AI Regulation: ECRs’ Perspectives is a CREATe blog series featuring the work of early career researchers who are exploring the contemporary challenges of AI regulation. Based on research presented at the AI Regulation ECR Conference, each post provides a concise and accessible insight into emerging legal and policy debates around artificial intelligence.
The series continues with a blog post by Cody Rei-Anderson. Cody holds a PhD from Victoria University of Wellington and an LLM and JD from the University of British Columbia. He is Lecturer in Law and Technology at Edinburgh Napier University, where his research focuses on copyright law in the digital age, platform regulation, and the impacts of artificial intelligence on law and legal education.
Copyright protection of prompted generative AI outputs in the United Kingdom: A preliminary taxonomy
The copyright debate on generative AI and its potential effects on creative industries have tended to focus on the issue of inputs and training. But it seems fair to assume that regardless of how the training issue plays out, generative AI applications will continue to be developed for and adopted in creative industries. This inevitably leads to the question: how will copyright law treat works produced by or with the assistance of generative AI?
Early indications from other jurisdictions such as the United States and European Union are that AI-generated works will be treated as unauthored and not protectable. The UK, however, has a specific statutory provision which deals with authorship of works which are “computer generated”: section 9(3) of the Copyright, Designs and Patents Act 1988.[1] However, to date, no decisions have been released on how this section might apply to generative AI works and the section raises some troublesome questions of interpretation.[2] With little concrete evidence of the value of the provision provided in the course of the AI and copyright consultation, the UK Government’s report suggests that section 9(3) may be destined for repeal.[3]
The computer-generated works provision does not, however, affect the status of works correctly described as “AI-assisted”, as the consultation report affirms.[4] Whereas section 9(3) addresses works without an identifiable author, works which are “AI-assisted” imply a human user who is primarily responsible for their creation. The task is therefore to determine where this boundary lies.
Good old-fashioned human originality: When might copyright subsist in generative AI works?
The simple answer is that if the work reflects the “author’s own intellectual creation” it will be protected even if AI assistance was used in its production, with no need to rely on section 9(3) at all.[5] Works made with tools such as photo manipulation software which happen to have AI features built in are the usual reference point here: they function as a tool in the same way that non-AI software does. However, a large category of works which may include both “AI-assisted” and “AI-generated” works are those produced by generative AI using a “prompt”: a set of words (often framed as instructions), sometimes accompanied by an image or other media.
There have already been unsuccessful attempts to assert copyright in prompted AI-generated works where the ostensible author is the person who provided the prompt (and settings) to generate a work.[6] The issue is that, where generative AI is used to produce an output, the relevant elements of the expression are not generally subject to the “free, creative choices” of the human user. To take a simple example, if asked to produce a picture of a “blue bird”, the image produced may be superficially complex, but all of the “creative” choices taken in producing it are merely the product of a statistical inference process which the model undertakes. We might call this a case of a “simple prompt“, “simple” here meant in the specific sense that it does not convey enough expressive content for the resulting generated work to be original if it is the AI model doing the expressive work.[7] Section 9(3) might extend copyright protection to these works, but in its absence they would be considered unauthored and unoriginal, and therefore unprotectable.
The discussion should not stop here, however: there are other possible routes to copyright protection for prompted AI works. I will consider three.
Original hand-drawn sketch by Cody Rei-Anderson. Reference image: Wikipedia (Credit: Andy Reago & Chrissy McClarren. License: CC BY 2.0)
Made with Gemini 3.5 Flash with the prompt “please colorise the drawing, preserving the original detail, size, and appearance as a pencil sketch. don’t correct my mistakes and don’t over-interpret what I’ve drawn, just try to colorise the shapes as they are” and the original hand-drawn sketch. Note that the instructions have not been followed closely: the bird’s beak is parted and additional detail has been added on the wings, tail and eye.
While a work which produced by prompting an AI model may seem like the archetypal case of an unprotectable “AI-generated work”, we would over-generalise to say that any work produced in this way is prima facie uncopyrightable. For example, the very text you’re reading came out of a large language model provided with a prompt, yet it is clearly copyrightable. Why? Because an author (myself) conceived and wrote every word—the model was merely used to transcribe my handwriting. Whatever “original” content exists in this blog post was human-authored and the copyright protection came into being as I put the words on the page; AI was merely an effective tool to transform my handwritten text into digital text.
