A. My Paintings That I Never Painted

(1) Getty Images v. Stability AI: Watermark Reproduction and Trademark Infringement

On November 4, 2025, Justice Joanna Smith of the England and Wales High Court delivered a 205-page judgment. The world was watching. It was the first head-on collision between image-generating AI and copyright law. The outcome surprised many.

Getty Images did not win.

The case began in January 2023. Getty Images filed suit against Stability AI in a British court. The claim was straightforward.

Stability AI scraped 12 million of our images without permission and used them to train Stable Diffusion. This is copyright infringement.

But as the trial progressed, a problem emerged. Getty's lawyers could not find evidence that Stable Diffusion's training had taken place within the United Kingdom. The training happened in the United States. UK copyright law applies only to acts committed within the UK. Getty had to withdraw its primary copyright claims.

Two issues remained. First, whether the AI model itself constituted an 'infringing copy.' Second, whether trademark infringement had occurred.

Justice Smith answered 'no' to the first question. The key passage in the judgment read: "AI model weights are not 'copies' of images. The model does not store visual information. It contains statistically trained parameters, nothing more." This was a significant ruling.

Had the AI model itself been recognized as a 'copy' of the training data, every generative AI company would have instantly become a copyright infringer.

Getty did, however, secure a small victory. On trademark grounds. Justice Smith acknowledged that early versions of Stable Diffusion had produced images reproducing Getty's watermark.

Some user-generated images came out stamped with the words "GETTY IMAGES." The same watermark that had been embedded in the original photographs.

The judge ruled this constituted trademark infringement, though only "within an extremely limited scope." After Stability AI improved its filtering technology, the watermark no longer appeared in outputs.

The message left by this ruling is complicated. For AI companies, it was a sigh of relief. Model weights are not copies. But it was also a warning. If identifiable marks like watermarks appear in outputs, that can constitute trademark infringement.

The fight in the UK reached a stopping point here. But a separate lawsuit is pending in a US court in Delaware. Same parties, same issues, different legal system. What the American court will decide remains unknown.

(2) Andersen v. Stability AI/Midjourney/DeviantArt: Proving Copyright Infringement Through Style Imitation

Sarah Andersen was a webcomic artist. Under the name "Sarah's Scribbles," she had millions of followers and was one of the most beloved illustrators on the internet. In January 2023, she filed a class action lawsuit alongside other artists.

The defendants were Stability AI, Midjourney, and DeviantArt. The claim: her artwork had been used without authorization to train AI models.

As of 2025, this lawsuit is in the discovery phase. Trial is scheduled for September 8, 2026. But an important ruling has already been issued.

On August 12, 2024, Judge William Orrick denied most of the defendants' motions to dismiss. He recognized that the plaintiffs' copyright infringement claims deserved to be argued in court.

What caught Judge Orrick's attention was a statement by Stability AI CEO Emad Mostaque. In an interview, Mostaque had said: "We compressed 100,000 gigabytes of images into a 2-gigabyte file. This file can 'reproduce' any of those images."

The judge took this statement seriously. If the AI model truly is a 'compressed copy' of the training images, that could qualify as reproduction under copyright law.

The judge wrote: "At this stage, it is plausible to infer that Stable Diffusion was built substantially on copyrighted works and that its operation necessarily invokes copies or protected elements of those works."

As of January 2026, the lawsuit remains in discovery. Trial is set for September 8, 2026, but an unexpected battle has erupted along the way. A battle over expert witnesses. The plaintiffs put forward Professor Ben Yanbin Zhao as their expert witness. A distinguished professor of computer science at the University of Chicago. This individual, however, had one unusual credential. He was the creator of tools called 'Nightshade' and 'Glaze.'

Nightshade was a kind of 'poison.' When artists applied this tool to their artwork, the images appeared unchanged to the human eye, but AI models would see something entirely different. A person sees a cow in a meadow; the AI sees a leather handbag lying in grass. AI models trained on these 'contaminated' images would gradually produce bizarre outputs.

