The Creative Frankenstein

The Legal Battle between AI and Copyright: News and Developments

Newsletter article — updated. Originally published in our weekly newsletter and later revised with the 2025 developments: the Thomson Reuters v. Ross Intelligence, Bartz v. Anthropic, and Kadrey v. Meta rulings, and the Disney and Universal v. Midjourney suit.

Last updated: June 2026

AI and Copyright: 2025 Moved the Question, Not the Answer

2025 was supposed to bring clarity to the relationship between artificial intelligence and copyright. It did the opposite: it moved the goalposts. The first substantial rulings didn't settle whether training a model on protected work is lawful — they changed which questions matter. No longer "is the use transformative?" but two more awkward ones: how did you get the data, and can you prove a market harm? And both of them, worth saying up front, reward scale and resources, not the individual creator.

The 2025 rulings: the battleground shifts

Three decisions, read together, tell the story of the shift.

Thomson Reuters v. Ross Intelligence (11 February 2025). Judge Stephanos Bibas held that using Westlaw's headnotes — Thomson Reuters' proprietary legal summaries — to train a competing legal-research engine was not fair use. The decisive point wasn't a judgment about AI as such, but the fact that Ross was building a direct market substitute for Westlaw. As Bibas put it, the public has no right to Thomson Reuters' legal analysis: copyright exists to encourage the creation of useful tools, not to hand them to a competitor. It's worth stating precisely, because the rest of the story confirms it: Ross didn't lose because "training AI is illegal," but because its product directly replaced the one it had drawn from.

Bartz v. Anthropic (23 June 2025). Judge William Alsup held that training Claude on legally purchased books is fair use — calling the process "spectacularly transformative." At the same time, he found Anthropic liable for downloading more than seven million books from pirate archives like LibGen. The distinction is sharp: training can be fair use, but illegally acquiring the material remains unlawful and answers on its own.

Kadrey v. Meta (25 June 2025). Two days later, Judge Vince Chhabria ruled for Meta on the training of LLaMA — but for a different reason. The authors, among them Sarah Silverman and Ta-Nehisi Coates, failed to show that the model was actually displacing their work in the market or causing concrete economic harm. Chhabria distanced himself from Alsup's emphasis on "transformativeness," pointing to market harm — not transformation — as the factor that ought to decide.

This is the point that holds the three rulings together. Transformativeness isn't "dead," and it hasn't been exposed as a fiction: in Bartz it won, and in glowing terms. But it has been demoted. On its own it's no longer enough, and on its own it no longer condemns: the ground on which cases are decided has moved to how the data was acquired and whether the harm can be proven. Anyone claiming that judges are "seeing through" transformativeness is describing a trend their own rulings don't support. The more useful truth is different: transformativeness has become necessary but not sufficient.

The fourth factor and the probatio diabolica

If the battleground is market harm — the fourth fair-use factor — the practical question becomes: who can actually prove it? Thomson Reuters won because the substitution was direct and obvious. Bartz and Kadrey held because the plaintiffs couldn't quantify the damage. And here a probatio diabolica emerges: how does an author mathematically demonstrate a diffuse harm, spread across millions of works and users, when the effect is real but hard to isolate? We're building a system in which protection depends on the ability to prove what is often intuitively obvious — and proving it is expensive. This is the first way the structure rewards scale: demonstrating market harm is within reach of a Getty or a Thomson Reuters, far less so for the individual illustrator.

Hollywood and the price of scale: Disney and Universal v. Midjourney

In June 2025 the heavyweights arrived. Disney and Universal sued Midjourney — the first time major Hollywood studios have taken an AI company to court for copyright infringement. The complaint, over a hundred pages long, accuses Midjourney of drawing on a vast number of protected works to train its software, iconic characters included, and describes the service as a "virtual vending machine" for unauthorized copies. With more than twenty million registered users and roughly $300 million in revenue in 2024, Midjourney is one of the largest image generators in the world.

The point isn't the Hollywood drama. It's that Disney and Universal have what individual artists lack: the resources to sustain years of litigation and the political weight to command legislative attention. As one industry observer summed it up, this isn't Hollywood trying to shut down generative AI; it's about compensation. The distinction matters — not stopping innovation, but getting paid — but it matters most for those who can afford to assert it.

Actors and the value paradox

Acting is where the question turns existential, because here the performer's identity is the job. Cloning a face, a voice, and a style turns a performance from a one-off act into a replicable template. Studios have already "resurrected" departed performers and digitally manipulated existing performances; the film Here, with its full digital recreations of Tom Hanks and Robin Wright, shows the opposite model, though — authorized use, with explicit consent and rights paid for. The problem, again, isn't the technology: it's consent and compensation.

There's an apparent paradox here. The most famous actors are the most exposed to algorithmic replication, because their very fame has produced the enormous body of material a model trains on. But that same fame also buys the defense: they're the ones with the lawyers, the Here-style contracts, and the political backing behind proposals like the NO FAKES Act — which Disney supports precisely to protect voice and likeness from unauthorized replicas. Fame is both the target and the armor. The truly defenseless creator, once again, is the one without scale.

