A. Germany: GEMA v. OpenAI Ruling (Key Excerpts)

November 11, 2025, Munich. When Presiding Judge Elke Schwager began reading the ruling in the 42nd Civil Chamber, reporters in the gallery set down their pens. The key sentence of the 65-page decision was too clear to miss. "The reproduction of memorized content is copying."

This case begins with song lyrics. Helene Fischer's "Atemlos," Herbert Groenemeyer's "Maenner," Reinhard Mey's "Ueber den Wolken." Songs any German can hum along to. GEMA is the organization that manages the copyrights to these lyrics. When they posed a simple question to ChatGPT, something strange happened. The AI began reciting the lyrics in full. Almost without a single character wrong.

OpenAI's legal team deployed a familiar defense. "Our model stores nothing. It only learns statistical correlations between words." Think of a faucet. To determine whether it leaks, you don't need to look inside the plumbing. If there's a puddle under the sink, it's leaking. The Munich court applied exactly this logic.

The technical explanation in the ruling is lengthy. It may be the most technically detailed copyright decision ever written by a European court. The operating principles of large language models, the meaning of parameter weights, the tokenization process. But all that technical reasoning comes down to one sentence. "Copying is copying, and memorization is memorization."

The court rejected OpenAI's arguments point by point. First, it dismissed the logic that "the user entered the prompt, so the user is responsible." The entity that designed the model, selected the training data, and operates the service is OpenAI. The immunity logic once applied to simple intermediaries does not work here.

Second, it excluded the application of the Text and Data Mining (TDM) exception. The TDM exception is similar to a researcher reading books in a library and taking notes. Analyzing the contents of a book and extracting patterns is permitted. But photocopying an entire book and stacking it in a warehouse is a different matter. The Munich court found that OpenAI's conduct was closer to the latter. If lyrics are "imprinted" inside the model and can be reproduced under certain conditions, that is not mere analysis but storage.

Third, OpenAI's claim of nonprofit research institution immunity was also rejected. Judge Schwager's reasoning was concise. To qualify for immunity, you must reinvest 100% of revenue in research and development or demonstrate a government-recognized public interest purpose. OpenAI proved neither. The sharpest passage concerns "machine readability." OpenAI argued: "The rights holders did not indicate a prohibition on use in a machine-readable format." The court responded: "The defendant boasts that its AI can understand the nuances of Shakespeare and analyze complex legal contracts. Yet it claims it cannot understand the phrase 'unauthorized reproduction prohibited' written on a website?"

The practical effects of the ruling fall into three categories. Injunction: further training on and output of lyrics belonging to the GEMA repertoire is prohibited. Disclosure: OpenAI must submit documents regarding the scale of past use and revenues. This is like being ordered to hand over a stack of receipts. It forms the basis for calculating damages. Damages: liability has been established, so now the amount must be determined.

There is one more thing. The court ordered OpenAI to publish a summary of the ruling in a local newspaper. A symbolic but powerful remedy. It means: publicly acknowledge that you were in the wrong.

OpenAI announced it would appeal. The Munich Higher Regional Court's judgment remains pending, and referral to the Court of Justice of the European Union is also possible. But even this first-instance ruling sends a clear message. In Europe, something being publicly available on the internet does not mean anyone can take it freely. If data is the new oil, that oil lies beneath land that has an owner. To extract it, you must purchase mining rights.

When this ruling meets Article 53 of the EU AI Act, the impact grows larger. The AI Act requires providers of general-purpose AI models to conduct "training in compliance with EU copyright law." The GEMA ruling shows concretely what that obligation means. Secure a blanket license, or give up the European market. There are only two choices.

B. United States: Anthropic/Meta Fair Use Rulings

June 23, 2025, San Francisco federal court. Judge William Alsup finishes signing. Two days later, on June 25, in another courtroom of the same building, Judge Vince Chhabria signs yet another ruling. Both decisions answer the same question. "Is AI training fair use?" The answer is the same. "Yes." But the paths to that "yes" differ, and the blanks left unfilled differ even more.

