Future Horizons

AI and Creativity: The Machine That Makes Art, Music, and Literature

The argument about whether machines can be creative is being settled by institutions that never ask it: a cert denial in March 2026, record labels signing licences instead of suing, and a jury trial on image-model training set for April 202

Gabriele Masetti · 2026-09-18

The question courts are answering instead of aesthetes

Every few months since 2022, someone has asked whether a machine can really be "creative," and every few months the discourse settles into the same stalemate: skeptics point to statistical pattern-matching, boosters point to surprising outputs, and the conversation ends where it started. That debate is not going to be resolved by essays.

It is being resolved, piece by piece, in copyright offices, federal courtrooms, and union contracts — institutions that don't care whether a diffusion model has an inner life, only who gets to own, sell, and be paid for what it produces. The aesthetic question is stalling out precisely because the legal and economic one is moving fast, and it's worth being blunt about why: generative AI is a spectacularly capable recombination engine operating inside existing conceptual spaces, and the entities deciding its cultural status have quietly agreed to stop asking it to be anything more than that.

Cognitive scientist Margaret Boden's taxonomy, laid out across decades of work culminating in her 2010 book Creativity and Art: Three Roads to Surprise, gives the vocabulary this argument needs. Boden distinguishes three kinds of creativity: combinational (unfamiliar combinations of familiar ideas), exploratory (generating novel ideas by working the boundaries of an accepted style or "conceptual space"), and transformational (altering the rules of the conceptual space itself, so that ideas previously impossible become thinkable).

The first two are what today's generative models do, at a scale and speed no human matches — they blend styles, remix genres, and explore the edges of the spaces defined by their training data. The third — the kind of rupture that gave the world Cubism or bebop, a genuine change in the rules of the game rather than a fluent performance within them — has no documented instance from a generative model, and nothing in current architectures suggests it's a matter of scale.

That gap is real, but it's also, for the people whose livelihoods and rights are actually at stake, beside the point. Nobody is being sued over transformational creativity. They're being sued, and suing, over combination.

Jason Allen won a ribbon and lost a copyright

The clearest illustration of that pivot is Jason M. Allen's Théâtre D'opéra Spatial, a Midjourney-generated image that won first place in the "digital arts / digitally manipulated photography" category at the 2022 Colorado State Fair fine arts competition. The win triggered a genuine aesthetic controversy — other entrants and commentators called it a form of cheating, a debate about whether prompting counts as artistry.

That argument went nowhere useful, as these arguments do. What actually mattered came a year later: in September 2023, the U.S. Copyright Office's Review Board issued a final refusal of Allen's application to register the image, on the grounds that it contained more than a "de minimis" amount of AI-generated material, which under the Office's authorship doctrine cannot be claimed by a human applicant unless disclaimed.

Allen wasn't refused because the Board judged the image derivative or unoriginal — it explicitly wasn't rendering an aesthetic verdict at all. It was applying a threshold rule about who, or what, counts as an "author" under the Copyright Act, a rule with the same lineage as the Office's earlier position on Stephen Thaler's AI-authored works. That lineage is now settled at the top: on 2 March 2026 the Supreme Court denied certiorari in Thaler v. Perlmutter, leaving no copyright without human authorship as final US law.

Allen filed an appeal in federal court in Colorado in September 2024, and it is still unresolved. He moved for summary judgment on 25 August 2025 and filed his reply to the government on 16 January 2026; the district court has yet to rule.

But notice what the dispute is actually about: not whether the machine was creative, but whether creativity performed by a machine can attach to a human legal person at all. That's a doctrinal question, and it's the one with teeth.

Style is the raw material, and it isn't fenced off

If Allen's case shows the law drawing a line around authorship, Greg Rutkowski's shows it declining, so far, to draw one around style. Rutkowski is a Polish digital fantasy painter whose work has illustrated Dungeons & Dragons, Magic: The Gathering, and Sony's Horizon Forbidden West.

When Stable Diffusion launched in 2022, his name became one of the most common prompt terms on the platform — MIT Technology Review reported in September 2022 that "Rutkowski" had been used as a generation prompt roughly 93,000 times, more than Vincent van Gogh, Pablo Picasso, or Leonardo da Vinci combined with each other.

Rutkowski told the magazine he worried the flood of imitative output would make his own work impossible to find and threaten his career; Stability AI responded by November 2022 by limiting the ability to invoke living artists' names as prompts. But the underlying legal fact didn't change: style, as such, is not protected subject matter under U.S. copyright law, only particular fixed expressions are.

A model trained to recognize and reproduce "the Rutkowski look" on demand is doing textbook combinational creativity — merging a named human signature with arbitrary prompt content — and doing it in a zone the statute wasn't built to police. Rutkowski's response was to become one of ten named plaintiffs, alongside Sarah Andersen and Kelly McKernan, in the 2023 class-action suit *Andersen v.

Stability AI*, which reframes the style question as a training-data question instead: not "can you imitate my style" but "did you copy my images to build the machine that imitates it." In August 2024, U.S. District Judge William Orrick allowed the artists' core copyright claims against Stability AI, Midjourney, DeviantArt, and Runway AI to proceed — a procedural win, not a verdict, but a signal that the claim was being taken seriously as a property dispute rather than dismissed as an aesthetic grievance. The verdict has not been taken. Andersen is the first US case set to put image-model training to a jury, and the trial, listed for 8 September 2026, was continued to 5 April 2027.