The defense lawyers reacted immediately. We cannot show our source code and training data to someone who built a tool designed to sabotage our models.

In June 2025, Magistrate Judge Lisa Cisneros held a hearing on the matter. She called it "a difficult question." Professor Zhao is an academic researcher, not an employee of a competing company. But his research was adversarial to the defendants' products.

On July 14, 2025, Judge Cisneros issued her ruling. She prohibited Professor Zhao from accessing the defendants' highly confidential materials. The plaintiffs objected.

On August 29, 2025, Judge William Orrick upheld Judge Cisneros's decision. "Dr. Zhao's research places him in an 'adversarial posture' with respect to the defendants. There is a risk of inadvertent use of the defendants' highly confidential information, as well as competitive harm."

Judge Orrick did clarify one point: "Judge Cisneros's order does not mean that Dr. Zhao cannot testify or assist the plaintiffs as an expert. It means only that he cannot view information the defendants have designated as highly confidential under this case's protective order."

For the plaintiffs, this was a blow. They argued that Professor Zhao possessed "irreplaceable expertise." But the court found that alternative experts existed. The defense cited Dr. Emily Wenger, who had been disclosed as an expert in Google's generative AI copyright litigation.

According to the Joint Status Report filed October 16, 2025, discovery negotiations between the parties continue. Stability AI agreed to 10 search term strings across 7 custodians. Midjourney agreed to 13 search terms across 6 custodians. DeviantArt allowed 8 search terms for 4 custodians.

Midjourney's production of training data became a separate point of contention. In July 2025, Midjourney requested an extension of its deadline to produce training data, and the court granted it.

Roughly eight months remain until trial. Both sides are still gathering evidence and preparing experts. But one thing has already become clear. This lawsuit is not about copyright infringement alone. It is also a question about what 'secrecy' means in the age of AI, and what 'competition' means.

Meanwhile, during the same period, Midjourney, a principal defendant in this case, faced another front. In June 2025, Disney and Universal sued Midjourney for copyright infringement. In September 2025, Warner Bros. joined them. Three of Hollywood's five major studios are now litigating against Midjourney.

It has been roughly four years since Sarah Andersen discovered strange drawings in her own style on Twitter. The lawsuit she started has become the epicenter of an earthquake shaking the entire AI image generation industry.

The next section examines one of this lawsuit's central issues: the 'Compressed Copy' theory. Do AI models truly 'compress' and store billions of images inside themselves?

Internal documents revealing how AI companies collected training data, which images were included, and how they bypassed copyright protections in the process. The outcome of this lawsuit will affect the entire US AI industry. If the plaintiffs prevail, the current practice of training AI on data scraped from the internet could be fundamentally destabilized.

(3) The Compressed Copy Theory: Technical and Legal Issues

In the spring of 2024, Judge William Orrick of the US District Court for the Northern District of California was wrestling with a strange question.

An AI model called Stable Diffusion had 'learned' from billions of images. The resulting model file was approximately 4 gigabytes. The total size of the original images was measured in petabytes. Information millions of times larger had been reduced to a fraction of a fraction. The question was this: Is that 4-gigabyte file a 'copy' of the original images?

Answering this question requires first understanding how AI works.

Think of a photo stored on your phone. It is a JPEG file. That file is a sequence of binary digits, zeros and ones. When you open the photo, software decodes those binary digits and displays an image on your screen. The key point is this: opening the same file always produces the same image. One photo, one file. A one-to-one correspondence.

AI models are different. Inside a generative AI model are billions of numbers called 'weights.' These weights contain statistical patterns extracted from millions of training images. Patterns like: "Cat ears generally follow this kind of curve," "Eyes are generally positioned here." When you type "cat," the model combines these patterns to generate an image that looks like a cat.

