The first clear labor-market effects of generative AI are not appearing evenly across freelance work. Research covering more than three million job postings found substantial declines in writing and translation tasks that AI could perform directly, while demand increased for several skills that became more useful because AI existed.
When ChatGPT arrived, one of the biggest questions was whether generative AI would replace knowledge workers or simply make them more productive.
Freelance markets are beginning to provide a more useful answer.
It depends on the work.
A peer-reviewed study published in the Journal of Economic Behavior & Organization analyzed more than three million job postings from a large global freelancing platform and found that demand for skills considered substitutable by generative AI fell substantially after ChatGPT’s public release.
Across the affected categories, demand fell around 20% to 50% relative to the counterfactual trend, with writing and translation among the clearest examples. The research into how ChatGPT changed freelance demand found the strongest losses in short projects that were expected to last no more than a few weeks.
The underlying results are more precise than the headline suggests. Across all skill clusters classified as substitutable, demand was about 25% lower relative to unaffected work after ChatGPT’s release. Certain writing categories fell much further. Real-estate writing declined 52%, writing for About Us pages fell 59%, and Western European language translation declined 23% relative to the estimated counterfactual.
At the same time, total freelance demand did not collapse.
Demand increased in areas that were unaffected by generative AI and in several categories where AI complemented human expertise. Machine-learning projects rose 24%, while demand for AI-powered chatbot development increased 179%.
The early impact of generative AI therefore looks less like the disappearance of freelance work and more like a redistribution of it.
The first work to disappear was the work AI could imitate most directly
Writing is not one task.
A product description, technical manual, investigative article, About Us page, legal analysis and sales email may all contain words, but the skills and information required to produce them are very different.
The research is useful because it avoided treating broad occupations as if every task inside them had the same exposure to AI.
Researchers grouped millions of job postings into 116 detailed skill clusters and classified them according to whether large language models could substitute for the work, complement the worker performing it, or had little direct effect.
Only 12 clusters were classified as primarily substitutable. Fifty-nine were considered complementary to AI, while 45 were classified as largely unaffected.
That distribution is important.
The study did not find that generative AI could simply replace most freelance work. It found a much narrower group of tasks where the technology was capable of producing something close enough to what clients had previously purchased from people.
Some forms of commercial writing fit that description particularly well.
A company that once paid someone to produce a basic property description can now generate a reasonable first version almost instantly. The same is true for a generic company profile, straightforward marketing copy or simple translation between languages where large models perform well.
The closer the deliverable is to text generation itself, the more directly the freelancer competes with the model.
That is very different from work where writing is merely the final format for research, expertise, reporting or judgment.
Short freelance projects took most of the damage
One of the most revealing findings in the study has little to do with job titles.
It is about project length.
The decline in AI-substitutable work was concentrated in projects expected to last one to three weeks. Longer engagements were less affected.
That tells us something important about what AI is replacing first.
Short freelance projects are often easier to define. The client has a relatively clear deliverable, limited coordination is required and the freelancer may need less organizational context to complete the assignment.
That makes them easier to automate.
If a business needs 20 short property descriptions, the task can be specified clearly enough for a language model to produce something useful.
A three-month writing engagement is more likely to involve changing requirements, meetings, research, feedback, subject knowledge and an evolving understanding of what the client actually wants.
Those activities are harder to compress into a single prompt.
The distinction is therefore not simply between writing and non-writing work. It is partly between discrete production tasks and relationships that require continuing context.
For freelancers, that difference may become increasingly important.
A professional whose income depends on a high volume of small, standardized assignments faces a different AI risk from someone embedded in longer projects where the client is paying for expertise as well as output.
Another major study found almost the same pattern
The Oxford-linked research is not the only marketplace study to detect a fall in AI-exposed freelance work.
A separate peer-reviewed study of the impact of generative AI on online freelancing published in Management Science found a 21% decline in job postings for automation-prone writing and coding work relative to manual-intensive categories within eight months of ChatGPT’s launch.
It also found a 17% decline in image-generation-related freelance postings after the introduction of widely used generative image tools.
The second finding is useful because it suggests the pattern is not unique to language.
When a generative technology becomes capable of producing the deliverable a customer previously had to buy from a freelancer, demand for that type of project can fall.
The research also found that competition increased among freelancers remaining in automation-prone categories.
