A study of more than 100,000 people published in Scientific Reports in January 2026 found that large language models now beat the average human on standard divergent-thinking tests. The same study found the top decile of human participants still beats every model tested, and so does the top half taken together. Meanwhile two marketplaces report clients buying more human writing, video and design, and nearly half of business leaders in Upwork's own survey say they would pay a premium for creative independent talent. The question is not whether AI is creative. It is which half of the distribution you are selling from.
"AI is more creative than humans" makes a better headline than what the research found. The accurate version is narrower and far more useful if creative work pays your rent.
What the January 2026 study found
Antoine Bellemare-Pepin, Karim Jerbi, Yoshua Bengio and colleagues at Université de Montréal compared GPT-4, GPT-4-turbo, GPT-3.5, Claude 3, Gemini Pro and several open models against 100,000 human participants. The paper appeared in Scientific Reports on 21 January 2026. The main test was the Divergent Association Task, where you list ten words as unrelated to each other as possible, scored by semantic distance. They added creative writing: haiku, story synopses, flash fiction.
The headline finding, in the authors' own words: models "can surpass average human performance on the DAT, and approach human creative writing abilities, yet they remain below the mean creativity scores observed among the more creative segment of human participants."
The part that got left out of the coverage is more precise. "The most creative humans, those in the top decile, quartile and above median, still achieve higher DAT scores than any model." And: "even the top performing LLMs are still largely surpassed by the aggregated top half of human participants."
An earlier, much smaller study points the same way. In February 2024, also in Scientific Reports, Kent Hubert, Kim Awa and Darya Zabelina matched 151 humans against 151 GPT-4 responses and found GPT-4 ahead on DAT semantic distance (84.56 against 76.95, t(300) = 13.65, p < .001). Different design, two years older, so do not blend the numbers. A 2023 paper in the same journal found the reverse ranking, which tells you how fast the ground moves.
The homogenisation problem, from a philosopher
Dorothea Winter researches AI and aesthetics at the Humanistische Hochschule Berlin. On stage at Freelance Unlocked 2026 she walked the audience through Margaret Boden's three levels of creativity from the 1990s: combinatorial, recombining known elements, which she puts at roughly 60 to 80% of all creative output; exploratory, pushing an existing system to its edges, roughly 20 to 35%; and transformational, changing the system or writing new rules, which she estimates at 1 to 5%.
She then described a study in which people wrote fictional texts with and without AI help. Those who had produced less original work before got better results with AI. But the results also converged, becoming more similar to one another and less individual. For the people who were already at the top, AI made almost no difference.
"My thesis is that AI homogenises the creative middle. If everyone in the middle uses AI, more similar output appears, and the differentiating force that actually defines the creative economy is lost. I think that is a real danger." (Dorothea Winter, translated)
That lines up exactly with the Montréal numbers. The models raise the floor and leave the ceiling alone.
Her second argument is about what happens to you over years, not to the market. She cited a study from medicine: after routine exposure to AI assistance, endoscopists doing colonoscopies without AI detected fewer adenomas than before. Checking the source, the Lancet Gastroenterology & Hepatology published it on 12 August 2025: the detection rate in non-AI procedures fell from 28.4% to 22.4% across 1,443 colonoscopies at four Polish centres, with 19 endoscopists who had each done at least 2,000 procedures. Winter's extrapolation is that outsourcing levels one and two of Boden's ladder for long enough costs you level three, because high-level creative work is trained by doing a lot of small creative work.
Her practical test is a question to ask yourself two or three times a week: is this tool enhancing me here, or am I just being lazy? And which parts do I keep doing by hand even when it is annoying, slow and less efficient in the moment?
Victoria Ringleb, managing director of the Alliance of German Designers, put the market consequence in one line:
"AI gives me kitsch. That is mediocrity. What art actually does, making me stumble over it, making me think, AI cannot deliver that." (Victoria Ringleb, translated)
Her worry is a visual monoculture, like only ever seeing grey sparrows at the bird feeder. She described a TV commercial captioned "created with generative AI" alongside a human name credited with the idea, which she read as the advertiser quietly conceding that the machine part was the weak part.