Clearly, where there is an underlying copyrightable work used as the basis for the “AI work” produced, the copyright protection in that original underlying work extends to the “AI output” insofar as elements originating with the human author are shared with the underlying work. (Any innovations contributed by the AI would be unprotectable.) Relatedly, there is nothing to stop copyright protecting a larger work of which the AI-generated work forms a part. For example, a book composed of (individually unprotectable) AI summaries could fall to be protected as a database if sufficient originality was used in its construction.[8]
Cases involving a copyrightable work separate from what the AI produces are the easiest to establish copyright protection in because the human creativity which goes into making them is clearly exercised outside the operation of the AI model. Yet if human creativity can survive being wrung through the AI inference process, by the same token a detailed prompt which specified sufficient creative choices in the expression might produce an output which would be original for the purposes of copyright—assuming that the model faithfully executed the human user’s instructions.
Case 2: Complex prompt
As discussed above, absent section 9(3) protection, there is little prospect for works created from simple prompts to receive protection. But a complex prompt might support a claim for authorship where the human author’s creative choices show through in the output and the AI model merely acts as the human’s “amanuensis”.[9]
Achieving originality this way may not be as easy as it sounds: there are many things which could be specified in a prompt which would not on their own constitute protectable expression. A prompt for an image like “A blue jay, perching on an oak tree in a verdant forest with light puffs of clouds in the sky, rendered in a realistic style recalling an oil painting” may be superficially detailed, but an image fitting that description could be expressed in many different ways, and the actual expressive choices would be made in the image’s execution: the simple combination of those elements would be unlikely to be protected.
However, if the user is able to specify in greater detail characteristics like the composition of the image, the lighting, or perhaps brush strokes or line weights, it is at least conceivable that the user’s inputs might rise to the level of making original creative choices which show through in the resultant image.[10] Yet a complex prompt like this would likely stumble over what I call the “bad amanuensis” problem: the more detailed instructions are, the less likely they will be followed faithfully by an AI model. I have proposed elsewhere that this analysis points to a two-step test for originality in AI outputs:[11]
- a sufficiently well-defined input (i.e., one which reflects the human user’s [own intellectual creation]), and
- the faithful execution of the human user’s instructions by the AI model.
This test sets a high standard for prompted AI outputs to be copyrightable where they are not based on an underlying work; it is possible that no currently existing generative AI models could reliably produce outputs which satisfy it. Perhaps it would be necessary therefore to engage more deeply with the process of generating the work to satisfy the requirements of copyrightability.
Case 3: Serial prompting
In the context of contemporary generative AI, the user’s interaction with an AI model need not be limited to a single, simple prompt. Repeated engagement with an AI model on the same work where the human operator exercises control over the creative choices through multiple rounds of interactions with the model could give rise to enough originality for a degree of protection—again, assuming their choices show through in the output.
Is this serial prompting really any different from the complex prompt case discussed above? Perhaps it would allow for more fine-grained control, and a kind of struggle with the limitations of the tool which might be analogous to artistic effort in traditional media. For example, an AI “co-written” book produced through a series of successive prompts to an LLM in which the human “author” guides the narrative and chooses how to develop characters could conceivably qualify as a literary work even if the AI-generated text lacked protection in its specific expression. Similarly, any subsequent human modifications to AI-generated works might qualify for copyright protection.
Conclusion
The repeal of section 9(3) CDPA would arguably produce uncertainty, since it would mean that the copyright status of AI works would fall to be determined on a case-by-case basis. This is scant reason to justify the continued protection of AI works that are not expressions of human creativity, however. Along with dispensing with some knotty interpretive problems, dispensing with the computer-generated works provision could be a good result for human creators—if one wanted to be certain of copyright protection, one would have to pay a human author.
Yet as the above discussion notes, the question of copyrightability for AI outputs is not quite so simple. Even if most would not qualify absent section 9(3), there is no specific bar to copyright protection subsisting in prompted AI outputs. Those based on an underlying copyright work are the easiest case, and should be easily identifiable since they rely on human creative labour outside the prompting of a model. The complex prompt case, on the other hand, seems narrow enough that AI works ought to be presumed to be uncopyrightable until the two-part amanuensis test is demonstrably satisfied. Properly applied, this test sets a very high bar for works produced through a “one-off” prompt to be considered original. In contrast, works produced through serial prompting might be more easily shown to have the necessary creative choices used in their production, particularly if the creator documents their process. Indeed, one takeaway from this analysis is that anyone who is intending to exercise their purported copyright interest in AI-generated/assisted works now should consider keeping good records to demonstrate the process by which a work was created. (This is probably good advice for human creators too, who increasingly face allegations of AI use.)