That creates a second effect beyond the disappearance of individual projects.
If the number of available assignments falls while many freelancers continue competing for them, clients gain more choice. Workers may need to submit more proposals, differentiate more aggressively or move toward work that AI cannot perform as easily.
The employment effect can therefore arrive before a profession disappears.
There are simply fewer straightforward opportunities available for the people competing in it.
The surviving work can become more complex
Fewer jobs does not necessarily mean the remaining jobs look the same.
The Management Science research found that the automation-prone projects remaining after ChatGPT’s arrival tended to be more complex and higher paying.
This is a particularly important result for writers.
AI can remove some demand from the bottom or middle of a market while increasing the relative importance of the work that remains.
A client who previously hired a writer to produce a basic article may decide that AI is sufficient.
A client who needs original interviews, technical expertise, credible sourcing, a strong point of view or editorial judgment still has a different problem.
The result is not necessarily a uniform reduction in writing rates.
The market can become narrower and more demanding at the same time.
That creates a difficult transition for freelancers whose careers developed around producing large quantities of reasonably standardized work. The work may still exist, but clients have a cheaper alternative they can compare it against.
Moving upward requires more than producing the same work faster with AI.
It means adding something the model does not already make abundant.
Translation shows why AI substitution is not uniform even inside one skill
Translation is another useful example because the technology does not perform equally well across every language.
The study found a significant 23% decline in Western European language translation relative to the counterfactual.
Translation involving less common languages did not show the same level of decline.
The researchers noted that this pattern was consistent with differences in large-language-model performance across languages. Models trained on enormous quantities of English and widely represented European-language text generally have much stronger capabilities in those languages than in languages with less training data.
That creates a more granular version of AI exposure.
Two freelancers can perform what looks like the same occupation and face very different levels of substitution risk depending on the language pair, subject area and level of accuracy required.
A generic English-to-Spanish translation of simple business material may face substantial competition from AI.
Medical translation, legal translation, literary translation or work involving a lower-resource language can present a very different technical and professional challenge.
The same principle applies well beyond translation.
AI rarely replaces an entire profession with equal effectiveness.
It becomes highly capable at particular tasks first.
The aggregate freelance market did not shrink
This is the part of the research that should prevent the headline from turning into an AI job-loss story.
The researchers did not find an overall decline in freelance demand after ChatGPT’s launch.
Demand increased in skill clusters that were largely unaffected by language models, and some AI-complementary categories grew quickly.
Machine-learning work increased 24%.
AI-powered chatbot development increased 179%, meaning demand was almost three times its previous level relative to the counterfactual.
The growth in those categories does not mean every writer losing basic assignments can simply become a machine-learning engineer.
Skills are not interchangeable.
A professional translator cannot instantly move into chatbot development because that category is growing. Retraining takes time, and some transitions are far more realistic than others.
Still, the aggregate result changes how the market shift should be understood.
Generative AI is simultaneously removing some work and creating other work.
The difficult question is whether the people losing demand possess skills close enough to the new opportunities to move into them.
Job creation does not automatically help the workers whose jobs are being substituted
This distinction is often lost in discussions about technological change.
An economy can create plenty of new work while individual workers still experience significant disruption.
Suppose AI removes 10,000 simple writing projects and creates 10,000 projects involving AI integration, machine learning and chatbot development.
The total number of projects has not changed.
The labor problem has.
The workers qualified for the first group of assignments are not necessarily qualified for the second.
The study found evidence of exactly this kind of unevenness. Although complementary categories did not decline overall, demand for novice workers within complementary work fell.
That result is particularly interesting because controlled AI productivity experiments often find that less experienced workers receive some of the largest productivity benefits from generative tools.
Those two findings can coexist.
AI may make a junior worker inside a company more productive while simultaneously reducing a company’s need to hire an external junior freelancer.
If an experienced employee with AI can perform work that previously required additional junior assistance, the technology complements the employee while substituting for the outside worker.
Whether AI creates or removes a job therefore depends partly on whose productivity is being improved.
Freelancers may experience AI disruption earlier than employees
Online freelance markets are unusually sensitive to technological changes.
A company with permanent employees cannot instantly remove a department every time a new tool becomes available. Employment contracts, organizational processes, management structures, training investments and institutional knowledge slow the adjustment.
Freelance projects are much easier to change.
