Almost everything written about AI and creativity is an argument dressed as a finding. One side cites surveys showing creators thriving. The other cites artists losing their livelihoods. Both are quoting real numbers.
They are also, almost always, quoting numbers about different people.
We build AI systems for a living, which makes this an uncomfortable subject to write about honestly. It is also the reason to write about it. If you ship AI development services and you cannot describe the costs of the technology accurately, you are not qualified to describe the benefits either.
So here is the evidence as of September 2026 — the studies, the settlements, the court rulings, and the platform data — with the caveats the headlines leave out.
Working through what AI means for your product or your team? Talk to our team.
Two surveys, two realities, one missing footnote
In June 2026, Adobe published its Creators' Toolkit Report: 87% of creators using creative AI said it had accelerated the growth of their business or audience, and 75% described it as integrated or essential to how they work. Harris Poll surveyed more than 16,000 creators across eight countries.
In July 2026, Creative Boom surveyed roughly 403 illustrators. Sixty percent said AI had affected their work or income over the previous twelve months. A third of that group said the effect was significant. Only 12% said AI had not touched them at all.
These look like contradictory findings about one profession. They are consistent findings about two.
Read Adobe's methodology and the tension mostly dissolves. Adobe defined creators as people who publish digital content several times a month to build an audience and earn across digital platforms — "emerging and professional social-first creators rather than individuals employed full-time in traditional creative industry roles." That is a population whose core scarce asset is *distribution*: attention, consistency, volume. Generative AI lowers their production costs without touching what they sell.
The Creative Boom respondents sell something else: commissioned execution, priced per project or per day. Their scarce asset is exactly the thing a diffusion model reproduces cheaply. When your product is the image, a machine that makes images competes with you. When your product is an audience, it does not.
Both numbers are true. Neither generalizes. Any argument that quotes one without the other is not making an empirical claim.
There is a further wrinkle inside the Creative Boom data that is easy to miss. Among illustrators who said AI had significantly affected their work, 16% felt fairly paid. Among those who said AI had not affected them, 38% did. Only a fifth of the whole sample felt fairly paid at all. UK day rates averaged £350, but the median was £425 for those who felt fairly paid and £280 for those who did not. AI is not creating the pay gap in illustration. It is widening one that was already there.
The effect everyone gets backwards: AI lifts the floor
The best-controlled experiment on AI and creative output is Doshi and Hauser's 2024 study in *Science Advances*. Two hundred ninety-three writers produced short stories; some worked unaided, some could request one AI-generated story idea, some up to five. Six hundred independent evaluators rated the results.
Access to AI ideas raised ratings. With five ideas available, stories scored 8.1% higher on novelty and 9.0% higher on usefulness than the unaided control.
The interesting part is the distribution. The researchers pre-measured each writer's inherent creativity. For writers in the *lower* half, five AI ideas produced gains of 10.7% on novelty, 11.5% on usefulness, 26.6% on how well written the story was judged, and 22.6% on enjoyability. For writers in the upper half, the effect was close to nothing. They were already performing at that level.
This is the finding that survives replication across domains, and it is almost always reported backwards. Generative AI is not a creativity multiplier. It is a floor-raiser. It moves people who were not very good toward competent, and it does approximately nothing for people who were already good.
Which tells you exactly who it threatens, and it is not who the discourse assumes. A technology that makes mediocre work competent does not endanger the mediocre. It endangers whoever was being paid a premium for the gap between mediocre and good.
And it takes the most from the people at the top
That prediction has been tested directly, and it holds.
Xiang Hui and Oren Reshef of Washington University, with Luofeng Zhou of NYU, studied a large online freelance marketplace around the releases of DALL·E in April 2022 and Midjourney in July 2022. Image-related freelancers — designers, illustrators, image editors — saw monthly jobs fall 3.7% and monthly income fall 9.4%.
A 9.4% income drop is a bad year, not an extinction. The distribution is the story again. The researchers found that for every 1% increase in a freelancer's *past* earnings, that freelancer suffered an additional 0.5% decline in job opportunities and an additional 1.7% decline in monthly income.
The better you had been doing, the worse it hit you.
This inverts the standard automation narrative, in which technology displaces the least skilled and the talented adapt. Here the mechanism runs the other way. A top freelancer's rate was underwritten by a reliability premium — the client's confidence in getting something good without supervision. When a cheap tool makes acceptable output broadly available, that premium is the thing that evaporates. The floor rises to meet the ceiling, and the people standing on the ceiling paid for it.
Homogenization is not an aesthetic complaint. It's a measurable effect.
The same *Science Advances* study measured something beyond quality: how similar the stories were to each other.