"Fair Use" can be compared to borrowing someone else's ingredients to cook a new dish. A student quoting a book in a paper. A critic analyzing a scene from a film. These are acceptable without the original creator's permission. Section 107 of the U.S. Copyright Act opens this door. AI companies are trying to walk through it. "We read books and learned from them. It's the same as a human writer reading thousands of books and developing their own style."

The background of Bartz v. Anthropic is this. Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson. Three writers sued Anthropic. Their claim: Anthropic used millions of books without permission to train its Claude model. There is something unusual. Anthropic legitimately purchased some books, but downloaded over 7 million from pirate sites like LibGen and Pirate Library Mirror for free.

Judge Alsup brought the blade down precisely. One edge cut Anthropic; the other spared it. "Copying for the purpose of training is quintessentially transformative and constitutes fair use." But in the same ruling, he adds: "Downloading pirated copies to build a permanent centralized library is not transformative."

The concept of "transformative" needs explanation here. If you don't serve the raw material as-is but transform it to create something new, the chance of being recognized as fair use increases. Judge Alsup found that AI training qualifies as such transformation. "Authors cannot prevent others from reading and learning from their works. For centuries, people have read and reread books. AI training is not trying to replace books; it is trying to make something different."

The pirated copies issue, however, is separate. Judge Alsup's logic is clear. "Downloading pirated copies to build a research library and keeping them around in case they might be useful for something later is a separate use in itself. It is not transformative."

After this ruling, the situation shifts rapidly. In August 2025, Judge Alsup certifies the case as a class action. The class represented by the three authors includes the rights holders of nearly 500,000 works contained in the LibGen and PiLiMi datasets. Under U.S. copyright law, statutory damages for willful infringement are up to $150,000 per work. Multiply that by 7 million titles, and potential damages could reach hundreds of billions of dollars.

Facing a December trial, both sides move toward settlement. The settlement amount announced in September 2025 is $1.5 billion. The largest AI copyright settlement in history. Approximately $3,000 is distributed per work.

Kadrey v. Meta reaches a different conclusion. Richard Kadrey, Sarah Silverman, Ta-Nehisi Coates, and ten other authors sued Meta. Their claim: Meta used pirated books to train its LLaMA model. Judge Chhabria ruled in Meta's favor.

There is a difference. Judge Chhabria was more cautious than Judge Alsup. He warns multiple times in his ruling. "This decision does not establish the proposition that Meta's use of the works was lawful. It establishes only the proposition that the plaintiffs made flawed arguments and failed to build a record supporting the correct ones."

The crux is "market harm." The fourth factor in fair use analysis. Did it damage the market value of the original work? The authors argued: "If AI mass-produces text similar to our books, our book market gets eaten away." Judge Chhabria found this claim too abstract and hypothetical. Just as "teaching a student to write does not automatically erode the market for existing authors."

Place the two rulings side by side and a pattern emerges. "Training" and "hoarding" are different. Training can be transformative. But downloading pirated copies and stacking them in a warehouse is not. If the output copies the original and harms its market, there is a problem. The burden of proof lies with the plaintiff.

In the Anthropic case, Judge Alsup separated the hoarding of pirated copies and imposed liability for it, while in the Meta case, Judge Chhabria viewed downloading and training as a single integrated process. The difference in perspective between the two judges will become a contested point in future litigation.

The conclusion is complicated. The United States leaves open "fair use for training" while preserving "operational illegality." Where the data came from, how it was stored, how closely the output resembles the original. These three questions will determine the coordinates of every AI copyright lawsuit going forward.

C. China: AIGC Platform Rulings

February 8, 2024, Guangzhou. Two images appear side by side on the Internet Court's screen. On the left is the Ultraman character submitted by the plaintiff. On the right is an image generated by the defendant's AI platform. The judge compares the two images. The silver and red body, the distinctive eye shape, the Color Timer on the chest. They look alike. No, they are nearly identical.

This ruling is the world's first final judgment recognizing copyright infringement liability of a generative AI platform. The time from filing to judgment: one month and three days. That is the speed of the Guangzhou Internet Court.