It is doubtful that withholding copyright protection from most AI-generated works will sufficiently cushion creative workers from AI disruption. Another avenue which deserves more exploration is labelling requirements for AI-generated and -assisted content. These have been discussed with respect to including machine-readable metadata on content, which are important to enable filtering and automatic labelling of works on online platforms.[12] But AI works are not only encountered digitally. Imposing requirements for human-readable disclosure through advertising standards, for example, could provide clarity about the copyright status of works as well as allowing consumers to make informed choices about the content they choose to watch, read, listen to, and support.[13] The partial taxonomy of different cases provided above may be helpful here, and the copyright analysis can ultimately help to demonstrate the amount of human creative labour involved in an “AI-assisted” work.
Footnotes
- A few other jurisdictions have equivalent provisions, including the interesting case of Ireland, which similar to the UK finds itself in tension with the European Union’s harmonised originality standard: see “Ireland at odds with EU on copyright and AI” Irish Legal News (24 February 2026), online: https://www.irishlegcom/articles/ireland-at-odds-with-eu-on-copyright-and-ai.
- See Patrick Goold, ‘The Curious Case of Computer-Generated Works under the Copyright, Designs and Patents Act 1988’ [2021] IPQ 120; “Copyright and AI Report” at 105-06. ↩
- Department for Science, Innovation and Technology, “Report on Copyright and Artificial Intelligence” (March 2026), online: https://www.guk/government/publications/report-and-impact-assessment-oncopyright-and-artificial-intelligence/report-on-copyright-and-artificial-intelligence [“Copyright and AI Report”] at 111. ↩
- “Copyright and AI Report” at 105. ↩
- Infopaq Int v Danske Dagblades Forening, Case C-5/08 [2009] ECR I–6569. ↩
- These decisions have been rendered by copyright offices and lower courts; I am not aware of any appellate decisions on the copyrightability of prompted works, but one will surely crop up soon. See AG München, Judgment of Feb 13, 2026 – 142 C 9786/25, online: https://www.gesetzebayde/Content/Document/Y-300-Z-BECKRS-B-2026-N-1513?hl=true; Andres Guadamuz, “No, the US Supreme Court did not declare that AI works cannot be copyrighted” (6 March 2026), online: https://www.technollama.co.uk/no-the-us-supreme-court-did-not-declare-that-ai-works-cannot-becopyrighted (noting that an appeal dealing with the copyrightability of the prompted artistic work “Opèra Spatiale” is currently pending in the United States). ↩
- An avenue for further exploration is whether the prompt itself might be an original literary work, in line with the short textual excerpts in Infopaq. ↩
- See e.g., Matt Bruenig, The National Labor Relations Book: A Labor Law Introduction (2026), online: https://www.nlrbedgcom/p/check-out-my-new-book. ↩
- See Jane C Ginsburg and Luke Ali Budiardjo “Authors and Machines” (2020) 34 Berkeley Technology Law Journal 343. ↩
- As was helpfully pointed out at the conference, this analysis should be expanded to media formats other than visual art. The use of AI in music production in particular is ripe for investigation. ↩
- Cody Rei-Anderson, “Protection of AI-Generated Images in Canadian Copyright Law: Charting a Narrow Path to Originality” in Graham J Reynolds, Alexandra Mogyoros and Teshager W Dagne, Intellectual Property Futures: Exploring the Global Landscape of IP Law and Policy (University of Ottawa Press, 2025) 99 at 114. ↩
- “Copyright and AI Report” at 72-74. ↩
- A salient point raised at the conference was whether we can really object to AI-generated works if they satisfy consumer preferences. Perhaps not, but it seems reasonable to insist that consumer choice be informed, and labelling requirements (which have been resisted by both technology firms and advertisers) are the most obvious way to inform consumers. A recent study (yet to undergo peer review) suggests that people display a strong aversion to AI-generated and -assisted works when they are labelled as such: see Graelin Mandel and Alex Imas, “Art and the Machine: Why People Devalue AI-Generated Creative Work” (SSRN), online https://papers.ssrn.com/sol3/papers.cfm?