If a company no longer needs someone to write ten short pages, it simply stops posting that project.
There is no layoff announcement.
No organizational redesign is required.
No occupation needs to disappear from official labor statistics.
Demand simply fails to appear.
That makes freelance platforms useful as an early indicator of which tasks businesses are beginning to automate.
It also means freelancers can feel technological changes before the conventional labor market does.
The study itself describes online labor markets as highly flexible environments that can react quickly to outside shocks.
That is one reason the results are important.
They do not prove traditional employment will experience the same magnitude of change. They show what businesses do when they have relatively little friction preventing them from substituting technology for a particular task.
Other research finds reductions in both freelance employment and earnings
A third body of evidence reaches the question from the worker side rather than counting job postings.
A peer-reviewed study of freelancer employment after generative AI became available found that workers in occupations highly exposed to generative AI experienced reductions in both employment and earnings after ChatGPT’s release.
The researchers also examined image-generating technologies and found similar negative effects among workers exposed to those tools.
Perhaps more surprisingly, strong historical performance did not fully protect freelancers.
The study did not find evidence that workers with better past performance escaped the employment effects. It found suggestive evidence that some top freelancers were disproportionately affected.
That result is worth taking seriously because it challenges a comforting assumption about AI disruption.
Being excellent at a task does not necessarily protect a freelancer if technology dramatically reduces how often clients need to buy the task in the first place.
Quality matters when customers are choosing among workers.
It has less power when customers are deciding whether to hire anyone at all.
Writers may need to separate the task from the profession
A useful way to interpret all three studies is to stop asking whether AI replaces writers.
The better question is which parts of writing clients still need to purchase from another person.
AI can already generate competent sentences, summarize documents, rewrite text, translate common languages and produce familiar content structures at extremely low marginal cost.
Those capabilities put direct pressure on work where producing text is most of the value.
Professional writing can contain many other activities.
A journalist reports.
A technical writer understands a product and the reader who must use it.
A subject specialist knows which facts matter.
An editor identifies weak arguments and unsupported claims.
A brand writer understands how a company wants to sound and why.
A researcher finds evidence that does not already exist inside the model’s context.
These activities eventually produce words, but the words are not the entire product.
That distinction is likely to become increasingly important as generated text becomes abundant.
Clients can automate the first draft and still hire someone later
Falling demand for straightforward writing does not necessarily mean human involvement disappears from the workflow.
It can move downstream.
A company may generate the first draft internally and later hire an editor, subject specialist or strategist to improve it.
A business may use AI to translate thousands of words and then pay a professional to review legally sensitive or customer-facing material.
A marketing team may generate many concepts before hiring someone to decide which ones are worth developing.
Flexable’s research into the growth of jobs fixing AI-generated work shows how visible this second-stage market is becoming. Businesses are increasingly hiring professionals specifically to correct, refine and finish output that has already been produced by AI.
That does not perfectly replace the work being lost.
Editing an AI-generated article may pay differently from writing one from scratch. The required skills may change. Clients may expect faster turnaround because a draft already exists.
But it illustrates how automation can rearrange a workflow rather than simply removing every human from it.
The client stops paying for the part AI can produce cheaply and concentrates human spending on the stages where errors, context and judgment matter more.
Trust creates another reason human writing does not simply disappear
There is also a difference between being able to generate content and being willing to rely on it.
For low-risk material, that distinction may barely matter.
If a company needs dozens of basic descriptions that can be checked quickly, AI may be economically attractive even if the output occasionally needs correction.
The calculation changes when mistakes are expensive.
Research, financial communication, medical information, high-value marketing, legal material and executive thought leadership all create greater consequences when information is wrong, unsupported or inappropriate.
Flexable’s report on why clients trust human and AI work more than AI acting alone shows the commercial importance of that distinction.
This suggests a likely segmentation of the writing market.
At one end, clients care mainly about obtaining acceptable text cheaply and quickly. That work faces direct pressure from AI.
At the other, the client needs someone to be accountable for whether the content is accurate, defensible, original or appropriate. AI can still play a major role in production, but human responsibility remains part of what is being purchased.
The market between those two extremes may be where competition becomes hardest.
The most vulnerable work has three characteristics
The research does not provide a universal formula for predicting which freelance jobs will decline next.
It does suggest a useful pattern.
