How Metaphors Can Make or Break AI Copyright Cases

Metaphors are crucial to our (mis)understanding of generative AI (GenAI). For instance, AI is systematically conceptualized in human terms such as neural networks that learn, know and memorize. This blog briefly demonstrates how such metaphors resonate in AI copyright disputes and may affect their outcomes. It does so by addressing the ‘big three’ AI debates: training (learning), internal copies in an AI model (memorization) and copyright protection of AI-generated outputs (creation).

The learning debate

Are artificial neural networks allowed to learn from unauthorized works? In other words, can works be used freely to train AI models? The two most explicit judgments on commercial training of AI models were both rendered in the US.

In Anthropic v Bartz (see also here) Judge Alsup ruled that training AI models is, in principle, fair use. He elaborated as follows:

Authors cannot rightly exclude anyone from using their works for training or learning as such. Everyone reads texts, too, then writes new texts. They may need to pay for getting their hands on a text in the first instance. But to make anyone pay specifically for the use of a book each time they read it, each time they recall it from memory, each time they later draw upon it when writing new things in new ways would be unthinkable.

Therefore, AI training is transformative. ‘Like any reader aspiring to be a writer, Anthropic’s LLMs trained upon works not to race ahead and replicate or supplant them—but to turn a hard corner and create something different.’

So, in Alsup’s analysis, the first factor of the fair use test – the purpose and character of the use – favored Anthropic. And so did the fourth fair use factor, which refers to the effect of the use on the potential market. After all: ‘‘Authors’ complaint is no different than it would be if they complained that training schoolchildren to write well would result in an explosion of competing works. This is not the kind of competitive or creative displacement that concerns the Copyright Act.’

In conclusion, training GenAI is fair use.

Only two days later though, Judge Chhabria came to the opposite conclusion in Kadrey v Meta (see also here), namely that training a GenAI model is, in principle, not fair use. Interestingly, Chhabria motivated this by explicitly refuting the learning/teaching metaphor (and analogy):

But when it comes to market effects, using books to teach children to write is not remotely like using books to create a product that a single individual could employ to generate countless competing works with a miniscule fraction of the time and creativity it would otherwise take. This inapt analogy is not a basis for blowing off the most important factor in the fair use analysis.

In general, Chhabria deems it unlikely that unauthorized GenAI training is fair use. The fourth fair use factor is crucial in that assessment, as he argues that even substantially dissimilar output can lead to copyright-relevant market dilution (‘indirect substitution’). After all, training AI is nothing like teaching humans.

The memorization debate

Next is the second debate in which a mental metaphor plays a dominant role. Do AI models ‘memorize’ and can such memorization legally qualify as a copy inside a trained model?

Now what is ‘memorization’ to begin with? According to computer scientists, an AI model ‘memorizes’ when it ‘encodes details from its training data, such [that] it is capable of generating outputs that closely resemble its training data’ (Cooper et al.). That models ‘memorize’ in this technical sense has been uncontroversial among computer scientists for a while now (Carlini; Cooper et al.). It’s the legal evaluation, however, that sparks debate in copyright.

The difficulty is that the term ‘memorization’ takes on a life of its own outside this technical context. After all, for most people memorization isn’t defined as a technical AI phenomenon but as a cognitive term just like knowing and learning. And in that sense, it’s an anthropomorphic metaphor, which in legal reasoning may have the same effect as the learning metaphor. For instance, Guadamuz commented on the ‘memorization’ issue in Getty v. Stability as follows: ‘That is indeed the case, models can memorise, but as I have been reminding people endlessly, memorisation is not an exclusive right of the author (otherwise memorising a poem could land you in court). Reproduction is an exclusive right of the author’.

Such arguments are very similar to Alsup’s circular usage of the learning metaphor: first AI is anthropomorphized – it learns or memorizes - after which the metaphor circles back like a boomerang to reveal the absurd consequences of applying copyright to ‘memorization’ or ‘learning’. If learning and memorizing AI is a copyright-relevant activity, then we might as well sue schoolchildren for learning or memorizing a poem. In the process of such reasoning, humans and artificial neural networks are implicitly – and in my view, wrongly - equated with one another.