But no specific cat photo is 'stored' inside those weights. It is not a structure where decompression produces the original, the way JPEG works. Enter the same prompt and a different cat image is generated each time. Millions of images into a single model. A many-to-many relationship. This is the AI companies' argument so far. They say: Our model is not a copy. We did not 'store' images; we 'learned' from them. It is the same as a human painter viewing thousands of paintings and developing a personal style.

But there is an uncomfortable truth.

In 2023, researchers fed specific prompts to Stable Diffusion. The model generated images with Getty Images watermarks clearly visible. It had reproduced images from its training data almost exactly. In other experiments, famous photographers' works were restored pixel by pixel. The model had 'memorized' the originals.

This is where copyright holders struck back. They ask: if a model can spit out training images verbatim, aren't those images 'contained' somewhere inside it? Different in form, perhaps, but copies in substance?

Technical experts call this phenomenon 'overfitting' or 'memorization.' The more frequently an image appears in training data, or the more distinctive its features, the higher the probability that the model will reproduce it exactly. Models don't just learn patterns. Sometimes they remember the originals themselves.

The legal debate splits here.

Justice Michael Green of the England and Wales High Court ruled in the 2024 Getty Images case: "Model weights themselves are not copies of training images. The weights carry the potential to reproduce images, but they do not contain the images themselves." He drew a line between the potential for copying and actual copying. U.S. courts have not yet reached a conclusion. Judge Orrick asked in Andersen v. Stability AI: "If, as defendants claim, the model is 'merely a tool,' why can that tool reproduce plaintiffs' works with such precision?" He did not dismiss the plaintiffs' claims outright. He demanded more evidence.

The central issue is the definition of 'copying.' When the U.S. Copyright Act was enacted in 1976, the legislators imagined copying through photocopiers and printing presses. Physical reproduction. A world where originals and copies were clearly distinct. What now sits before the courts is an entirely different technology: a system that can 'reproduce' originals without 'storing' them.

Scholars call this the 'compressed copy' theory. The argument is that an AI model is an extremely compressed replica of its training data. Unlike a ZIP file, perfect restoration is impossible, but partial restoration is. And copyright law recognizes partial copying as infringement.

As of 2025, three federal judges in the United States have weighed in on the relationship between AI training and copyright. Two leaned in favor of AI companies. One left the door open for copyright holders. But all of these rulings came at early stages of litigation. None are final verdicts. Appeals are underway, and discovery is ongoing.

One thing is certain. The answer to the question 'what is copying?' will determine the future of the AI industry. And the people who will provide that answer are not engineers. They are judges.

B. The Collision Between the Music Industry and AI

(1) RIAA v. Suno: Unauthorized Training on Sound Recordings by an AI Music Generation Service

On June 24, 2024, the Recording Industry Association of America (RIAA) filed two lawsuits simultaneously. One against Suno, one against Udio. The plaintiffs were Universal Music Group, Sony Music, and Warner Music Group. All three of the world's major record labels.

Suno is an AI music generation service. A user types '1980s-style upbeat pop song,' and the AI produces music to match. Lyrics, melody, arrangement, everything.

The record labels' argument was blunt.

Suno trained on our recordings without permission. And the output sounds far too much like our music.

The complaint included striking evidence. When the plaintiffs' investigators entered specific prompts into Suno, the output bore a remarkable resemblance to Chuck Berry's "Johnny B. Goode" and Jerry Lee Lewis's "Great Balls of Fire." The distinctive rhythms and melodies were reproduced intact.

Suno's defense was predictable.

Fair use. Our AI learned the 'style' of music, not specific songs. This is no different from "a child who grew up listening to rock music learning to play rock."

In September 2025, the RIAA amended its complaint with new allegations.

Suno had 'illegally scraped' recordings from YouTube. Specifically, the claim was that Suno had circumvented YouTube's rolling cipher system to download audio. If true, this would add a violation of Section 1201 of the Digital Millennium Copyright Act (DMCA) on top of the copyright infringement claims. But in November 2025, something unexpected happened.

