Creative Industries and Art

Karla Ortiz builds worlds for a living. A concept artist whose brushwork has shaped Marvel films and video games, she spent two decades learning to conjure a creature or a city that had never existed before. Then, in late 2022, she typed her own name into an image generator and watched it produce pictures "in the style of Karla Ortiz" — competent, uncanny approximations of a craft she had spent her life building, generated in seconds, by a model that had ingested her portfolio without asking. She had never licensed her work for that purpose. She had never been paid. She had never been told.

In January 2023, Ortiz became a lead plaintiff in Andersen v. Stability AI, one of the first major lawsuits to argue that training an image model on billions of scraped pictures is not innovation but infringement at industrial scale. She has since testified before the U.S. Senate. Her case is still unresolved. In the years it has spent winding through the courts, the technology it concerns has improved by several generations, and the market she worked in has continued to contract around her.

That gap — between how fast the machines move and how slowly the law responds — is the story of this chapter.

The lawsuit wave

By early 2026, the creative-industries fight had become the largest wave of copyright litigation in the history of the medium. Depending on how consolidated cases are grouped, trackers counted somewhere between roughly forty and sixty active suits in the United States alone, with more filed in the United Kingdom, Germany, Canada, and elsewhere. The defendants are the biggest names in the field: OpenAI, Meta, Google, Anthropic, Stability AI, Midjourney, Suno, and Udio. The plaintiffs range from individual illustrators and novelists to The New York Times, Getty Images, Universal Music Group, and — as of mid-2025 — the Walt Disney Company.

The legal questions cluster into two. First, the input question: does copying millions of copyrighted works to train a model infringe, or is training a "transformative" fair use? Second, the output question: when a model reproduces a passage of text, a recognizable melody, or a named artist's style, is that output an infringing derivative of the works it learned from?

Through 2025, courts began to answer — and the answers were split, sometimes within a single ruling. The most consequential came in June 2025 from Judge William Alsup in Bartz v. Anthropic. Alsup held that training a large language model on books was "exceedingly transformative" and, in that respect, fair use — a genuine victory for the AI industry. But he drew a hard line at how the books were obtained. Anthropic had downloaded millions of titles from pirate "shadow libraries," and building a permanent library from stolen copies, Alsup ruled, was not fair use and would go to trial. The exposure was existential: statutory damages for willful infringement run up to $150,000 per work, across millions of works. Rather than gamble, Anthropic settled in September 2025 for $1.5 billion — roughly $3,000 for each of about 500,000 works, the largest copyright recovery on record.

Days later, Judge Vince Chhabria reached the opposite bottom line in Kadrey v. Meta, ruling for Meta — but on strikingly narrow grounds. The authors had lost, he wrote, not because Meta was in the right but because they had failed to prove market harm. He went out of his way to flag the theory he thought could win: not that a model memorizes and regurgitates a specific book, but that a machine trained on an author's work can flood the market with cheap substitutes and dilute the value of the human original. That "market dilution" argument is now the frontier of the next round of cases. Earlier, in February 2025, Judge Stephanos Bibas had rejected a fair-use defense outright in Thomson Reuters v. Ross Intelligence, though that involved a non-generative legal-research tool trained on Westlaw headnotes.

The picture that emerges is not clarity but a mosaic. Training itself may often be fair use; using stolen copies to do it is not; and whether the outputs infringe depends on facts a jury has yet to weigh in most cases. Meanwhile the scraping continues, because nothing in a pending lawsuit stops it.

What the Copyright Office actually said

Into this uncertainty stepped the U.S. Copyright Office, which spent 2024 and 2025 producing a three-part report on copyright and artificial intelligence. Part 2, released in January 2025, addressed authorship, and its conclusion was blunt: material generated by a machine without meaningful human control is not copyrightable, and typing a prompt — however elaborate — does not make the resulting image or song your creative work in the eyes of the law. That position was reinforced by the courts the same year, when the D.C. Circuit affirmed that an AI system cannot be listed as an author.

Part 3, released in a pre-publication version in May 2025, took on the harder question of training and fair use — and this is where the Office parted ways with the industry's core defense. It accepted that some AI training is genuinely transformative. But it drew a line precisely where the creators had been pointing: when a model is trained on copyrighted works in order to generate expressive content that competes with those works in the same market, the Office concluded, the use "goes beyond established fair use." Feeding a music model a catalog of songs so it can produce songs that displace them is not, in the Office's reading, the same as a person learning by listening.

The report's release was overshadowed by an extraordinary sequel: within days, the administration removed the Register of Copyrights, Shira Perlmutter, who had overseen it — a firing she challenged in court and that critics read, fairly or not, as retaliation on behalf of an industry the report had just ruled against. The episode underscored how political the question has become. A Copyright Office report is not binding law; it is guidance courts may weigh. But it framed the disagreement with unusual precision, and it framed it in the creators' favor.