Adobe: a bet, not a virtue

While the others defend themselves in court, Adobe has tried to sell the "ethical" alternative: Firefly as "commercially safe AI," trained — it says — on images licensed from Adobe Stock and on public-domain content, with tools like Content Credentials and a "Do Not Train" tag. It's best read for what it is: a product and positioning strategy, not a virtue.

The numbers show the limits. According to Bloomberg revelations from April 2024, around 5% of Firefly's dataset consisted of images generated by competing AI, Midjourney included; within Adobe Stock, 57 million images are labeled as AI-generated, 14% of the total. Adobe responds that every image passes through a moderation process that excludes recognizable trademarks and characters, and that the model stays cleaner than those built on entirely unlicensed data. Maybe so. But "cleaner than the competition" isn't "ethical": it's a different degree of the same problem, and for the end user it amounts, at most, to lower legal exposure — not none. It works as a plausible industrial bet that licensing will become the norm. It doesn't work as a moral license.

Europe and mandatory transparency: the AI Act

While the United States picks its way through the fair-use labyrinth case by case, Europe chose the opposite route with the AI Act, in force since August 2024 and now being implemented. The law requires providers of general-purpose models to publish a "sufficiently detailed summary" of their training data, protected material included; in January 2025 the Commission released a template to help draft it. Three pillars: transparency about sources, respect for EU copyright law regardless of where the training takes place, and respect for the opt-out of rights-holders.

The effect is extraterritorial: the law applies to anyone placing a model on the EU market, regardless of the jurisdiction in which the copyright-relevant acts occur. On one hand this creates standards incompatible with American case law and risks fragmenting the global market. On the other, mandatory transparency is itself a cost the large players absorb more easily than the small ones: the same logic of scale, in regulatory form.

What can be protected: the Copyright Office report

In January 2025 the U.S. Copyright Office published Part 2 of its report on AI, clarifying what is protectable. The principles: only works whose expressive elements are determined by a human author can be protected; merely supplying a prompt is not enough; AI assistance does not by itself preclude protection; works generated entirely by AI cannot be protected.

The line is thin and partly arbitrary — what really separates someone selecting among a thousand model outputs from someone selecting among a thousand lines of code? — but it remains decisive, and drawing it requires the ability to document the human contribution. Here too, an advantage for those with processes and records, not for those improvising.

The international mosaic

The global picture is fragmented. A Beijing court, back in November 2023, recognized protection for an AI-generated image provided it showed originality and human intellectual input. A Czech court, in 2024, denied protection to an image created via prompt, aligning with the American position. Different states, different principles: fragmentation isn't a transitional phase, it's the condition companies have to operate in.

Conclusion: who's winning, for now

So what's left of 2025? Not the answer everyone was waiting for. The rulings converge on two principles — piracy is never fair use, and market harm is the real battleground — and diverge on almost everything else. But the thread running through them is single: the system meant to protect creators is rewarding scale. The winner is whoever can prove the harm, sustain the lawsuit, negotiate the licenses. Thomson Reuters, Getty, and Disney are actively monetizing their archives as training data and building a market — on the music industry's ASCAP/BMI model — that risks shutting out precisely the smaller creators in whose name the battle began. It's the heterogenesis of ends: lawsuits born to defend the individual that end up consolidating the giants.

The question to carry forward is no longer whether AI can infringe copyright. It's whether legal systems can build coherent rules fast enough to govern a technology moving at this pace — and whether those rules will protect the people they claim to, or merely ratify a market between titans. Applying twentieth-century laws to twenty-first-century technology, the risk isn't only that we stop protecting whom we meant to protect: it's that we obstruct forms of expression that don't fit the old categories. 2025 didn't close the case. It only clarified who, for now, is winning it.

P.S. The Frankenstein paradox usually gets told wrong: Frankenstein is the creator, not the creature — a common mistake among those who haven't read the novel. It's worth recalling, because Mary Shelley's question wasn't "who is the monster?" but "what does the one who creates and then turns away answer for?" In 2025 the courts began naming the creators — the companies — instead of chasing the creature. But they called them to account only for the most conspicuous crime, Anthropic's piracy, not for the systematic use of others' work waved through under the label of fair use. It's a start. It's also the measure of how far there is to go.

Sources and resources:

• Rulings: Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc. (D. Del., Feb. 2025); Bartz v. Anthropic (N.D. Cal., June 2025); Kadrey v. Meta (N.D. Cal., June 2025); Disney and Universal v. Midjourney (June 2025); Andersen v. Stability AI (ongoing).
• Regulation (EU) 2024/1689 (AI Act) — European Commission official site.
• U.S. Copyright Office, report on AI, Part 2 (January 2025).
• For pending cases: BakerHostetler's AI-copyright case tracker.