Meanwhile, clients keep buying human creative work
Two marketplaces published demand data in 2026. Both are vendor data about their own platforms, not the market.
Upwork's In-Demand Skills 2026, published 4 February 2026 and based on freelancer earnings on its own marketplace across calendar 2025, lists ten design and creative skills in most demand: graphic design, video editing, presentation design, video production, image editing, product and industrial design, 3D animation, logo design, illustration, brand identity design. Fastest growing: AI video generation and editing at +329%, AI image generation and editing at +95%, and logo design at +44%. The sentence that matters most for pricing: "nearly half of business leaders say they would pay a premium to work with independent talent who are creative and innovative."
Fiverr's Business Trends Index, published 3 August 2026, tracks search demand on its platform, comparing May to October 2025 with November 2025 to April 2026. Translation up 55%, formatting up 71%, book editing up 28%, essay writing up 17%. Video and animation up 278% as a category, with YouTube thumbnails up 52%. Logo design up 76%, brand identity up 26%. It is search data, so it measures intent to buy rather than money spent, and Fiverr is not a neutral observer of Fiverr.
Read together with the research, these are not contradictory findings. AI passes the average on a standardised test, and clients keep paying humans for work where the average is not good enough.
Copyright: still a human requirement
If your differentiator is human authorship, the legal position is currently on your side and is moving. The European Parliament's research service published a briefing on AI-generated works in December 2025. EU copyright requires a work to be the "author's own intellectual creation" resulting from "free and creative choices", with a "personal touch". Purely machine-generated output with no or very little human intervention is unlikely to qualify, and an AI system cannot hold copyright. A Czech municipal court ruled in 2023 that writing a prompt does not by itself make you an author. Italy passed a law in October 2025 confirming that AI-assisted works stay protected where they reflect genuine human intellectual work.
Pending, not settled: an own-initiative draft report by MEP Axel Voss insisting AI-generated content stay ineligible was expected for a plenary vote in early 2026, and the Copyright Directive review cannot begin before June 2026. The contract detail is in AI and copyright for creative freelancers.
The positioning move
Marco Janck has been self-employed for 25 years, built an agency to 20 people and deliberately shrank it back to four. His session was about brand building, and his answer to the homogenisation problem was the least polished thing said all day:
"Everything out there is only being built by AI now. So the art is to push back against it and say: no, I still make my own typos. And if that is the difference in future, that I have to make typos, then fine, I make typos." (Marco Janck, translated)
His underlying rule is that being noticeable requires leaving the norm, and that this has a price you can charge for. He told a story about being fully booked, wanting to stay in a bidding process without winning it, and sending a quote far above his own internal scale: a €50,000 retainer. The signed order came back ten minutes later.
"The recipient, the client, sets the value. Not me with my strange internal grid." (Marco Janck, translated)
That is the practical bridge between the research and your invoice. The top decile is not a personality. It is a position you occupy for a specific client on a specific problem, and it is expressed in the choices only you would have made.
What to do on Monday
- Sort last month's billable work into Boden's three levels. If more than 80% was recombination, that is the part clients will eventually buy cheaper.
- Pick one part of your craft you will keep doing by hand, on purpose, and write down why. Winter's test: annoying and slower now, better judgment later.
- Look at your last five deliverables and mark the choices a model would not have made. If you cannot find any, that is the finding.
- Add one thing to your portfolio page that is unmistakably yours: the process, the wrong turns, the reason for the decision. Human authorship is also the copyright argument.
- Reprice one offer as the outcome for the client rather than the hours it takes you, and see what comes back.
For what clients are actually ordering right now, the 2026 Fiverr and Malt data has the category detail, and the AI premium or AI discount question covers what happens to rates on both tracks. If you want the work that pays for judgment, a free 9am profile puts your portfolio in front of companies across DACH.
Freelance Unlocked is co-organized by 9am together with Uplink and freelancermap. This article draws on the sessions of Victoria Ringleb and Dorothea Winter and of Marco Janck at Freelance Unlocked 2026. Watch the full talks above, and join us at the next edition: freelanceunlocked.com.