The core legal dispute, then, comes down to a single analogy and whether it holds. AI companies argue that a model learning statistical patterns from millions of works is doing what every human artist does — studying predecessors and internalizing technique — and that the law has always permitted learning. Creators and the Copyright Office answer that a human who studies Hemingway produces, at most, one imperfect Hemingway imitator, while a machine that ingests Hemingway can produce infinite Hemingway substitutes on demand, at zero cost, aimed at the exact market Hemingway's estate depends on. Learning that competes at industrial scale, they argue, is a different thing than learning, and pretending otherwise smuggles the whole question past the point where copyright is supposed to bite.

The compensation crisis

Underneath the legal fight sits an economic logic that explains why the fight is so bitter. Once a work has been absorbed into a model, the marginal cost of generating another image, song, or paragraph in its manner is essentially nothing — a few cents of electricity. A business that can manufacture unlimited substitutes for free has no market incentive to pay the humans whose work made the substitutes possible. Payment, in this structure, is not a cost the technology forces; it is a cost the technology removes. That is the crux of question after question in this domain: the near-zero marginal cost of synthetic content severs the link between the value a creator's work generates and any return flowing back to the creator.

Some money has begun to move, but unevenly and on terms that favor the largest rights-holders. Major music labels and news organizations have negotiated licensing deals; a handful of AI firms have signed content agreements worth tens of millions. Indie-focused organizations such as Merlin and Kobalt have at least given their artists and writers the right to opt in or out. But two problems recur. First, even where licensing happens, the money typically arrives as a one-time lump sum rather than an ongoing royalty. A session musician whose recordings once paid out every time a track was streamed, synced, or performed now receives, at best, a single check for the right to be trained upon — an ongoing revenue stream converted into a severance payment. Second, and more fundamentally, the vast majority of working creators have no label, no publisher, and no seat at the negotiating table. Their work was scraped from the open web and absorbed into commercial systems generating real value, with no mechanism for consent, credit, or compensation to reach them at all. For them, the "compensation crisis" is not a bad deal; it is the absence of any deal.

The obvious remedy — route consent and payment back to creators at scale — runs into the fact that the systems were built precisely to avoid that. Proposals exist: collective licensing bodies that treat AI training the way radio royalties are handled, statutory training levies distributed to registered creators, opt-out registries. But each requires the industry to accept, as a baseline, that training data is something owed for rather than taken. The $1.5 billion Anthropic paid establishes that pirated data can carry a price. Whether lawfully scraped data ever will is the open question.

The displacement wave

The creators most exposed are not, at first, the famous ones. They are the mid-market professionals who did the steady commercial work that kept the lights on: freelance illustrators, book-cover and album-cover designers, concept artists, stock-photo shooters, copywriters, translators, and the composers who supplied library and background music. This is the layer where "good enough for the price" is the whole business, and where a $30 monthly subscription now competes directly with a $500 commission. Session musicians and jingle writers face tools like Suno and Udio that generate finished, produced tracks — melody, lyrics, instrumentation, vocals — faster than a person can listen to them. Voice actors face cloning systems that reproduce a performer's voice from seconds of audio.

Quantifying the damage precisely is genuinely hard, and honesty requires admitting it. A widely cited 2024 study commissioned by CISAC, the international confederation of authors' societies, projected that music creators stood to lose roughly a quarter of their income and audiovisual creators around a fifth by 2028, even as the generative-AI content market swelled toward tens of billions of euros — a transfer, in effect, from the people who made the training data to the companies that trained on it. But projections are not measurements, and here the epistemic problem is acute: creative markets were already being hollowed out before generative AI arrived, by platform economics that pushed prices toward zero, by content commoditization, and by industry consolidation. Freelance illustration rates were under pressure a decade ago. Separating the AI signal from those long-running trends is difficult, because they compound one another — cheap AI supply accelerates a commoditization that platforms had already begun. The most defensible claim is not that AI single-handedly caused the contraction, but that it sharply accelerated an existing one and removed the floor beneath it. Where earlier disruptions redistributed creative work, this one threatens to remove the market for the human input entirely.