AI-assisted stories were significantly more alike than unaided ones — a similarity increase of roughly 10.7% of the measured range in the one-idea condition and 8.9% in the five-idea condition. Stories also drifted about 5% closer to the AI's own suggested ideas, which the authors describe as anchoring.
The authors frame the result as a social dilemma, and the framing is precise. Each individual writer is better off using the tool. Collectively, the set of stories produced is narrower. Nobody defects; the diversity loss is an emergent property of everyone rationally accepting help from the same model.
For anyone building with this technology, that is not a philosophical observation. It is a product risk with a measurable signature. If your content pipeline, your recommendation copy, your generated designs, and your competitor's all pass through a handful of frontier models, convergence is the default outcome, and differentiation becomes something you must engineer against rather than something you get for free. In the AI agent and content systems we build, the guardrails that matter most are usually not about correctness. They are about not producing the same thing as everyone else.
The money moved. Very little of it reached artists.
2025 and 2026 were the years the AI industry started paying. Following where the money went is more instructive than any position paper.
In June 2025, Judge William Alsup ruled in *Bartz v. Anthropic* that training large language models on books was fair use — "exceedingly transformative," in his words. He also ruled that downloading pirated copies to build a permanent internal library was not fair use, even when the books were later used for training. Converting lawfully purchased print books to digital was fine. Acquiring them from pirate libraries was not.
Anthropic settled the surviving claims for $1.5 billion, and Judge Araceli Martínez-Olguín granted final approval on 20 July 2026 — around 500,000 works at roughly $3,000 each, the largest copyright settlement in US history.
Read that sequence carefully, because it is the single most misreported fact in this debate. The record-setting payout was not compensation for training AI on people's work. On that question, the authors *lost*. The payout was for how the files were obtained and warehoused. An AI company that had bought the same books legally and trained the same model would, under this ruling, have owed nothing.
Music followed a similar shape with a sharper ending. Warner Music settled with Suno in November 2025 and with Udio shortly after; Universal settled with Udio and signed on for a licensed platform. Suno agreed to retire its existing models and train only on licensed works going forward. Udio's deal reshaped the product into a walled garden where nothing users generate leaves the platform. Sony has not settled, and its fair-use claims remain live.
Then, on 5 June 2026, the American Federation of Musicians sued Universal and Warner — not the AI companies. The union alleges the labels licensed members' recordings into these AI deals without paying, crediting, or even disclosing which recordings were used, triggering the "new uses" clause of the collective bargaining agreement. An amended complaint followed on 24 July. The labels moved to dismiss, arguing the clause does not cover generative AI at all.
Set the pieces side by side. Rightsholders sued AI companies. AI companies paid rightsholders. Musicians then had to sue the rightsholders to see any of it. At no point in that chain did a mechanism exist to route money to the people who made the work.
The law has not settled this — and the first rulings went the other way
For visual artists specifically, the courtroom record so far is worse than the coverage suggests.
Getty Images v. Stability AI produced the UK's first substantive judgment on 4 November 2025. Getty abandoned its principal copyright claim mid-trial, unable to establish that training occurred in the UK. Justice Joanna Smith rejected the secondary infringement claim, holding that the Stable Diffusion model is not itself an "infringing copy" — a model's weights are not a container of the works it learned from. Getty won narrowly on trademark, because the model could emit images bearing Getty and iStock watermarks. Justice Smith herself called the findings "extremely limited in scope." Getty received permission to appeal the secondary infringement point on 16 December 2025.
*Andersen v. Stability AI*, the class action brought by Sarah Andersen, Karla Ortiz and Kelly McKernan and now naming Stability, Midjourney, DeviantArt and Runway, survived dismissal and moved to discovery. But an order entered 15 June 2026 pushed the jury trial to 20 September 2027, with class certification and dispositive motions due 2 June 2027.
Filed January 2023. Merits decided, at the earliest, late 2027. Nearly five years — during which the models at issue were superseded several times over. Whatever the verdict, it will land on a technical landscape that no longer exists.
If you are an artist waiting for the law to resolve this, the honest read is that it will not resolve in time to matter for the current generation of models.
What happens when supply becomes free
The economics show up fastest in music, because distribution there is frictionless.
Deezer began detecting fully AI-generated uploads in January 2025. In January 2026, such tracks were 39% of daily uploads, about 60,000 a day. By April, roughly 75,000 a day and 44%. In June 2026, AI-generated tracks passed 50% of daily uploads for the first time, at nearly 90,000 per day. Deezer began tagging them for listeners that same month.
Spotify removed more than 75 million spam tracks over roughly twelve months and introduced an impersonation policy prohibiting unauthorized voice clones.
The fraud figures explain the volume. Deezer has reported that a large majority of streams on fully AI-generated tracks show signs of being fraudulent. Most of this material is not competing for listeners. It is competing for royalty pool disbursements — content as a mechanism for extracting fractions of a cent at scale.
