Two stories tend to dominate talk about AI and freelance work. One says independent creative and technical work is heading for collapse. The other says the tools are simply a productivity boost, with no structural change underneath.
Both miss what is happening across the market. The effect is uneven. Some work has already largely gone. Some work has grown. Most work has changed shape without changing volume, so the person doing it feels the job becoming different rather than disappearing.
We run a freelance marketplace and an AI products business, so we see the issue from both sides. This is the picture as we see it.
Some standardised freelance work has already gone
Certain categories have contracted sharply and are not coming back. They have a shared pattern. The output was standardised, the quality bar was functional rather than exceptional, and the buyer could evaluate the result themselves.
That matters because the buyer did not need deep expertise to decide whether the work was good enough. If the result looked fine, read fine or matched the template, the job was accepted. AI tools are strongest where that kind of judgement is enough.
The categories under the most pressure are clear:
- Basic copywriting at volume. Product descriptions, meta descriptions, category blurbs and listing text. This was work bought by the hundred at low rates and judged on whether it read acceptably. That market has largely gone, and the honest thing is to say so plainly rather than reassure people.
- Straightforward translation of non-critical content. This excludes legal or medical translation and literary work. It covers the enormous middle of documentation, support articles and marketing collateral where a competent result is sufficient.
- Simple graphic production. Resizing, format variants, background removal and template population. These are the parts of design that were execution rather than decision.
- Basic transcription and data entry. These were already under pressure before this wave, and are now essentially finished as a category.
What connects these areas is the buyer’s ability to look at the output and know whether it is fine. Where the buyer can self-evaluate, the intermediary disappears fastest.
Other categories have grown because AI creates follow-up work
The growth side gets less attention, and it is larger than expected. AI has created work around correction, connection, specification and responsibility. These are not small side effects. They are visible categories in their own right.
- Fixing AI output at scale. Organisations generated enormous volumes of content and code, then discovered it needed editing, verifying, restructuring or replacing. This is real, well-paid and slightly demoralising work, and it is currently one of the fastest-growing categories on any marketplace.
- Integration and plumbing. Every business wants a model connected to something. The model is the easy part. Connecting it to a CRM built in 2014, with authentication, error handling and a sensible failure path, is ordinary software work and there is far more of it than there was.
- Specification and evaluation. This means turning a vague requirement into something measurable. This barely existed as a freelance category two years ago and is now a genuine speciality, because it is the skill that makes AI work sellable.
- Anything with accountability attached. Work where a named person has to stand behind the result has become more valuable, not less. The cheap alternative got cheaper, which widened the gap and made the accountable version more clearly worth its price.
The common thread is that AI output still needs someone to decide what is correct, usable and safe to rely on. Where that decision matters, the work has grown.
Much freelance work has barely moved
The largest group gets the least attention. A lot of freelance work has barely moved, because the hard part was never producing a first draft of something.
Some work is protected by the nature of the job itself:
- Work involving physical presence.
- Work requiring a relationship built over years.
- Work where the difficulty is political rather than technical, such as getting six stakeholders to agree.
No model has any purchase on that last problem. The barrier is agreement, trust and responsibility, rather than a shortage of generated output.
Creative work with a distinctive voice also sits differently. In that market, the buyer is purchasing a specific person’s judgement rather than a category of output. If anything, that market has strengthened, because the volume of competent-but-generic work rose and distinctiveness became scarcer by comparison.
Highly regulated work is another case. Anything where a professional body, an insurer or a regulator requires an accountable human is structurally protected, and will stay protected for reasons that have nothing to do with capability.
The real dividing line is whether the buyer can judge the work
Set out this way, the pattern becomes clearer. The effect is not distributed by skill level or by seniority. It is distributed by whether the buyer can evaluate the output.
Where the buyer can judge quality themselves, they no longer need the freelancer in the same way. The model produces something they can assess directly. Product descriptions, simple formatting and basic transcription fall into that pattern.
Where the buyer cannot judge the output, they need someone accountable. That need has grown. A client may know they want an AI feature, a safer data process or a better generated document, but they may not be able to tell whether the result is correct. In those cases, they are buying judgement as much as output.
This explains why some experienced people have lost work while some juniors have gained it. The question is not only how good you are. The question is whether your client can tell.
The market is splitting, and the middle is under pressure
The consequence is a market splitting into two, with the middle thinning fast.
At one end is volume work at low margin, where the freelancer is competing with a tool the client already has. Surviving there means being dramatically faster. That means using the same tools, which means competing on price against people doing the same. It is a poor position.
At the other end is work sold on accountability, judgement and relationship. Rates in that part of the market have held or risen. There is smaller volume, better margin and a much harder path in.
