A US federal court has ruled for the first time that training AI models with copyrighted books is permissible as “fair use”.
This landmark case, Authors v. AI developer Anthropic (Claude AI), could fundamentally change the rules for AI platforms in the US. What does this ruling mean for the AI industry in concrete terms? Furthermore, how would such use of copyrighted works be assessed in Germany, where there is no “fair use” doctrine? This blog post will examine the US case, its significance for AI companies, and highlight the clear differences from German copyright law.
Fair Use Ruling in the USA: AI Training with Books Permitted
A US federal judge has ruled that using copyrighted books to train an AI language model falls under the principle of fair use. This decision emerged from proceedings in the Californian District Court, presided over by Judge William Alsup.
Authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson had sued the AI start-up Anthropic. They alleged Anthropic used millions of books, including pirated copies of their works, to train the chatbot Claude AI without permission.
The court’s decision favored Anthropic. The judge determined that using legally acquired books for AI training was “highly transformative” and thus permissible as fair use.
The AI algorithm serves a different purpose than the original book. Its goal is to generate new, independent text material, not to reproduce the original work. This transformative character is akin to a person reading literary classics, appropriating their style, and creating something new.
Such human activity would not infringe copyright. The court applied this reasoning to an AI model that forms new sentences from many read texts.
Remarkably, much of the training data’s origin played a role in the case. Anthropic initially downloaded over 7 million books from pirate sources like Library Genesis and “Pirate Library.”
Recognizing the legal risk, the company later altered its strategy. Anthropic purchased millions of print copies, dismantled them, and scanned the pages to acquire digital training data legally.
The judge deemed even this mass digitization of purchased books permissible under fair use. The court reasoned that the mere conversion of a printed book into a searchable digital file already constituted a transformative use.
However, the fair use ruling does not grant Anthropic a complete reprieve. Judge Alsup clarified that using pirated copies can still constitute copyright infringement. Fair use only applies when material is lawfully obtained.
Consequently, the court ordered a separate trial. This trial will determine Anthropic's liability for the initial use of illegal copies. If willful infringement is proven, this could lead to damage claims of up to 150,000 US dollars per work.
Despite the pending damages trial, Anthropic welcomed the landmark ruling. For the first time, courts have recognized that AI training can be a transformative, fair use of copyrighted material.
Significance of the Ruling: A Milestone for AI Platforms in the USA
The Anthropic decision marks a significant milestone for the AI industry. For the first time, a court has explicitly confirmed that “extraction” of copyrighted works to train AI models can fall under the fair use doctrine.
This provides US AI developers with substantially greater legal certainty. As long as they obtain training data legally, AI systems can use protected content without requiring licenses from every rights holder.
This is a major relief for the burgeoning generative AI industry, encompassing everything from chatbots to image generators. It echoes the principle from the famous Google Books case, where scanning millions of books for a search engine was also deemed fair use.
Conversely, this ruling represents a setback for authors and rights holders in the USA. Existing lawsuits against AI companies for unauthorized use of works may lose considerable traction if the fair use argument gains wider acceptance.
Several similar lawsuits are currently pending in the US, including cases by bestselling authors against OpenAI and by major film studios against AI image generators. The Anthropic ruling could set a precedent, potentially leading to the dismissal of such lawsuits if no piracy-like conduct is involved.
Authors’ associations are concerned this ruling grants AI companies a carte blanche to commercialize works. However, AI companies contend that their models merely emulate human learning from reading books.
It is crucial to note that this ruling originates from a court of first instance. Therefore, it remains subject to appeal.
Higher courts, or even the Supreme Court, may soon need to provide final clarity on how fair use applies to AI training. Despite this, the tech industry is celebrating the decision.
It establishes a novel legal foundation for AI developers to formulate their data strategies in the USA.
No Fair Use in Germany: What Applies in This Country?
While the US ruling is groundbreaking, it cannot be directly applied to Germany. German (and European) copyright law lacks a general fair use doctrine.
Instead, any use of a work without the rights holder’s consent is prohibited, unless a specific legal exception applies. German copyright law includes several narrowly defined limitations.
Examples include the right to quote, private copying, or reporting on current events. However, it does not feature an open “fair use” clause like US law.
Specific Legal Exceptions vs. Fair Use
Consider this example: A US court would assess whether a use is “transformative” to determine fair use. A German court, conversely, would examine the applicability of specific sections like Section 51 UrhG (quotations) or Section 60d UrhG (text and data mining).
If no such exception applies, the use is inadmissible, regardless of its creative or transformative nature. This strict legal framework aims to ensure certainty.
However, it also means certain types of use remain prohibited in Germany that would be permitted in the USA.