The plaintiff is Shanghai Xinchanghua Cultural Development Co., Ltd., a company that holds an exclusive license for the Ultraman series in China from Japan's Tsuburaya Productions. The defendant is an AI platform operator referred to by the pseudonym "Tab." When users entered Ultraman-related prompts on this platform, images substantially similar to Ultraman were generated.

Here we need to understand "platform liability." Think of a parking lot. Someone parked illegally. Is only the person who parked the car at fault, or is the manager who left the entrance open without any warning signs also at fault? Chinese courts ask the latter question. "Is the platform a mere conduit, or an active participant?"

The Guangzhou Internet Court's answer is clear. The platform is an active participant.

The court's reasoning proceeds in two stages. First, the question of infringement. Ultraman images have high recognition in China. They are easily accessible on major streaming platforms like iQiyi. The defendant's access to these images is sufficiently probable. The generated images substantially reproduce the original creative expression of Ultraman. Therefore, the reproduction right and adaptation right have been infringed.

Second, the attribution of liability. The defendant argues: "We only provided the tool. The user entered the prompt." The court does not accept this argument. Through its AI model, the defendant directly generates infringing images in response to the keyword "Ultraman." This is not mere intermediation but content creation.

The court cites the "Interim Measures for the Management of Generative Artificial Intelligence Services" implemented in August 2023. Under these regulations, generative AI service providers have three obligations. First, establish a mechanism for rights holders to file complaints. Second, explicitly warn users about the risk of copyright infringement. Third, label content as AI-generated. The defendant fulfilled none of these obligations. The damages awarded are 10,000 yuan (approximately $1,400). The amount is small. But the significance is large. The court ordered the defendant to implement "filtering measures for Ultraman-related keywords." Not a reactive measure, but a preventive one.

This ruling is only the beginning. A second lawsuit filed by the same plaintiff against a different AI platform was heard at the Hangzhou Internet Court in September 2024. This time it is a "LoRA model" case.

LoRA stands for "Low-Rank Adaptation." If you repeatedly paint the outline of a specific character onto a large canvas (the base model), a layer of paint forms that can reproduce that outline at any time. In the Hangzhou case, the defendant platform allowed users to upload Ultraman images to train LoRA models and share them.

The Hangzhou court's analysis differs from the Guangzhou court's. In this case, the platform did not directly generate the images, so there is no direct infringement. However, because the platform neglected its duty of care to monitor whether user-uploaded models infringed copyrights, it bears contributory infringement (indirect infringement) liability. Damages awarded: 30,000 yuan (approximately $4,200). On December 30, 2024, the appellate court affirmed this ruling.

Combining the two rulings reveals China's approach. It is a "tiered classification" liability framework. At the input and training stages, courts are relatively lenient. Space is left for technological innovation. But at the output and distribution stages, courts are strict. When you are making money, you must not infringe on others' rights.

The practical implications are clear. A platform's duty of care is proportional to its profitability. The higher the revenue from the service, the higher the standard applied. Keyword filtering, complaint mechanisms, AI-generated content labeling. These three items form the mandatory checklist. A platform lacking these is deemed to have "failed to exercise reasonable duty of care."

Chinese courts ask: "Who could have stopped it?" The answer to that question determines where liability falls.

D. Beijing Internet Court: AI Text-to-Image Ruling

November 27, 2023, Beijing. An image appears on the Internet Court's screen. A young Asian woman standing in twilight. Flawless skin, dreamlike dark eyes, auburn hair flowing over her shoulders. This woman does not exist. It is not a photograph. It is an image created by a man named Li Yunkai using AI software called Stable Diffusion.

There is a reason this ruling generated intense interest. At the same time the U.S. Copyright Office was declaring that "images not drawn by a human have no copyright," the Beijing court reached the opposite conclusion.

The background of this case is straightforward. On February 24, 2023, Li Yunkai used Stable Diffusion to generate images. He entered more than 150 prompts, set negative prompts, adjusted parameters, and changed seed values until he obtained the results he wanted. He then posted one of these images on Xiaohongshu, a social media platform, under the title "Spring Breeze Brings Tenderness."

