Context: Dozens of Generative AI (GenAI or GAI) copyright lawsuits are pending in U.S. courts from the West Coast to the East Coast, and they take time to be resolved (see, e.g., September 28, 2024 ai fray article). Some other jurisdictions are known for relatively speedy decisions, such as Germany.
What’s new: The Hamburg Regional Court, a German court of first instance has now handed down a much-awaited GenAI copyright decision. What proved outcome-determinative is that the defendant is a beneficiary of an exception from EU copyright rules that applies to scientific text and data mining (TDM). The decision is appealable and the matter may ultimately reach the European Court of Justice (ECJ). It isn’t limited to the single outcome-determinative question of the TDM exception, but also addresses other aspects of this case that are going to be found in other AI copyright cases, such as whether robots.txt files prohibiting crawling must be respected (the court thinks so) and the need for a clear distinction between collection, input and output.
Direct impact: This case was brought over one image made by a professional photographer and used in image-keyword pairs for GenAI purposes. That doesn’t make it a high-stakes case in commercial terms as far as this particular plaintiff’s interests are concerned.
Wider ramifications: It’s clear that the case is more about the transcendental questions surrounding the use of copyrighted material by GenAI systems than about the alleged harm to the plaintiff. Therefore, an appeal is a given. There is a possibility that the Hamburg Higher Regional Court or the Federal Court of Justice of Germany will refer certain AI-related copyright questions to the ECJ. The next court must hear the appeal, but has no obligation to make a referral to the ECJ; the Federal Court of Justice may decline to hear the case, but if it does, it has an obligation to refer, within reason, legal questions of first impression to the ECJ.
The ruling was published in German (Openjur link), and German court rulings are published so quickly only when there is an exceptionally strong interest by a large group of stakeholders, such as an entire industry.
With the greatest respect, a decision by a German court of first instance on a legal question of first impression is of limited formal influence. Unlike in common law jurisdictions, the ruling is not even binding on that same court in any way. But in this case, with courts in different jurisdictions (note that differences between U.S. and EU copyright law are huge) grappling with certain questions for the first time, any formal ruling gets more attention (and may influence other decisions to a greater extent) than it otherwise would.
The way the ruling is written reflects that the court was well aware of the pioneering role it had here. That’s why it didn’t take a shortcut, but went beyond the call of duty in discussing aspects of the case that the court would have (certainly or at least likely) resolved in plaintiff’s favor:
- A general TDM exception under German copyright law, which implements Art. 4 of the 2019 EU Copyright in the Digital Single Market (DSM) Directive, would probably not have gotten the defendant off the hook. It was only the fact that the data collector (LAION, though the name was redacted out of the official redacted version of the ruling) is a scientific organization that made a TDM exception under Art. 3 of the EU’s Copyright in the DSM Directive apply. But the same case against a purely commercial operator might have succeeded.
- It also played a role that LAION has been in existence since the 2000s, long before GenAI became a big thing. That’s why the argument that (EU and German) lawmakers didn’t have GenAI in mind when defining certain exceptions from copyright was deemed unavailing.
- The defendant’s argument that the use of the copyrighted material in question was merely “transient” didn’t fare better, though.
- The court made a pretty strong (though not legally definitive) statement on whether data miners must respect crawling prohibitions in norobots.txt type files. The decision says that those providing data for GenAI purposes must also use AI to interpret such files, so they can’t just argue that such prohibitions aren’t machine-readable. (In fact, it takes a lot less than AI to properly parse such files.)
- The court insisted on a proper distiction between data collection (which is what the defendant in this case did), input for LLM training purposes and the use of the copyright material for the generation of new material (output). Note that the terms emphasized in the previous sentence are not used in the same form, and particularly not emphasized, in the court ruling. But the distinction is there.
Arguably, the plaintiff in this case tackled the legally (even if not in financial terms) most difficult opponent: a non-profit that collects data. They tried to argue that it’s not really a non-profit because of the financial interests of some of its key players, and they raised other points, but in the end it appears that the court believed a scientific non-profit to be a clearly intended beneficiary of a TDM exception from EU copyright law.
Regardless of the the fact that this ruling isn’t binding on anyone, it contains some obiter dicta that could take center stage in other cases, provided that those other cases are brought against commercial operators. But in such cases a number of other questions, and particularly some other defenses, will have to be addressed. That’s why one can’t just conclude from the LAION decision that the same fact pattern in a case against, say, OpenAI would have necessarily succeeded. The plaintiff would, however, have surmounted the hurdle that proved dispositive in this case.
EU copyright law, with which German copyright law must be consistent though it can still have its own features, does not come with a fair use defense the way it exists in the U.S., but there are some rules in Europe that collectively have a somewhat similar effect. They are called limitations and exceptions (and in some EU languages, a single term that means a combination of both is used, such as “Schranken” in German). There also are rules in the EU for highly transformative derivative works. But it’s still rather different from the mulifactorial fair use analysis under U.S. copyright law. On balance, U.S. fair use is stronger than the limitations and exceptions that exist in the EU, though in this GenAI copyright case in Germany, one of those limitations disposed of the complaint, for the time being.