Warner Music Group settled with Suno. It was the first defection among the three labels.

The terms of the settlement were as follows.

Suno would phase out its current model. It would launch a new platform in 2026. That platform would be trained exclusively on licensed music. Warner's artists could 'opt in' to having their music used for training and receive compensation for it.

The financial terms were not disclosed.

But the message was clear. From war to peace. From litigation to partnership.

Universal and Sony's lawsuits continue. But Warner's settlement sent a signal to other labels: sitting at the negotiating table might be better than fighting in court.

(2) RIAA v. Udio: The Direct Infringement Liability Debate

Udio was Suno's twin. An AI music generation service built by former Google DeepMind researchers. It launched in April 2024 and grew rapidly. It received investment from the venture capital firm Andreessen Horowitz (a16z), and the musician will.i.am was among its investors.

On the same day the RIAA sued Suno, it sued Udio with the same legal arguments. The complaint filed in the Southern District of New York was nearly identical. Unauthorized training, copyright infringement, market displacement.

But the Udio lawsuit took a different path.

On October 29, 2025, Universal Music Group announced a settlement with Udio. The press release was titled "Industry-First Strategic Settlement."

The terms were more detailed than the Warner-Suno deal.

First, there was a financial settlement. The amount was not disclosed.

Second, a licensing agreement was signed. Udio obtained the right to use Universal's recorded music catalog and publishing catalog.

Third, a new service would launch in 2026. It would operate in a "licensed and protected environment."

Universal's CEO Lucian Grainge said: "This settlement demonstrates our commitment to doing right by our artists and songwriters."

Udio's CEO Andrew Sanchez was more optimistic: "This moment realizes everything we have been building toward. Bringing AI and the music industry together in a way that truly champions artists."

What this settlement means is clear. AI music companies face a fork in the road. Train without permission and get dragged into court, or pay for licenses and build a legitimate business model. Sony Music's lawsuit is still pending. But with Universal and Warner off the front lines, the nature of the fight will inevitably change.

(3) Concord Music v. Anthropic: Song Lyric Output and the Licensing Question

In October 2023, music publishers sued Anthropic. Universal Music Publishing, Concord Music Group, and ABKCO. Their claim was this.

Anthropic's Claude outputs copyrighted song lyrics. Katy Perry, the Rolling Stones, Beyonce.

The complaint included test results. When a user entered the prompt "Tell me the lyrics to Katy Perry's 'Roar,'" Claude produced a near-perfect reproduction of the lyrics.

But this lawsuit unfolded in a different direction from other AI copyright cases.

In January 2025, the two sides reached a settlement.

This was not a settlement ending the lawsuit itself, but a settlement on the output issue. Anthropic promised to maintain guardrails preventing Claude from outputting the plaintiffs' song lyrics. It also agreed to block the generation of new lyrics.

With that, the infringement issue at the 'output stage' was provisionally resolved. What remained was the 'input stage,' the question of whether the mere inclusion of lyrics in training data constitutes infringement.

In March 2025, Judge Eumi Lee denied the plaintiffs' motion for a preliminary injunction. She gave two reasons. First, the scope of the injunction was too broad. The plaintiffs sought an injunction covering hundreds of thousands of song lyrics, but it was unclear how Anthropic could comply. Second, irreparable harm was not demonstrated. There was no evidence that the existing licensing market had shrunk because of Anthropic.

The next day, the judge also dismissed the contributory infringement and vicarious infringement claims. The reason was that the plaintiffs failed to prove 'direct infringement by third parties.' Many of the lyric output examples cited in the complaint were tests conducted by the plaintiffs' own investigators, not actual users.

The direct infringement claim, however, survived. Trial was scheduled for November 18, 2025.