The feedback loop

Beyond individual livelihoods lies a subtler worry: that a culture which increasingly trains its machines on itself might slowly narrow the range of what gets made. The mechanism is specific. A model learns from a vast corpus of human work and, by design, converges toward the statistical center of that corpus — the median of everything it has seen. Its outputs are, on average, more conventional than the tail of human eccentricity that trained it. Audiences consume those outputs, acculturate to them, and adjust their expectations. Human creators, competing for attention, drift toward what now reads as "normal." And the next model trains partly on that — on a corpus already tilted toward the machine's own preferences.

graph LR
  A[Human creative corpus] --> B[Model trained toward the median]
  B --> C[Synthetic output: conventional, optimized]
  C --> D[Audiences acculturate, taste narrows]
  D --> E[Human creators mimic AI aesthetics]
  E --> A
  C --> A

Is this speculation, or is it observable? Some of it is now measured. In a 2024 Science Advances study, Anil Doshi and Oliver Hauser found that giving writers access to generative AI made each individual story more creative on average — but made the pool of stories collectively more similar to one another. Individual enhancement, collective homogenization: exactly the shape the feedback-loop hypothesis predicts. On the technical side, a 2024 Nature paper by Ilia Shumailov and colleagues demonstrated "model collapse" — when models are trained recursively on their own synthetic output, the tails of the distribution vanish and quality degrades toward a bland mean. That is a fact about training data, not proof about human taste, and the two should not be conflated. But together they move the cultural-convergence worry from pure speculation toward a hypothesis with early empirical support. What remains genuinely unknown is how strong the effect is at the scale of a whole culture, and whether human creators' appetite for novelty is enough to keep breaking the loop.

Authorship, value, and the audience

If the loop tightens, a philosophical puzzle becomes an economic one. Who authored an image assembled from a million uncredited inputs? Under current U.S. law, often no one — a purely machine-generated work has no copyright, which means it can be freely copied by anyone, including the person who prompted it. That is a strange new category: cultural output that is expensive to want and free to own, with no author to credit and no rights to defend. It scrambles the entire apparatus by which culture has been made, attributed, and paid for.

Whether audiences ultimately care is the pivotal unknown, and it is harder to measure than it looks. Some studies find that people rate identical works lower once told a machine made them — a bias against AI authorship. Others find that in genuinely blind tests, listeners and viewers cannot reliably tell the difference and rate the two similarly. Both can be true at once, and that is the measurement trap: a stated preference for human work ("I'd never listen to AI music") can coexist with revealed behavior that ignores authorship entirely (streaming an AI track because it surfaced on a playlist). To know whether audiences truly don't care, you would need to watch what they choose when authorship is disclosed, when it is hidden, and when disclosure costs them nothing — and those experiments are only beginning. The honest position is that we do not yet know whether human authorship is a preference audiences hold and simply aren't expressing through the market, or one they are quietly abandoning.

Depending on the answer, one of two market structures for human creativity emerges. If authorship matters, human work survives as a premium niche — the vinyl record, the handmade table, the live performance sold partly on its provenance. If it does not, human creators may instead form a parallel economy built on relationship and patronage rather than product: direct audience support, membership, presence, the things a model cannot fake because they are not about the artifact at all. Most likely both appear at once, and the mass middle — the commercial commodity tier — is what disappears.

The pipeline problem

The longest shadow falls on people not yet in the field. Creative careers have always been economically precarious, but they were possible: enough illustrators, musicians, and writers could earn a living to justify the training, the apprenticeships, the years of unpaid practice. Remove the economic base, and you do not merely impoverish today's professionals — you change who can afford to become one tomorrow. Creative fields would increasingly belong to those who can practice without income: the independently wealthy, the subsidized, the hobbyist. The voices that enter narrow toward the voices that can afford to. A culture that prices out everyone who needs to be paid does not stop producing art, but it produces art from a shrinking and less representative slice of humanity — which, in a domain whose entire value is the diversity of human perspective, is a quiet catastrophe.

The pace problem — and the window

None of this is the first time a machine has upended a creative trade, and the history is instructive about timing. The comparison that matters is not whether disruption happened before — it always has — but how long the gap ran between the disruption and the legal response that resolved it.

Two patterns stand out. Historically, resolution took years to decades — and generative AI's diffusion is faster than any of these, reaching hundreds of millions of users within months of launch. The disruption compressed; the legal machinery did not. Second, the settlements that actually protected creators were structural — a compulsory license, a trust fund — not one-off court wins. The player-piano fight, in particular, is the closest analogy: a machine that reproduced performances threatened to destroy the sheet-music market, and Congress answered not by banning the machine but by inventing a mechanism that made it pay automatically. That is the kind of instrument the current moment lacks.

Which is why the near-term legal forecast is, most plausibly, continued contestation rather than clarity. The rulings so far point in different directions; the pivotal output-infringement and market-dilution cases are years from final judgment; appeals and likely circuit splits could push a definitive answer toward the Supreme Court well into the 2030s. The Anthropic settlement priced pirated data but set no rule for lawfully scraped data. In the interval, the practical status quo holds: models keep training, creators keep losing ground, and the law keeps arriving after the fact. That lag is not a footnote to the story — for the people living inside it, it is the story.