On March 2, a blogger named Liu Yuchun used this image in her own poetry post. She had removed the original watermark. Li Yunkai filed a lawsuit, alleging infringement of his right of attribution and right of information network dissemination.

Three issues were in dispute. First, is an AI-generated image a copyrightable work? Second, who is the author? Third, did the defendant's actions constitute infringement?

The Beijing Internet Court applied four criteria to determine copyrightability. Does it belong to the field of literature, art, or science? Does it have a form of expression? Does it possess originality? Is it the product of human intellectual effort? The first two were easily satisfied. The image is a visual artwork similar to a photograph or painting. The difficulty lay in the latter two.

"Intellectual achievement" means the result of human intellectual activity. The court analyzed Li Yunkai's creative process in detail. He conceived the character's appearance. He selected an art type: "ultra photorealistic" and "color photo." He set the subject as "Japan idol" and described skin tone, eye color, and hair color in detail. He specified the environment with "golden hour" and "dynamic lighting." He determined the pose with "cool pose" and "viewing at camera." Then, examining the initial output, he added prompts and modified parameters until he obtained the image he envisioned.

The court's conclusion: "The plaintiff made a considerable degree of intellectual investment from conception to final selection. This image is the plaintiff's intellectual achievement." Originality was also recognized. Under Chinese copyright law, originality does not require absolute novelty; the question is whether the creator's individuality is reflected in the work. The court found that Li Yunkai's prompt design and parameter adjustments reflected his "aesthetic choices and individualized judgment." This image was not mechanically generated. It was a creative work carrying the individual expression of a human being.

Who is the author? Article 11 of China's Copyright Law limits authorship to natural persons, legal persons, and unincorporated organizations. An AI model cannot be an author. Nor can the AI developer. The developer's intellectual contribution lies in building the AI tool, not in creating this particular image. The author is Li Yunkai, who personally entered the prompts, adjusted the parameters, and selected the output.

Whether defendant Liu Yuchun's actions constituted infringement was easily determined. She used the plaintiff's image without consent and removed the watermark. She infringed both the right of attribution and the right of information network dissemination. The court ordered a public apology and damages of 500 yuan (approximately 70 US dollars).

The amount is small. But the door this ruling opens is enormous.

Compare the approach taken by the U.S. Copyright Office. In March 2023, the Office partially canceled the copyright registration of "Zarya of the Dawn," a graphic novel created by Kristina Kashtanova using Midjourney. The reason: "the human is not the mastermind of the final output." AI-generated images themselves are not eligible for copyright protection; only the text Kashtanova wrote and her arrangement of the images received copyright recognition.

The Beijing court chose a different path. AI is a tool. It is like a brush or a camera. When the camera was invented in the 19th century, people asked, "How can a photograph, taken by merely pressing a shutter, be art?" Today, photography is art. The Chinese court says AI is the same.

The policy implications of this ruling are clear. China encourages content creation with AI as a "creative activity." A person who lacks the skill to draw can now visually realize their ideas through AI. And the resulting work receives legal protection. This is a detonator for explosive growth in the AI tools market.

There is a caveat. The court emphasized "disclosure of AI use." It is a matter of good faith and public notice. One must not pass off an AI-made image as something personally hand-drawn. In September 2025, the Beijing Internet Court, in a follow-up case, tightened the standard: "To claim copyright in an AI-generated work, one must demonstrate creative effort."

Place the four rulings side by side and the world comes into view. Germany chose "regulation and protection." Training is copying, and reproduction is infringement. The United States is in the process of "re-examining fair use." Training is transformative, but it is not piracy. China chose "encouragement under control." Strict toward platforms, generous toward users.

AI companies must satisfy all three standards simultaneously. They must balance like acrobats. In the process, enormous legal fees and lobbying funds will be spent. Those who design the system, those who use the system, those who seek to regulate the system. Their war has only just begun.