Text and Data Mining: German Approaches for AI Training
In 2019, Europe introduced new exceptions specifically for AI training data, subsequently incorporated into German law in 2021. These provisions aim to facilitate research and innovation by simplifying copyright considerations.
Specifically, two key regulations are relevant:
- Section 60d UrhG – Scientific Text and Data Mining: This rule permits research institutions, such as universities, to copy and evaluate copyrighted material on a large scale. This is allowed provided it serves non-commercial scientific research, and copyright holder consent is not required.
- The term scientific research is interpreted broadly. The Hamburg Regional Court recently ruled that creating large AI training datasets, even by non-profit organizations like LAION, can fall under this research freedom. Importantly, even if commercial companies later utilize such data, the initial dataset remains legal if its creation served knowledge acquisition.
- In the Kneschke ./. LAION case, the downloading of millions of images for AI training was deemed permissible. The court did not consider this a copyright infringement, but rather a permitted act under Section 60d UrhG.
- Section 44b UrhG – Text and Data Mining for Other Purposes: This broader exception permits text and data mining for all purposes, including commercial ones, but under stricter conditions. A key prerequisite is that the user must have lawful access to the material. This means the work must be publicly accessible, licensed, or purchased by the user.
- Rights holders can prohibit this data mining through an opt-out mechanism. Practically, this is implemented via terms of use or technical measures, such as a robots.txt file, which disallow automated reading. In the Hamburg LAION case, a stock photo website's no-scraping clause in its terms and conditions was considered an effective opt-out.
- This opt-out was irrelevant for LAION due to Section 60d (Research) applying, which does not allow rights holders to opt out. However, for a commercial AI company in Germany, such an opt-out would be binding. If a provider prohibited scraping newspaper articles or stock photos for AI training in its terms, Section 44b UrhG would not permit such use.
Key Differences: USA vs. Germany
In summary, the USA and Germany approach AI use differently. In the USA, courts flexibly apply fair use criteria—like transformative use, scope, purpose, and market impact—to determine permissibility. Conversely, in Germany, use must either fall under rigid legal exceptions or be contractually licensed.
There is no general exception for “transformative AI use” in Germany, regardless of the innovation. Nonetheless, new European rules on text and data mining provide a partial equivalent. They facilitate the mass analysis of content, but strictly within defined requirements, such as a research context or the absence of an opt-out for legally accessible sources.
Conclusion: Opportunities and Risks for Companies and Authors
The US ruling favoring Anthropic signals a new direction for AI developers, at least in America. Courts have clearly recognized that AI systems can use extensively protected works for training if they generate new content.
This could significantly simplify data utilization for AI platforms and accelerate innovation. However, the case also established a crucial boundary: Willful use of illegal sources still constitutes copyright infringement, even in the US.
Therefore, AI companies should prioritize lawfully acquiring their training data. It is advisable to buy or license books rather than sourcing them from illicit parts of the internet.
In Germany and Europe, the situation differs significantly. Lacking a fair use principle, companies must meticulously verify the permissibility of using copyrighted works.
While the new TDM exceptions offer some flexibility, they apply only under specific conditions, such as research contexts or the absence of an opt-out. A commercial AI startup in Germany cannot claim a blanket freedom of transformation when using millions of books or images for training.
Instead, it must adhere to barrier regulations or secure license agreements with rights holders. This entails greater effort and legal uncertainty compared to the USA. Conversely, European authors maintain enhanced control over their works. They can opt out of content collection by web crawlers and, if unauthorized use occurs, are entitled to injunctive relief and compensation.
Potential clients, including AI companies and rights holders, should closely monitor legal developments. It is advisable to seek expert advice before processing protected content for AI purposes on a large scale.
In the USA, the Anthropic judgment is fostering a liberalization favoring the AI industry. Germany, however, is still in an exploratory phase. Initial court rulings, like those in Hamburg, indicate how courts are applying the new TDM rules. Nevertheless, it remains uncertain whether higher courts or the ECJ will affirm this liberal approach.
Simultaneously, political discussions persist regarding a fair balance, including potential remuneration models for authors whose works contribute to AI training.
One certainty is that the topic of AI and copyright remains highly dynamic. Companies training AI systems must adopt an international perspective, as what is permissible in the USA may be prohibited in Germany.
Conversely, European rules, such as the TDM exceptions, could ensure that AI innovation remains viable here without infringing copyrights in the medium term. The current US decision represents a significant step for AI, but it is not a worldwide free pass.
The legal area dictates permissibility, and the distinctions between US fair use and German copyright law are undeniable. For a future-proof AI strategy, therefore, it is prudent to monitor both legal landscapes and seek legal certainty when in doubt.