the final result.
Software is showing the shift most clearly
Software development provides the strongest signal in the marketplace data.
AI remediation jobs in software and web development have grown 8.3 times since 2023.
The reason is not difficult to understand. Code can appear finished long before it is actually ready.
An AI coding tool can produce a plausible function in seconds. That function still has to work with the rest of the application. It has to behave properly when users do something unexpected. It needs to be maintainable, secure, testable and compatible with decisions already embedded elsewhere in the codebase.
Those later stages are much harder to judge from a clean-looking block of generated code.
The 2025 DORA research on AI-assisted software development makes that tension visible. Its survey of nearly 5,000 technology professionals found that 90% were using AI at work and more than 80% believed AI had improved their productivity. At the same time, 30% reported little or no trust in AI-generated code. The research also found that AI adoption still had a negative relationship with software delivery stability, even as its relationship with delivery throughput improved.
Those findings are not contradictory.
A developer can produce code faster while the team still spends meaningful time reviewing, testing and integrating what was produced. More output can also expose weaknesses elsewhere in the development process, particularly when testing and feedback systems cannot keep pace.
This is why measuring AI productivity only at the point of generation can be misleading.
A software product is not finished when the code appears on screen. It is finished when the code works reliably enough to ship.
Faster does not always mean less work
A 2025 experiment from METR made this point from another direction.
Researchers ran a randomized controlled trial involving 16 experienced open-source developers completing 246 real tasks in repositories they already knew well. Before the experiment, the developers expected AI to reduce their completion time by 24%.
Instead, with the early-2025 tools tested, tasks took 19% longer when AI was allowed. Even after completing the experiment, the developers still believed the tools had made them faster, according to the study of AI and experienced developer productivity.
That result needs context. The study involved a small group of experienced developers working inside mature projects they already understood particularly well. The researchers explicitly warned against treating the finding as evidence that AI slows software development generally.
The technology has also improved.
When the researchers attempted to repeat the study with newer AI tools, they encountered a different problem. Developers had become so accustomed to working with AI that recruiting people willing to complete tasks without it introduced serious selection bias. In February 2026, the research team said AI was probably providing more benefit than it had during the earlier experiment, but that the newer data was not reliable enough to estimate the size of that improvement.
The larger point remains useful.
AI tools can become faster and more capable while the cost of reviewing their output remains a separate question. Model performance alone does not tell us how much time a professional spends understanding, checking or correcting what the model produced.
The growth of remediation work suggests clients are increasingly paying for exactly that part of the process.
Better AI can still create a bigger repair market
It is tempting to interpret rising remediation work as evidence that AI produces poor results.
That is not necessarily what the data shows.
Even if AI becomes more accurate, businesses are also using it to produce much more material. A lower error rate applied to a far larger volume of output can still create more work for the people responsible for quality.
Software provides a concrete example of why review matters.
A 2025 IEEE study of more than 500,000 human-written and AI-generated code samples compared Python and Java code across defects, vulnerabilities and complexity. The researchers found different weaknesses in each group. AI-generated code tended to be simpler and more repetitive, while human-written code was structurally more complex and showed more maintainability issues. The researchers also found more high-risk security vulnerabilities in the AI-generated samples.
That does not mean AI-generated code is inherently unsafe.
Other research shows how quickly the models themselves are improving at detecting and repairing vulnerabilities. A 2026 study of newer AI models repairing security issues tested GPT-4.1, GPT-5 and Claude Opus 4.1 against vulnerabilities found in real developer interactions. The models detected roughly four-fifths of the confirmed vulnerabilities in the experiment, with successful repairs somewhat below or around that level depending on the model. Their performance had improved substantially compared with an earlier GPT-4o test.
The important detail is that none of the systems caught everything.
The emerging workflow is therefore unlikely to be as simple as either human-written or AI-written work. AI may create part of the output, help inspect it and even assist with repairs, while a professional remains responsible for deciding whether the final result is fit for use.
That is a different role, but it is still valuable work.
AI remediation is becoming a real freelance category
Freelance marketplaces are particularly useful for spotting shifts like this because clients do not need to wait for a new occupation or formal job title to become established.
If a company needs someone to clean up AI-generated code, improve a generated design or fix an AI-built prototype, it can hire for that specific problem immediately.
For freelancers, the opportunity therefore goes beyond learning how to use the latest AI tool. In many fields, the more durable advantage may be knowing how to recognize when the output is not good enough.
A developer who understands debugging, security, architecture and production systems can inspect generated code in ways a casual user cannot. A designer with strong typography, composition and brand judgment can recognize why an AI-generated image looks polished but still fails to fit the business. An editor with subject expertise can tell the difference between fluent writing and writing that is unsupported, inaccurate or simply generic.
These are not entirely new skills. They are existing professional skills being applied to a new source of work.
That makes the trend particularly relevant to specialists using development and technology freelance platforms and design freelance platforms, where clients can bring in specific expertise without creating a permanent position.
It also fits the broader shift toward project-based specialist work documented in Flexable’s report on the global rise of freelancing.
AI remediation suits that model particularly well because demand is often specific and uneven. A company may not need a full-time employee to repair generated illustrations, review an AI-built prototype or debug an application created partly by AI. It may still need an expert for several days or several weeks.
The value of professional work is moving further downstream
The larger story is not simply that AI sometimes creates work that needs fixing.
It is that AI changes where businesses need expertise.
When the first version of something becomes inexpensive to produce, producing that first version becomes less distinctive. More commercial value can move toward the parts of the process where mistakes become expensive or where context matters.
In software, that may mean testing, security, architecture and deployment. In design, it may mean brand consistency, art direction and production quality. In research and writing, it may mean source verification, subject knowledge, accuracy and editorial judgment.
The implications extend beyond freelancing.
Companies may discover that the employees becoming more valuable in an AI-heavy workplace are not necessarily the people generating the most output. They may be the people who can quickly distinguish strong work from weak work, understand why something failed and know what needs to change before it reaches a customer.
That is a different kind of productivity.
It is harder to demonstrate through the number of words written, images generated or lines of code produced. It becomes visible when fewer mistakes reach production, fewer weak outputs move to the next stage and less time is spent repairing problems after they become expensive.
The growth of AI remediation work offers an early market signal that businesses are beginning to pay for that difference.
What the 70% figure does and does not tell us
The methodology behind the headline number is more substantial than a simple keyword count.
The analysis used a 50% sample of job posts across 12 marketplace categories between January 2023 and June 2026. Each post was independently assessed by three different open-weight language models. A job was classified as AI remediation only when all three agreed that AI had created the work the client wanted a professional to repair or finish.
There are still important limits.
The findings describe activity on a freelance marketplace, not every job across the wider economy. Project-based markets can respond to technological changes faster than traditional employment, so the pattern may be more visible there.
The classification was also performed by AI models rather than by human researchers manually reviewing every post.
Most importantly, a 70% increase in remediation work does not tell us whether AI has created more jobs than it has displaced. It measures one specific form of demand that has grown as businesses use AI more extensively.
That narrower finding is still significant.
Three years into the generative AI boom, businesses are not only paying people to use AI. They are increasingly paying professionals to improve work that AI has already produced.
For workers, that suggests a practical change in where expertise may hold its value. Producing the first draft is becoming easier. Knowing whether that draft is accurate, secure, useful and ready for the real world remains a professional skill.
The market is starting to put a price on that difference.