Pastiche?
Regular readers will know that I have been interested in the pastiche exception for a while now, mostly because it has always struck me as the most underused exceptions under modern copyright law in Europe and the UK. Perhaps it is not surprising that parody gets all the attention, caricature gets a polite nod, and pastiche has tended to sit in the corner being ignored by everyone except a handful of German courts and the occasional bewildered UK judge trying to work out what Parliament meant when it copied and pasted the InfoSoc list into section 30A of the CDPA.
That may be about to change. The Grand Chamber has just handed down its judgment in the long-running and highly-awaited Pelham II case (C-590/23), and while the headlines will focus on the fact that Kraftwerk have, after roughly three decades of litigation, finally found themselves on the losing end of a pastiche argument, the more interesting story for those of us watching the AI copyright space is what the Court has done with the concept of pastiche itself, which may have substantial implications for the outputs of generative AI.
The Pelham saga
A quick recap for those who have not been following the case. The Metall auf Metall case began, to put it in proper perspective, before Napster was a thing. Kraftwerk founders Ralf Hütter and the late Florian Schneider-Esleben sued Moses Pelham and his colleagues over a two-second rhythm sample from the 1977 track Metall auf Metall that was looped throughout the 1997 song Nur mir. The case has been up and down the German court system more times than I care to count, been to the Federal Constitutional Court on fundamental rights grounds, and of course to the CJEU in 2019 in the original Pelham decision, which held that a sample, however short, could engage the phonogram producer’s reproduction right unless modified beyond recognition.
That would seem to have been game over for Pelham, but between the original CJEU reference and the current one, Germany transposed Article 5(3)(k) of the InfoSoc Directive by enacting a new section 51a UrhG on 7 June 2021, introducing explicit exceptions for caricature, parody, and pastiche. A court in Hamburg then held that for the period after 7 June 2021, Nur mir was lawful as a pastiche. Kraftwerk appealed, the case was referred to the CJEU again, and here we are.
The referred questions were, essentially: is pastiche a catch-all covering any artistic engagement with a pre-existing work, or is it subject to limiting criteria like humour, stylistic imitation, or tribute? And does pastiche require a subjective intention on the user’s part, or is it enough that the pastiche character is recognisable to someone familiar with the source?
The Pelham II decision
The CJEU’s answer to the first question is carefully balanced. Pastiche is not a catch-all. But it is also not limited to humour, stylistic imitation, or tribute as closed categories. Instead, the Court held that pastiche covers:
“…creations which evoke one or more existing works, while being noticeably different from them, and which use, including by means of sampling, some of those works’ characteristic elements protected by copyright, in order to engage with those works in an artistic or creative dialogue that is recognisable as such and that can take different forms, in particular the form of an overt stylistic imitation of those works, of a tribute to them or of humorous or critical engagement with them.”
This is interesting for many reasons. First, the Court rejects the catch-all reading that the German courts tentatively floated, on the sensible grounds that if pastiche covered every creative use of protected material it would swallow both parody and caricature and render the rest of Article 5 redundant. Fair enough. But it also rejects the restrictive reading that would limit pastiche to, say, stylistic imitation, because that too would undermine the effectiveness of the exception and would not achieve the fair balance between property rights under Article 17 of the Charter and freedom of the arts under Article 13.
What is left is a middle path defined by four elements: (1) evocation of an existing work; (2) noticeable difference from it; (3) use of characteristic, copyright-protected elements of that work; and (4) engagement in an artistic or creative dialogue that is recognisable as such. And crucially, on the second question, the Court rejected the subjective test: it does not matter whether the user intended to create a pastiche. What matters is whether the pastiche character is recognisable to someone familiar with the source. This is an objective test, grounded in legal certainty.
Pastiche, in other words, is not defined by what it excludes (plagiarism, concealed imitation) or by a closed list of permitted purposes, but by the relational quality of the new work to the old. If the new work is in dialogue with the old, and that dialogue is visible to someone who knows the source, then it is pastiche. The forms this can take are explicitly open-ended: stylistic imitation, tribute, humour, criticism, “in particular” but not exclusively.
The main difference with parody is that it for a work to be covered as such, it has to be humorous (see Deckmyn).
So what does this mean for AI outputs?
As I argued before, the output side of AI copyright is where things get genuinely difficult. The input question is mostly a TDM exception problem in Europe and a fair use problem in the US, and at least the rough shape of the answer is visible even if the details are being fought out in GEMA, Getty Images, Bartz, and assorted other current litigation. The output question is different. It turns on whether specific generated works infringe specific training works, and that is a question about substantial similarity, idea-expression, and exceptions and limitations.
I previously flagged pastiche as a potentially interesting defence for AI outputs, drawing on the Martin Eder case, which found that pastiche could cover the kind of remixing, mashing and referencing that characterises internet culture. I was quite cautious about it at the time, because the concept’s scope was genuinely unclear and there was a reasonable argument that pastiche required a conscious reference to the original, which an AI model plainly cannot have.
Pelham II has, I think, removed both of those objections.