The work most vulnerable to direct AI substitution tends to be relatively easy to specify, easy for the model to imitate and easy for the buyer to evaluate without maintaining a long relationship with the worker.
A short About Us page fits those conditions well.
The client can describe the company, request several hundred words and judge quickly whether the result is acceptable.
A complex research assignment is different. The client may not know in advance which evidence matters, which sources will prove reliable or what conclusions the research will support.
The professional has to make decisions during the work.
That distinction provides freelancers with a more useful way to assess AI exposure than simply asking whether a model can perform part of their occupation.
The question is how much of the client’s problem remains once the model has generated its answer.
If almost nothing remains, the service is highly exposed.
If substantial verification, context, judgment or implementation still remains, there is more space for professional value.
Moving into AI-complementary work is easier said than done
The obvious career advice is to move away from tasks AI can substitute and toward work AI complements.
That advice is directionally sensible but incomplete.
Skills take time to build.
A copywriter cannot simply declare themselves an AI systems consultant. A translator cannot become a machine-learning engineer because the demand data looks attractive.
Real career transitions happen through nearby skills.
A writer might move deeper into subject specialization, editorial strategy, original research, interviewing or content that requires access to people and information a model does not possess.
A translator might specialize in regulated material, localization, cultural adaptation, quality assurance or languages where model performance remains weaker.
A designer might move toward art direction and brand systems.
A developer might move from producing basic code toward architecture, integration, security and AI implementation.
The relevant question is therefore not only which jobs AI is creating.
It is which of those jobs are realistically reachable from the skills a worker already has.
That is a much harder labor-market problem than simply counting job creation.
What the research actually measured
The headline study examined daily job-posting data from a large global online freelancing platform.
The dataset covered more than three million postings, with the broader collection spanning January 2021 to September 2023. The core analysis compared demand before and after ChatGPT’s public release on November 30, 2022.
Researchers used topic-modeling techniques to organize postings into 116 detailed skill clusters based on the skills employers requested.
They then classified those clusters into three groups: work that large language models could substitute, work that the technology could complement and work that was largely unaffected.
The final classification contained 12 substitutable clusters, 59 complementary clusters and 45 unaffected clusters.
The researchers used a difference-in-differences design, treating the unaffected clusters as a comparison group. This matters because freelance demand can rise or fall for reasons unrelated to AI.
The reported declines therefore are not simply the raw difference between the number of writing jobs before and after ChatGPT.
They estimate how demand for AI-substitutable work changed relative to how those categories would have been expected to move alongside unaffected freelance work.
That is what “relative to the counterfactual” means in the headline finding.
The headline range should not be read as a universal decline in writing
This limitation is especially important for a report built around the 20% to 50% figure.
The study does not establish that the entire freelance writing industry contracted by somewhere between 20% and 50%.
It found substantial declines across specific skill clusters categorized as substitutable by large language models.
Across the substitutable group as a whole, demand fell about 25% relative to the control group after ChatGPT’s release.
Individual categories varied considerably.
Real-estate writing fell about 52%.
About Us page writing fell about 59%.
Western European language translation fell around 23%.
Other forms of writing and translation were less affected.
The estimates also describe one large online freelance market during the relatively early period following ChatGPT’s launch. They should not automatically be generalized to every freelance marketplace, every country or traditional salaried employment.
The data ended in September 2023, when generative AI was considerably less capable than it is today. That makes the findings historically important, but it also means they are evidence of the first phase of the transition rather than a measurement of today’s entire market.
The deeper shift is from selling output to selling what surrounds it
The most useful conclusion from the research is not that writers and translators have no future.
It is that producing a familiar output is becoming a weaker professional moat when software can produce a similar first version almost instantly.
That changes where freelancers need to create value.
Some will move toward harder subject matter. Some will build deeper client relationships. Others will specialize in verification, editing, research or strategy. Many will use the same AI systems that reduced demand for simpler assignments to become faster at the work that remains.
Meanwhile, entirely new categories of freelance work will continue to appear around implementing, supervising and improving AI itself.
The study captures both sides of that transition.
Demand for several easily substituted services fell sharply.
Demand for AI-powered chatbot development nearly tripled.
Those are not contradictory outcomes. They are two consequences of the same technology changing what businesses still need to buy from people.
For freelancers, the important question is no longer simply whether AI can perform part of their work.
It is whether clients still need enough expertise around that output to justify hiring them.