In the meantime, only one ‘memorization’ decision has been rendered so far: OpenAI v GEMA (Regional Court of Munich), which held that ‘memorized’ song texts qualify as copies in the AI model. In legal literature this case is often contrasted with Getty v Stability AI. However, this wasn’t an actual ‘memorization’ case like OpenAI v GEMA. Getty ultimately did not contend that any specific work was copied inside the trained models, it instead argued that merely making the model weights – had it occurred in the UK – was infringement (para 559-560).

The creation debate

Let’s move on to the ‘creative’ side of GenAI and copyright. The question is whether and to what extent AI-generated output is copyrightable. Most of the existing cases concern text-to-image systems like Midjourney (that use diffusion-based models). Are the resulting synthetic images protected? In general, Chinese courts are inclined to answer this question affirmatively because ‘in essence, it is still humans that use tools to create’ (Beijing Court 2023; see also here) . Briefly put, Chinese courts and various scholars across the world equate GenAI with other ‘tools’ such as cameras. And why should AI-generated output be denied copyright if photographs do enjoy protection? Because GenAI is equated with a (passive) tool, there’s no control or causality issue concerning the output of AI, which should be attributed to the user.

On the other hand, there are scholars claiming that GenAI isn’t a (passive) tool but rather functions autonomously (from a copyright perspective). Because of this, the human user merely controls the idea or theme but not the actual expression, rendering the output unprotected (Wang; Smit). This depiction of GenAI as autonomous closely aligns with the approach of the US Copyright Office. In the Zarya of the Dawn case, the user argued that she ‘authored every aspect of the work, with Midjourney serving merely as an assistive tool’. The USCO responded that ‘Rather than a tool that Ms. Kashtanova controlled and guided to reach her desired image, Midjourney generates images in an unpredictable way.’ The images are therefore made by the technology rather than its user, resulting in a denial of copyright for the individual images.

In other cases, such as Opéra Spatial, this line of reasoning was repeated. To illustrate the lack of control of the user, the Office also drew an analogy with Kelley v. Chicago Park District. In this case, the plaintiff – a designer of a large living garden – claimed copyright on the garden, but the court rejected this primarily because of nature’s indomitability and the concomitant lack of control over the garden by the plaintiff. These views were reiterated in the official USCO report on copyrightability.

In the first German decision on this matter, regarding three AI-generated logos, the judge stated that in order for AI output to be protected, the AI system should - ‘figuratively speaking’ - be more like an aiding instrument (tool) than something autonomous. In this case it wasn’t and copyright was denied. Moreover, the judge emphasized that mere manual craftsmanship – referring to prompting - does not reflect individual personality (for commentary, see here).

Final reflections

As is true for all three debates: the legal questions around the world are the same, copyright principles are similar, and they concern the same AI models and systems. And yet, legal analyses and judgments diverge significantly. To understand this fully, we need to grasp the importance of language. Metaphors in particular. The metaphorical terms in which we conceptualize GenAI precede legal analyses and can (silently) affect their outcome.

Moreover, metaphors aren’t just relevant to (short-term) litigation. The rhetoric in AI copyright debates, on a deeper philosophical level, reveals ‘our’ own human self-image and its importance to copyright law. The traditional copyright worldview is – briefly put – Romantic, in that it presumes that a human being has an individual personality, spirit or soul, that can be expressed autonomously in a work. This justifies copyright protection and the concomitant ‘elevated’ terminology such as ‘creation’. AI output, on the other hand, deserves no protection because AI lacks spirit/personality and merely ‘produces’ or ‘generates’ output through algorithms. However, this line of reasoning isn’t necessarily convincing for someone with a different worldview such as computationalism (the human mind is a computational system).

Also, for the better part of Western history, human ‘artists’ conceived their efforts in the very terms that judges and scholars now use to legally devalue AI (output). Artists or writers were mere tools or instruments in divine hands. Art was therefore not created but produced. Not by artists but by artisans, and as a result of (technical) craftsmanship instead of human creativity (see Buydens). Concomitantly, the idea of human authorship – let alone copyright – didn’t exist. This illustrates, once again, that metaphorical framing may turn the autonomy and causality analysis completely upside down, and with it – of course – legal evaluation.

This blog is a short adaptation of the article ‘Metaphors we judge (AI) by: a rhetorical analysis of artificial copyright disputes’ by MA Smit. Published in: Journal of Intellectual Property Law & Practice, 2026;, jpag018, https://doi.org/10.1093/jiplp/jpag018

Illustration by Public domain vectors on Unsplash