What we owe, and what we want

The unresolved legal questions sit on top of unresolved moral and political ones. Do AI companies owe anything to the creators whose work formed the foundation of their models? The Copyright Office's reasoning implies yes, at least where outputs compete with inputs; the $1.5 billion settlement concedes it for stolen data. If the answer is yes in principle, the mechanism question follows: consent and compensation would need to flow back at a scale that individual negotiation cannot reach, which points toward collective solutions — training levies, statutory licenses administered by collecting societies, opt-out registries with teeth — modeled on the century-old machinery that made radio and mechanical reproduction pay.

Whether human authorship should get protection distinct from machine output is the harder normative question, and the defensible answer is probably modest but real. Outright bans on AI content are unworkable and arguably censorious. But three levers are defensible and mutually reinforcing: labeling, so that a market preference for human work can at least be expressed rather than hidden; preferential licensing and procurement, so that public institutions and platforms can choose to source human work where it matters; and public investment — the arts-funding and education base that historically underwrote culture that the market alone would never have paid for. None of these stops the technology. All of them widen the space in which humans can still afford to make things.

The deepest question is the one that is currently being answered by default. What kind of creative ecosystem does a society actually want — and is it willing to build it deliberately, through copyright rules, platform-design standards, and public investment, or will it accept whatever emerges from millions of individual cost-minimizing choices? At present the second path is winning, not because anyone chose it, but because no one chose the first. That is the real stake of the litigation, the reports, and the settlements: not merely who gets paid, but whether the culture of the coming decades is designed or merely defaulted into.

Summary

Generative AI has triggered the largest wave of copyright litigation in the history of creative media, and as of early 2026 the law is fractured rather than settled.

1. The litigation landscape is large and split. Dozens of active U.S. suits (roughly forty to sixty by common counts), plus cases in Europe and beyond, target every major AI firm. Courts have ruled both ways: training was deemed transformative fair use in Bartz v. Anthropic, but using pirated books was not — and Anthropic settled for a record $1.5 billion. Kadrey v. Meta went for Meta on narrow grounds while keeping a "market dilution" theory alive for future plaintiffs.

2. The Copyright Office sided partly with creators. Its 2025 report concluded that AI output which competes with the works it was trained on goes beyond fair use, and that purely machine-generated work has no human author and thus no copyright. The report's release was followed by the removal of the Register of Copyrights.

3. The core legal disagreement is one analogy. AI firms say training is learning, which the law permits; creators and the Office say learning that manufactures infinite market substitutes at zero cost is a categorically different act.

4. Near-zero marginal cost removes the incentive to pay. Where licensing exists it favors large rights-holders and typically replaces ongoing royalties with a one-time lump sum; most independent creators, scraped from the open web, receive nothing.

5. Displacement is real but hard to isolate. Mid-market illustrators, designers, session musicians, copywriters, and voice actors are most exposed; a CISAC study projected significant income losses by 2028. But platform economics and consolidation were already hollowing these markets, so AI is best described as accelerating an existing collapse, not solely causing it.

6. The feedback-loop worry has early evidence. A 2024 Science Advances study found AI made individual work more creative but collective output more homogeneous; a 2024 Nature study demonstrated technical "model collapse." Neither proves cultural monoculture, but both move the hypothesis beyond pure speculation.

7. Authorship is now murky and audience preference is unmeasured. Machine-only works often have no owner; whether audiences genuinely don't care about human authorship, or care but don't yet express it in the market, remains an open empirical question.

8. The pace gap is the crux. Historically, creative-tech disruptions took years to decades to resolve, and the durable fixes were structural (compulsory licenses, trust funds), not one-off court wins. AI diffused faster than any predecessor while the legal response lags — most plausibly, contestation will continue for years while scraping does too. The choice is whether the resulting creative ecosystem is designed deliberately or defaulted into.

Sources

(Canonical primary references, cited from knowledge and not re-fetched in this session — verify before publication.)

• U.S. Copyright Office — Copyright and Artificial Intelligence (report series)
• Bartz v. Anthropic — N.D. Cal. fair use ruling (June 2025)
• Kadrey v. Meta — N.D. Cal. ruling (June 2025)
• Thomson Reuters v. Ross Intelligence — D. Del. ruling (Feb 2025)
• Andersen v. Stability AI (Karla Ortiz et al.)
• Doshi & Hauser, "Generative AI enhances individual creativity but reduces the collective diversity of novel content," Science Advances (2024)
• Shumailov et al., "AI models collapse when trained on recursively generated data," Nature (2024)
• CISAC — Global economic study on generative AI and human creators (2024)
• The Verge — AI copyright lawsuits tracker

Last updated: 2026-08-02