On scope, the Court has given pastiche a definition that is broad enough to cover a substantial chunk of what generative AI actually does. Consider what a diffusion model like Stable Diffusion or Midjourney does when prompted for “a llama in the style of Gustav Klimt” or “Superman as painted by Caspar David Friedrich”. The output evokes existing works (both Klimt’s style and the general iconography of Superman), is noticeably different from any single one of them, uses characteristic elements (the gold leaf and ornamental patterning; the red cape and S-shield), and engages those works in a recognisable artistic dialogue. The dialogue is often explicitly laudatory or stylistic rather than critical or humorous, which is precisely the category of “non-parody laudatory outputs” that previously had nowhere to go under Article 5(3)(k). Parody requires humour or mockery per Deckmyn. Caricature requires exaggeration. Pastiche, post-Pelham II, covers the rest, provided the relational and recognisability (is that a word) criteria are met.
On intention, the Court’s objective test is even more significant. The referring court explicitly asked whether a pastiche finding required a subjective intention on the user’s part, and the Court said no: it is sufficient that the pastiche character is recognisable to someone familiar with the source. This matters enormously for AI, because the perennial objection to applying expressive exceptions to machine outputs is that machines do not have intentions. ChatGPT does not “intend” to parody Richard Brautigan when it generates a poem titled ‘All watched over by machines of mugging face’. Midjourney does not “intend” to pay tribute to Klimt when it outputs a golden llama. But under the objective test, it does not matter. What matters is whether the user looking at the output can recognise the dialogue with the source. If they can, the exception is engaged.
One could push this further. The user prompting the model, unlike the model itself, does have communicative intentions. When I type “a llama by Banksy” I am deliberately invoking Banksy’s visual vocabulary for some expressive purpose of my own, whether that is mere amusement or a more pointed reference. If we treat the user rather than the model as the relevant “author” for exception purposes (which is, for better or worse, the direction UK and EU case law has been heading), then the subjective and objective limbs of the Pelham II test can both be satisfied. The user intends the dialogue, and the dialogue is recognisable. That looks like textbook pastiche.
But there are limits of course.
I do not want to oversell this. Pelham II does not give AI outputs a blanket exemption, and there are at least four significant limits worth flagging.
Firstly, the Court was explicit that pastiche does not cover concealed imitations or plagiarism. The dialogue must be overt. An AI output that reproduces a training image closely enough to be mistaken for it, as can occasionally happen through memorisation, is not going to be saved by the pastiche exception. Memorisation is not dialogue.
Secondly, the elements used must be characteristic of the source. This cuts against the most ambitious reading of AI-as-pastiche, because a lot of what generative models produce draws on aggregated stylistic tendencies rather than recognisable characteristic elements of any single work. A generic “fantasy landscape” output is not in dialogue with anything in particular; it is in dialogue with the ambient aesthetic average of its training data. That is probably not enough.
Thirdly, Article 5(5) and the three-step test are still lurking in the background. Even an otherwise-qualifying pastiche must not conflict with normal exploitation of the work or unreasonably prejudice legitimate interests of the rightholder. Large-scale commercial AI generation of outputs that substitute for the original in its natural market is vulnerable on this ground, and the Grand Chamber’s silence on the three-step test in Pelham II should not be read as having abolished it.
And finally, and most obviously, the UK is no longer in the EU. Section 30A CDPA implements a pastiche exception in nearly identical terms, but UK courts are not bound by Pelham II, but I can see a world where UK courts, when presented with a pastiche case, may rely heavily on Pelham, it’s good law. Section 6(2) of the Retained EU Law (Revocation and Reform) Act 2023 permits consideration of post-Brexit CJEU decisions, and given how little domestic guidance we have on pastiche, I would expect UK courts to find Pelham II quite persuasive, particularly as the alternative is to leave section 30A without any meaningful content.
Concluding
The reason this decision matters for AI is not that it decides any AI question directly. It does not. It is a case about two seconds of rhythm sampled from one phonogram into another, decided by reference to a Charter balancing exercise. But the definitional move the Court made, from a closed-category understanding of pastiche to a relational and objective one, is exactly the move that a court would need to make to apply the exception sensibly to AI outputs. And the particular way the Court framed the test, as a dialogue that is recognisable to someone familiar with the source, maps onto the realities of AI generation in a way that the older humour-or-tribute framings did not.
European AI copyright litigation is going to reach the output question eventually. GEMA is still primarily an input case, but the substantial similarity questions in Getty Images and Disney v Midjourney are going to force courts to ask whether the outputs themselves infringe, and if so whether any exception applies. When that happens, Pelham II will be sitting there, waiting to be cited. The judgment does not solve the output question on its own. It does something more useful: it removes the conceptual obstacles to treating a substantial chunk of AI generation as pastiche.
Whether that convergence is good news depends on where you stand. Rightsholders will not love it, and I expect a round of commentary arguing that the Grand Chamber has quietly unpicked thirty years of strict construction and dressed fair use in a dirndl. AI developers will breathe a cautious sigh of relief. Users who have been mashing up cultural artefacts on the Web for two decades will note, correctly, that the law has finally caught up with what they have been doing all along.
The cat, as I am fond of saying, is well and truly out of the bag; the genie is out of the bottle, the bell has been rung, the arrow has left the bow… pick your cliché.
For now, I wonder what a llama by Klimt looks like, there’s only one way to find out.
0 Comments