This is the part of the story that has least to do with creativity. When production cost approaches zero, the binding constraint moves from making things to filtering them. Artists are not primarily losing a quality contest to AI. They are being buried in an index, and the platforms' response — detection, tagging, demonetization thresholds — is an admission that discovery, not generation, is now the scarce good.
What still holds value
Assemble the evidence and a consistent pattern appears. The things holding their value are the ones a model cannot produce from a prompt.
Authorship of the brief, not execution of it. AI collapsed the cost of rendering. It did not touch the judgment of what is worth rendering. The illustrators reporting the least damage are consistently those selling art direction and conceptual work rather than finished assets.
Verifiable provenance. Every settlement in 2025–26 turned on data lineage — where files came from, whether acquisition was lawful, what was in the training set. Provenance moved from a compliance checkbox to the thing that determines liability.
Relationships and rights, not files. Adobe's thriving 87% are people with audiences. That is not incidental. Distribution and trust are the assets that generative abundance does not deflate.
Deliberate difference. If homogenization is a measurable effect of shared models, then work that visibly is not model-shaped acquires scarcity value for exactly that reason.
What this means if you build with AI
We are on the other side of this ledger, and the honest position is not that AI is fine for artists. The evidence says it is not fine for a specific and identifiable group: people whose income comes from executing commissioned visual work at a professional standard. That group is measurably worse off, and the people who were best at it are worst off.
What follows for anyone building AI products is practical rather than moral.
Treat training-data provenance as an engineering requirement. The Anthropic ruling drew the line at acquisition, not use. That is a line you can engineer to: know what is in your data, know how it was obtained, keep the receipts. It is far cheaper than discovery.
Assume licensing is the direction of travel. Suno agreed to retire its models and train only on licensed material. Whatever the courts eventually decide, the commercial settlement is arriving first, and products built on unlicensed corpora carry a repricing risk that products built on licensed ones do not.
Design against convergence. If your differentiation depends on output from the same models your competitors use, you have no differentiation. This is an architecture problem — proprietary data, real evaluation criteria, human judgment at the points that matter — and it belongs in the design phase, not the polish phase.
Pay people for the parts models are bad at. Concept, direction, taste, and the decision about what is worth making at all. These are not sentimental categories. They are the categories the research says machines have not moved.
A disclosure, since it would be hypocritical to omit it: this blog runs on an automated pipeline we built, and generative AI is in it. This particular article was written by a person. The illustration at the top was composed algorithmically from a kit of vector elements our designer drew by hand — assembled by software, but every shape in it made by a human being who was paid for it. That arrangement is not a compromise we settled for. It is roughly what the evidence says a defensible one looks like.
The honest summary
Generative AI raises the floor of creative output and does close to nothing for the ceiling. It reduces the collective diversity of what gets made, measurably. It has cost professional visual freelancers real income, with the heaviest losses falling on the highest earners. The largest copyright settlement in history was won on a piracy technicality, not on the principle that training requires consent — and on that principle, so far, rightsholders have mostly lost. Money is now flowing, but the routes to individual creators barely exist, which is why musicians are suing their own labels. And in the one market where distribution is free, machine-generated work now exceeds half of daily supply.
None of that means the technology should not be built. It does mean that anyone building it who claims artists are simply fine has not read the numbers, and anyone claiming AI has ended creativity has not read them either.
The defensible position is narrower and less satisfying than either: a tool that helps most those who need it most, harms most those who were best, homogenizes what everyone produces, and has so far routed almost none of its returns to the people whose work made it possible. The first three are properties of the technology. The last one is a choice, and it is still open.
The evidence on AI and creativity is neither the catastrophe nor the liberation it gets sold as. Generative models raise the floor of creative output and barely move the ceiling. They measurably narrow the range of what gets produced. They have taken real income from professional visual freelancers, and taken the most from the ones who were doing best. Meanwhile the largest copyright settlement in history turned on how files were acquired, not on whether training requires consent — and on that second question, so far, artists have mostly lost in court.
What follows is not a moral posture but an engineering one. Know the provenance of your training data, because that is where the liability actually landed. Assume licensed corpora are the direction of travel, because the commercial settlements are arriving faster than the rulings. Design deliberately against convergence, because shared models produce shared output and that is a differentiation problem, not a philosophical one. And pay people for concept, direction and judgment — the parts the research says the models have not moved.
Building an AI product and want these questions handled in the architecture rather than in a compliance memo after launch? Talk to our team about your project.
“Generative AI is not a creativity multiplier. It is a floor-raiser — which tells you exactly who it threatens, and it is not who the discourse assumes.”
