Agentic non-works in German law

Three months have passed since the Higher Regional Court of Düsseldorf issued its (I-)20 W 2/26 decision. The second notable German court decision on the copyrightability of AI outputs. Building on the previous decision 142 C 9786/25 of the District Court of Munich, the North Rhine-Westphalian court (re)stated that not all AI outputs are copyrightable. As not every AI-generated output qualifies as a work in itself.

Picture from @bastianp. Used under the Unsplash License.

Only the AI outputs from human prompting [freie kreative Entscheidung(en)] that carry sufficient creative weight to shape the output to visibly express the distinct taste of the author may be copyrighted (para. 30).

On reflection, I've been following the inside discussions in the legal community about (I-)20 W 2/26 pretty closely. The way the judges imported the Mio/konektra case into their decision sparked some what-if thinking for me. As it's hard for me to imagine how the Mio/konektra logic would work with the self-learning agents, I'd like to explore this further here. Because, as far as I can tell, there isn't any relevant German case law that would directly address the specifics of this type of AI agents; yet.

Fairly simple case?

While the mutatis mutandis application of the Mio/konektra case (Düsseldorf paradigm) is anything but simple, the facts behind it are actually quite clear. The background is as follows:

Photographer (of a unique photo of his dog underwater) asked the Regional Court of Düsseldorf to issue an injunction against an individual who used AI to create and publish a comic-style version of the photo. After some consideration(s), the judges ultimately rejected the injunction application (no infringement), which was later affirmed by the Higher Regional Court of Düsseldorf.

Not so fast with adaptations!

For self-learning agents, what matters most is how the judges (of the Higher Regional Court of Düsseldorf) defined a work in the AI domain under the German Copyright Act.

„Adaptations or other transformations of a work, including, in particular, of a melody, may be published or exploited only with the author's consent. If the newly created work maintains sufficient distance to the work used, this does not constitute adaptation or transformation within the meaning of sentence 1.“ [Section 23(1) of the UrhG]

In their decision, the judges concluded that the AI-generated comic-style image of the dog doesn't qualify as an adaptation under Section 23(1) of the UrhG, and thus can't be legally considered as a work.

My reading of the decision suggests that this assessment rests on the absence of demonstrated distinct design patterns in the work, leading to the legal fiction that the model relied on a collection of existing more or less distinctive design patterns from other authors. Although the Higher Regional Court of Düsseldorf appears to express this view implicitly, it seems that I'm not alone in interpreting the decision in this way.

Mio/konektra twist

While this argument would be a new and refreshing approach to tackling what seems to me a wicked problem in German law, it's actually an imported idea. It comes from the Mio/konektra case in which the CJEU dealt with tables. Is a shape dictated by function, not a creative choice capable of bearing copyright?

That's the train of thought that the judges decided to apply to the definition of a work in the AI era. In Mio/konektra, the focus was on furniture: if a design's shape is partly determined by physical factors such as load and ergonomics, is it still copyrightable? CJEU thinks so, but only for those elements that result from the designer's free choice. After excluding those dictated by physical necessity.

Picture from @ninjason. Used under the Unsplash License.

It's similar in the Düsseldorf decision. Just substitute the legs of tables for all the training data. If models are internalising countless design patterns from their training, then the output of minimal high-level conceptual prompting must be determined by the training data, not by the „creative“ prompt itself. Right just as the presence of legs in tables is determined by gravity!

Following this paradigm, the judges needed to distinguish between what was created by a person and what was produced by an algorithm. As any copyright protection is inherently limited to the creative expression that remains after all non-human or deterministic influences are excluded. But since the defendant couldn't demonstrate a link [the burden of proof lies on the one who claims that the work is a work per (I-)ZR 16/24 of the Supreme Court of Germany], the judges concluded that it was a non-work.

Düsseldorf paradigm in a nutshell

What's unique to the Düsseldorf paradigm is its approach to the burden of proof. Which would, in my opinion, often lead to one party bearing most of the burden of proof in litigation. Because, to succeed in a possible civil lawsuit, the litigant would have to show that the key creative decisions directly caused a specific feature of the AI output (causa efficiens from Aristotle essentially).