The infrastructure cases don't ask if it's art

Scale up from individual artists to institutional plaintiffs and the pattern holds even more starkly: the biggest AI-and-creativity lawsuits now pending are almost entirely indifferent to the quality or originality of AI output. The New York Times sued OpenAI and Microsoft in December 2023, alleging that ChatGPT was trained on millions of Times articles and can reproduce them near-verbatim; in March 2025, Judge Sidney Stein declined to dismiss the core copyright claims, and through mid-2026 the case remained in contested discovery, with the Times and other papers asking the court in July 2026 to sanction OpenAI over how it searched its own systems for evidence.

Getty Images' suit against Stability AI took the argument furthest and hit the clearest wall: in a landmark ruling on November 4, 2025, England's High Court rejected Getty's central claim that the Stable Diffusion model itself is an "infringing copy" of Getty's photographs, holding that a model's trained weights are not a reproduction in the sense the UK's Copyright, Designs and Patents Act requires — they're statistical parameters, not stored images.

Getty had already abandoned its claim that the underlying training occurred in the UK, so the court explicitly declined to rule on whether training itself infringes; Getty won only a narrow trademark finding, over AI-generated outputs that carried a recognizable Getty watermark. The ruling stands, but not finally: at a consequentials hearing in December 2025, Mrs Justice Joanna Smith granted Getty permission to appeal the secondary-infringement point, calling it a "pure question of law" on which "the minds of reasonable lawyers may differ." Read together, the Times case and the Getty case are fights over reproduction, derivation, and jurisdiction — lawyerly categories that route entirely around the question of whether what comes out the other end is any good, or new, or "real" writing and imagery.

The creativity debate isn't being won or lost in these courtrooms. It's being ignored, because the statutes being litigated were never built to ask it.

Music makes the same fight audible, plus a second one: your voice

The music industry compresses this into an even sharper split. In June 2024, the RIAA filed twin lawsuits on behalf of Universal Music Group, Sony Music, and Warner Music — one against Suno in Massachusetts federal court, one against Udio's parent Uncharted Labs in the Southern District of New York — alleging that both AI music generators were trained on copyrighted recordings without a license, pointing to outputs the labels said echoed songs like Green Day's "American Idiot."

Suno's defense, laid out publicly in August 2024, was fair use: the same doctrinal argument OpenAI and Stability are making about text and image training. Most of the plaintiffs stopped waiting for a ruling on it. Universal settled with Udio in October 2025, Udio committing to a platform trained only on authorised and licensed music; Warner settled with Suno in November 2025, under a deal its artists and songwriters can opt into and be paid for; BMG licensed its recorded-music and publishing catalogues to Suno in August 2026, settling past use in the same document. Universal and Sony are still suing Suno. The fair-use question survives, but in music it is being priced rather than litigated.

The second fight is different in kind, and it's the one that cuts closest to the creativity question without actually being about it: identity. In April 2023, a TikTok user going by Ghostwriter977 released "Heart on My Sleeve," a song they wrote and produced themselves, with AI-generated vocals built to sound like Drake and the Weeknd. It racked up millions of plays before Universal Music Group filed takedown notices on April 17, 2023, pulling it from streaming platforms.

Recording Academy CEO Harvey Mason Jr. initially told the New York Times the track was "absolutely eligible" for Grammy consideration because a human had written it — then reversed himself days later, ruling it ineligible on the grounds that the vocals weren't legally cleared and the song wasn't commercially available. That flip-flop is instructive: nobody disputed that a human wrote the song, meaning the human-authorship test that sank Jason Allen's copyright claim was satisfied here.

The problem was never authorship or creativity — it was that the song borrowed two real people's voices without consent, a right-of-publicity and likeness problem, not a copyright-of-expression problem. It's the same distinction the Copyright Office relies on for Allen and that the Getty court relies on for training data, applied to the one piece of a person that a diffusion model can now convincingly counterfeit: how they sound.

Hollywood's unions got there before the courts did. The SAG-AFTRA strike that ran through much of 2023 was resolved in November of that year and ratified in December, producing a contract that requires studios to obtain explicit, informed, compensated consent before creating an "Employment-Based" or "Independently Created" digital replica of a performer's voice or likeness, and gives the union the right to bargain before any "synthetic performer" is used in place of a human role.

That agreement doesn't settle whether an AI-voiced performance is creative any more than the RIAA suits or Getty's loss do. It settles who has to be paid and asked, which turns out to be the only question labor actually needed answered.

What actually gets decided this decade

Put the cases side by side and a shape emerges that no single one of them states outright. The Copyright Office needs an answer to "who is the author," so it built a threshold test around the proportion of human input, and Allen's appeal is testing that test. English and American courts need an answer to "was something copied," so Getty's case turned on whether trained weights count as a copy at all, and the Times case turns on whether outputs reproduce protected expression.

The RIAA needs an answer to "was the input licensed," which is a training-data question wearing the same clothes as Getty's. SAG-AFTRA needed an answer to "who consents and gets paid when a likeness is used," and negotiated it directly rather than waiting on any of the above. None of these bodies asked Boden's question — combinational, exploratory, or transformational — because none of them are built to.

Two of those questions closed while the essay was being written. The Supreme Court shut the authorship one in March 2026 by declining to hear Thaler. The record industry shut its own version of the training-data one by signing licences. The third is set to go to a jury in San Francisco in April 2027: twelve people, none of them asked what creativity is.

They're building a settlement instead: a set of rules about authorship thresholds, training-data licensing, and consent-for-likeness that will govern how AI-assisted creative work gets made, credited, and monetized for years, independent of whether anyone ever agrees on what "creative" means. The philosophical question about machine creativity was never actually undecidable — Boden's own framework answers it reasonably well, model by model. It's just turned out to be the wrong question to bring to the rooms where the outcomes that matter are actually being decided.