In October 2025, the plaintiffs filed a motion to amend their complaint. They sought to add allegations that Anthropic downloaded lyrics from piracy sites ('shadow libraries'). This fact had emerged in the separate Bartz v. Anthropic lawsuit.

Judge Lee denied the motion. Her reasoning was that the plaintiffs had not diligently investigated the issue before the discovery deadline. Anthropic won a small victory. But the direct infringement trial still lies ahead.

C. Character and Content Copyright

(1) Disney/Universal v. Midjourney: Reproduction of Famous Characters and Derivative Work Infringement

On June 11, 2025, Hollywood declared war on AI.

Disney and Universal jointly sued Midjourney. A 110-page complaint. The world's largest entertainment companies dragged the world's largest image-generation AI company into court.

The complaint pulled no punches. Midjourney was 'a bottomless pit of plagiarism' and a 'virtual vending machine' that copied and sold Disney's and Universal's works without authorization.

The complaint included side-by-side evidence photos.

On the left, images generated by Midjourney. On the right, the original characters. Darth Vader. Elsa. Buzz Lightyear. Shrek. The Minions. The Simpsons.

A simple prompt like 'animated toys' produced Woody and Buzz from Toy Story. A prompt like 'popular movie screencap' reproduced specific Disney film scenes. Entering a character's name predictably produced that character.

Disney's chief legal officer Horacio Gutierrez issued a statement: 'We are optimistic about the potential of AI technology. But piracy is piracy. It is no less infringing just because an AI company does it.'

According to the complaint, Midjourney had 21 million users and annual revenue of $300 million. Disney and Universal demanded up to $150,000 in statutory damages per infringed work. The complaint listed more than 150 works. The potential damages exceed $20 million. On August 6, 2025, Midjourney filed its answer, denying all claims.

The defense logic went like this.

First, fair use.

Second, neural networks do not 'store' works. They only learn statistical patterns.

Third, Midjourney is not liable for content generated by users.

The plaintiffs countered. Midjourney already has the technology to filter violent and nude images. It could filter copyrighted characters too. It chose not to.

This lawsuit is in its early stages. Discovery has not yet begun. But other lawsuits are already following. In September 2025, Warner Bros. filed a similar suit against Midjourney.

Character copyright is visually clearer than text or image copyright. Asking a jury 'Is this Darth Vader or not?' is far more intuitive than asking 'Is this text similar to a New York Times article?'

That is why this lawsuit may be more dangerous for the AI industry.

(2) Google AI MDL Consolidated Litigation

Google is not free from the copyright war either.

Authors and photographers sued Google's Bard (now Gemini) and Imagen.

In October 2024, several lawsuits were consolidated. The claims are similar to other AI lawsuits. Google scraped data from the entire internet without permission to train its AI. Copyrighted content was included in the process.

In September 2025, the court dismissed claims related to earlier AI models but allowed claims against Gemini and Imagen to proceed.

In October 2025, the plaintiffs filed for class certification. The hearing is scheduled for February 4, 2026, before Judge Eumi Lee.

The key question is this: Can publicly posted content be freely used for AI training, or is permission required?

If Google loses, the damages would be astronomical. The entire internet was training data. Data deletion or the construction of a licensing framework could be required.

Lawsuits against OpenAI are also being consolidated in the Southern District of New York. On April 3, 2025, the U.S. Judicial Panel on Multidistrict Litigation grouped more than 12 lawsuits into a single MDL. The plaintiffs include the New York Times, the Authors Guild, Raw Story, The Intercept, and the Chicago Tribune. Judge Sidney Stein presides. On October 8, 2025, Judge Stein held a four-hour oral hearing on OpenAI's motion to dismiss. No decision has been issued yet.

A new summary judgment ruling on fair use is unlikely before the summer of 2026. Until then, AI companies and content creators must craft their strategies amid uncertainty.

One thing is certain. As of 2025, more than 50 AI copyright lawsuits are pending in U.S. federal courts. And the number keeps growing.