The middle was the largest part of the freelance market. It was competent execution of well-specified tasks at moderate rates. That is where the pressure is concentrated.
Most people actually work in that middle layer. That is why the aggregate statistics look calm while individual experience is turbulent. The total can appear stable while the work underneath changes sharply.
Stable aggregate numbers can hide a changed market
Industry reports keep finding that freelance earnings are broadly stable. That sits oddly beside the number of people describing their category collapsing. Both are true, and the reason is important.
Totals are dominated by the top of the market. A modest rise at the accountable end can offset a severe fall across a much larger number of smaller engagements, and the aggregate barely moves. The distribution changed shape while the total held.
Survivorship makes the picture look calmer than it feels. Surveys reach people who are still working. Someone whose category disappeared eighteen months ago and took a salaried job is no longer in the sample, and their absence reads as stability.
There is also a lag in how work is described. Someone who used to write product descriptions and now edits generated ones may still be categorised as a copywriter. The label held, the work changed completely, and the statistics cannot see it.
So treat reassuring aggregates with suspicion. Treat individual accounts of collapse as real even when the totals disagree. They are measuring different things.
Three common outcomes show how uneven the change is
These are composites drawn from patterns we see, rather than individual users. They show how different the outcomes can be for people who all appear to be in the same broad freelance market.
The volume copywriter
The volume copywriter was producing forty product descriptions a day for e-commerce clients at a per-piece rate. That work is gone.
The ones who adapted did not become faster copywriters. They moved into brand voice definition and editing. They now sell judgement about what good looks like rather than the output itself. That means fewer clients, higher rates and considerably more competition for each engagement.
The generalist developer
The generalist developer built small business websites and simple integrations. Volume held up, but the nature of enquiries shifted.
Clients now arrive with something half-built by a model and need it finished, secured or rescued. Rates are similar. The work is less pleasant and requires more diagnostic skill than it did.
The specialist consultant
The specialist consultant advises on regulated data handling in a specific sector. Demand has risen noticeably, because every organisation now has an AI question and very few have anyone who can answer it credibly.
This person has done nothing differently and is busier than ever, purely because scarcity moved.
The instructive part is that only one of them changed their skills. The other two were repositioned by the market around them.
Our own data shows more value and fewer very small jobs
We are being careful here, because our platform is young and the sample is not representative of the whole market.
Even with that caution, we see a clear pattern. Average engagement value has risen while the count of very small engagements has fallen. Categories requiring a named accountable person have grown as a share. The fastest-growing single category is work that describes itself as reviewing, correcting or replacing previously generated output.
That last finding is the one we would flag to anyone thinking about this. A large amount of the current work is cleaning up the first wave of enthusiastic adoption. Whether that is a permanent category or a transitional one is the interesting question, and we do not know the answer.
The useful move is toward work the client cannot fully judge alone
Advice in this area is usually either learn to use the tools or find work AI cannot do. Both are too vague to act on.
The more concrete direction is to move toward work where the client cannot evaluate the output themselves. That is where the need for you is structural rather than economic.
The original advice breaks down into a few related moves:
- Move toward work where the client cannot evaluate the output themselves. That is where the need for you is structural rather than economic.
- Sell outcomes rather than hours, because hours are the unit being compressed and outcomes are not.
- Get closer to the problem definition, because specification is the part that has become more valuable rather than less.
- Build a reputation that is portable and specific, because a general claim of competence is exactly what has been commoditised.
None of that is comfortable if your current position is in the middle of the market. It is more honest than telling people the tools are just a productivity boost.
Platforms should be honest about the structural change
Marketplaces have mostly responded to this by adding AI features and staying quiet about the structural effect on the people who use them.
We think the obligations are more concrete:
- Be honest about which categories are contracting rather than continuing to promote them.
- Make specification and accountability visible in the product, because those are what buyers should be paying for.
- Do not charge a percentage on work that has become low margin, because a fee designed for a different economy is a tax on a shrinking business.
- Make reputation portable, because someone whose category has collapsed needs to move without starting again.
That last set of positions is why Open Lance is built the way it is. It was not designed as a response to AI, but the reasoning turned out to point the same direction.
The open question is what happens when systems buy work
Everything above describes work bought by humans from humans. The open question is what happens when the buyer is itself a system.
If an organisation’s process can purchase a defined job programmatically, meaning through software rather than a person choosing each job, the market stops being a marketplace of people and becomes something closer to a supply chain. We are building for that with BotUp, and we are not confident about the timeline or the shape.
What we are reasonably confident about is that accountability remains the scarce good. A system can buy a completed task. It cannot carry responsibility for the result, and someone always has to.
