When Machines Paint Like Masters: Can Pastiche Cover AI-Assisted Art?

Generative tools like DALL·E, Midjourney and Stable Diffusion can now produce an image "in the style of" almost any artist in seconds. That capability sits uneasily with EU copyright law, which has always assumed a human being on the other end of the pen. The pastiche exception in Article 5(3)(k) of the InfoSoc Directive was drafted for exactly this kind of stylistic borrowing -- but for a human borrower. The question this post asks is narrower and more tractable than it first appears: how much of the human operator's own creative contribution has to be present in an AI-assisted output before it can invoke pastiche at all.

Pelham II gives pastiche a spine

Until this year, pastiche was the least developed of the InfoSoc exceptions, with member states applying wildly different thresholds. The CJEU's Grand Chamber judgment in Pelham II (C-590/23) changed that. The Court held that pastiche is an autonomous EU law concept requiring a work that evokes an existing one, is perceptibly different from it, and enters into an objectively recognisable artistic or creative dialogue with it. Recognisability is assessed objectively, from the perspective of a person familiar with the earlier work, so courts no longer need to dig into the user's subjective state of mind.

The word “dialogue” is key. A dialogue presupposes an interlocutor -- someone capable of engaging with, selecting from, and positioning themselves against the work being evoked. A generative model does not do that. It produces an output by interpolating statistically across a learned distribution of training images. However convincing the result looks as homage or critique, there is no engagement with the source work in any sense a court could recognise as creative choice. This human-interlocutor requirement is not an external policy limit imported into pastiche to keep AI outputs out; it follows from the Court's own definition. If dialogue is what pastiche requires, and dialogue requires an interlocutor, then a work produced without one does not satisfy the Pelham II test on its own terms.

Why the camera analogy fails

The obvious counter-argument is photography, because it raises the same underlying question: whether output that a machine mechanically produces can still carry the human creative input that pastiche's dialogue requirement is looking for. Cameras automate a technical process, yet photographs can be pastiche. If camera-mediated output can be pastiche, why not AI-mediated output? The CJEU's Painer judgment (C-145/10) supplies the answer. A portrait photograph carries its photographer's free and creative choices about background, pose, angle and lighting; the camera then executes those choices deterministically. It is that chain, human choice determining output, that lets a photograph function as pastiche's interlocutor requires: the photographer, not the camera, is doing the positioning that Pelham II's ‘dialogue’ calls for. Generative AI is structured differently. The system does not execute a user's instructions deterministically; it produces an output through probabilistic recombination, so the artistic character of the result is not fully specified by the user's prompt but emerges substantially from patterns learned during training. Selecting among several AI-generated candidates is not the same act as a photographer choosing what to point the camera at and when to release the shutter: in one case the human generates every candidate, in the other the model does.

That does not close the door entirely. The generative step itself stays non-deterministic throughout; no amount of iteration changes that. What can shift is who is doing the choosing: where an operator can show a documented pattern of selecting specific source elements, discarding successive outputs, and refining prompts toward particular features across iterations, the resemblance becomes traceable to that sequence of human choices rather than to a single stochastic roll. At that point, the tool-use analogy starts to hold: the model's non-determinism supplies the raw material, but the operator's own selections are what fix the work's imitative character.

A two-condition test

Building on this, a two-condition test for when an AI-assisted output can qualify for pastiche can be proposed. This operates cumulatively:

• Process-level human dominance: the human operator's creative input -- not the model's algorithmic recombination -- must be the dominant causal factor shaping the work's imitative character.

• Output-level internal distance: the resulting work must show deliberate aesthetic positioning relative to the source, rather than mere statistical resemblance.

Where both conditions are met, the output is, in the relevant legal sense, not “AI-generated” at all but a human-authored work made with an AI tool. Because Pelham II's dialogue requirement is precisely what an AI-generated output lacks, satisfying the two conditions supplies the missing interlocutor, and the ordinary pastiche analysis, evocation, perceptible difference, objectively recognisable dialogue, can then run as it would for any human-made work. Where a generic, single-shot prompt produces the resemblance, neither condition is satisfied, there is no interlocutor to supply, and the exception does not engage, however striking the stylistic echo.

Architecture offers a useful stress test for this framework: because so much of a building's form is dictated by structural and planning constraints rather than free choice, it asks how much room for the operator's own aesthetic positioning the two-condition test actually needs, and what happens to the analysis when that room shrinks toward zero. Comparing Germany's broad section 51a UrhG with France's more restrictive, author-centred droit d'auteur tradition shows that neither system has actually confronted human-AI collaborative authorship -- both simply presuppose a human maker. The CJEU applies the same originality standard to applied art as to any other category of work; there is no separate filter for applied art. However, external structural or planning constraints may be more likely. Where those constraints leave no room for the architect to make free creative choices, there is no protectable expression, human- or AI-assisted, for the two-condition test to apply to, regardless of how much the architect steered the AI tool.

What the EU AI Act does, and does not, solve

It is tempting to think the AI Act resolves the evidentiary gap this test leaves open. It does, partly. Article 53 requires providers of general-purpose AI models to publish training-data summaries, which helps rightsholders exercise their text-and-data-mining opt-out under the DSM Directive, and Article 26 places responsibility for outputs on human deployers rather than on the system itself, reinforcing that any pastiche claim has to be asserted by a human, not a machine. But these provisions address the input side of the problem: whether material was lawfully ingested, and who is accountable downstream. They say nothing about the question on which pastiche actually turns, how much of the operator's own creative judgement shaped this particular output. That evidentiary gap is not a drafting oversight; the Act is a product-safety instrument, not a copyright one, and Recital 105 expressly leaves output-infringement questions to existing copyright law.

Practitioners see the same divide

A short qualitative study[i] conducted with ten artists and architects using generative tools reinforced this doctrinal picture. Across the board, respondents described AI as a collaborator that expands possibilities but does not replace human judgement, while raising real concerns about uncredited training data and the erosion of creative labour. One architect made a point that maps neatly onto the process-level condition: AI design tools are only as good as the parameters and knowledge the human designer feeds in. That is a good working definition of dominant human causal input, arrived at independently of the doctrine.

Where this leaves practice

For lawyers advising artists, galleries or AI developers, the practical takeaway is evidentiary rather than binary. “I used AI” is not, by itself, an answer to a pastiche question in either direction. What matters is whether an operator can document the iterative choices, which elements were selected, refined, rejected, and why, that would let a court find dominant human causal input and genuine aesthetic distance from the source. Absent that record, generic-prompt outputs sit outside the exception, however recognisable the stylistic echo. Given how the CJEU's objective-recognisability test in Pelham II already relieves courts of investigating subjective intent, that documentation is likely to become the main evidentiary battleground in the disputes this technology will keep generating.