If proven, the AI output would be recognised as the author's original work and receive copyright protection. If not, it would be considered a non-work, effectively placing it in the public domain. See the distinction between:

1. creating a „work“ by giving AI a conceptual (higher rows of the abstraction ladder) constructive framework for AI to work within using loose, minimal prompts that get refined step by step; and

2. creating an actual work by giving AI an instrumental (lower rows of the abstraction ladder) constructive framework for AI to work within that is in its essence spelled out using clear and detailed instructions right from the start.

As it's nearly impossible to link a particular human decision to a specific expressive feature in conceptual, high-level iterative prompting, such outputs (1) aren't protected under this paradigm.

Düsseldorf or Munich paradigm?

Reading my interpretative summary (subjective) of the Düsseldorf paradigm, one might wonder if it isn't just an extension of the Munich paradigm. As it seems to build on 142 C 9786/25, it offers just a more conceptual explanation for the same outcome. Although both courts arrive at similar inferences, I do believe that the paths they take to reach those conclusions are notably different.

Picture from @markusspiske. Used under the Unsplash License.

For me, the difference lies in how they treat AI-generated works. Munich's court sticks to a literal interpretation. Copyright is for people, so if a work isn't made by a person, it doesn't qualify. This has long excluded software-generated outputs. Düsseldorf, however, adds a key nuance. It considers how models use the works of others and requires the author to clearly reshape the defaults, relying on the Mio/konektra logic. While I don't agree with this approach because of the circular causality problem, I do have to admit that it's pretty clear and elegant. And also the reason why I'm naming it the Düsseldorf, not Munich, paradigm.

Circular causality problem

My reservations towards the Düsseldorf paradigm mostly come from its mistaken assumption that digital mirrors the rules of analogue. In digital, axioms, forms and boundaries are almost inherently fluid. Legal paradigms must account for this. In my opinion, the Düsseldorf paradigm fails to do so.

For example: consider how the roles of machines and humans are separated in the paradigm. Agents offer a bank of patterns. Humans add a dose of creative input. Legibility from James C. Scott, huh? It works for basic scenarios like image prompting, but it falls short in more complex AI use cases. Particularly with self-learning agents, which I see as the most telling example.

Wicked linearity

While the Düsseldorf paradigm can adapt the Mio/konektra ideas to digital, it remains dependent on linear input-to-output interaction patterns. This presents a problem for self-learning agents, which do not process inputs as a function processes arguments. They act and observe the consequences of their actions instead. Adjusting their internal rules based on those observations. And incorporating all those adjustments into their subsequent actions.

Every output immediately becomes the next input; in a recursive loop. As the cause-and-effect blurs, any imported Mio/konektra logic collapses here. This paradigm wasn't designed to deal with agents that effectively co-create instrumental constructive frameworks with their users.

Dependency Burden of proof hell

Since self-learning agents evolve through a combination of their internal processes (know-what and know-when) and user expertise (know-how), comprehensive logging has limited usefulness. Same goes for meeting the burden of proof by taking a snapshot of the agent at a particular time. As you'd also have to establish what happened at X±n before the AI output, proving that all your co-thought-through actions were actually your creative decisions. Not just your vague input being completed by the more or less distinctive design patterns of others.

And even if you somehow met the required burden of proof per (I-)ZR 16/24 or the German courts gave you the benefit of the doubt, you'd still be opening your agent's internals to frivolous or vexatious litigants who can use them to obtain various unfair advantages at your expense. From mimicking your workflows (economic advantage) to mapping your weak spots (security/intelligence advantage).

So where does that leave us?

Probably at Herrenstraße in Karlsruhe. But for the right reason this time! Düsseldorf paradigm isn't wrong to obsess over human creative choices. It's just wrong to think a judicial tool built for legs of tables can hold weight in a recursive digital loop.