Copyright in AI-generated outputs

In a decision that has surprised absolutely no-one, the US Court of Appeals for the District of Columbia Circuit confirmed both the Copyright Office and Federal District Court’s decisions that Dr Stephen Thaler’s Creativity Machine (a general artificial intelligence model) could not be listed as the sole author of a copyright work entitled ‘A Recent Entrance to Paradise’ - shown below - on his application to register his copyright in the US (Thaler v Perlmutter).

A Recent Entrance to Paradise

The fundamental reason underpinning this conclusion was that, under the US federal Copyright Act 1976, a work must be authored by a human being which, of course, the Creativity Machine is not. By deciding the point under the Copyright Act 1976, the Court of Appeals also avoided the need to grapple with the argument advanced by the Copyright Office that the US Constitution itself requires human authorship. 

The decision merely reflects the position that the US has historically adopted that authorship means human authorship. Dr Thaler had disavowed any claim to human authorship (or at least “traditional” human authorship) in this case. He was testing the limits of the law. Ultimately, it was an easy conclusion for the court to reach, based on the pleaded case and the state of US law and practice. 

Leaving aside the normative approach, should a work generated solely by generative AI without any human element of authorship be protected? That is a far more complex question but if it is to be protected, copyright might not be the most appropriate form of protection. For example, the duration of copyright is typically calculated by reference to the life of the author. What is the ‘life’ of a generative AI model (what Dr Thaler refers to as a Creativity Machine)? 

One possible alternative is to introduce a new form of protection free from the traditional strictures of copyright doctrine – perhaps a copyright-lite right that is curtailed to reflect the relative ease with which such works can be generated (this is not my first ramble on this: see An agnostic observation on machine-originated works and copyright).

However, leaving aside for the moment the idea that a human being may play no creative role in the generation of an AI output, a more challenging question arises from the reality that humans routinely use generative AI in the creation/generation of works. Adobe’s Generative Fill is a good example, which can be used to enhance and embellish images.The conflation of human creativity with the generative power of AI makes it harder to work out whether the human input is sufficient to justify copyright protection under the normative approach. 

Human actors can also be involved an a number of ways in the creation of a work which involves the use of AI. These actors may include:

• Investors (i.e., those who invest in the development of the AI model or the specific project to produce the AI-generated work)
• Project managers (i.e., those behind the decision to deploy AI to generate a particular work or works and who set the objectives to achieve that)
• Software engineers (i.e., those responsible for producing the code which drives the generative AI model)
• Data selectors (i.e., the persons who decide what data is used to train the AI model and/or persons involved in Retrieval Augmented Generation)
• Prompters (i.e., those who devise and iterate the prompts to generate a particular output)

In the above examples, the prompters are probably the category of persons most proximate to the creation/generation of the work. Even with prompters, however, there are some challenges around whether creativity in the human-authored prompts is sufficiently evidenced in the associated AI-generated output. 

In its paper on ‘Copyright and Artificial Intelligence Part 2: Copyrightability’ (January 2025) the US Copyright Office concludes that prompts alone do not provide sufficient human control to make users of an AI system the author of the output and that prompts are essentially instructions that convey unprotectable ideas. The US Copyright Office does appear to accept that to the extent that the creative expression of a human-authored input (such as a sketch) is perceptible in an AI-generated output, it may be protected by copyright. It also appears to accept that where a human does have some control over the selection, arrangement and content of the final output (this could be where such a user is using Adobe Generative Fill, for example) then their creativity could be perceptible in the output. But prompts alone are not sufficient – not even the 600 plus prompts that generated Théâtre D’opéra Spatial. 

Théâtre D’opéra Spatial

Conversely, in Li v Liu the Beijing Internet Court granted copyright protection to a person who had devised the prompts that generated an image using Stable Diffusion. The Court recognised that the author in question had used a substantial number of prompts, set parameters and modified those parameters to elicit the desired image. The output was found to constitute the author’s intellectual achievement. The court put it like this: “as long as the AI-generated images can reflect people’s original intellectual investment, they should be recognized as works and protected by the Copyright Law”. 

The image in Li v Liu

Evidently there seems to be a difference between an output: (i) that reflects human intellectual investment (China); and (ii) in which a human input is perceptible (US). Under EU copyright law it is necessary to consider whether the output reflects the human’s personality and expresses his/her free and creative choices. These varying standards demonstrate the plurality of approaches to the copyright protection of AI-generated outputs. This is entirely unsurprising given that the standard of ‘originality’ (the benchmark for copyright protection) is not an international copyright norm under the Berne Convention, WIPO Copyright Treaty or the TRIPS Agreement, meaning that signatories remain largely free to determine the standard as they see fit.

The UK Intellectual Property Office is also considering copyright in AI-generated outputs as part of its consultation on Copyright & AI. It has noted that outputs will sometimes be co-creations of a human and a AI tool (referred to as AI-assisted works). In that case the human should be considered the author, provided that the creative essence of the work comes from the human. Where a work is generated without human authorship, it may presently be protected as a computer-generated work under UK copyright law. The author of a computer-generated work is deemed to be the person responsible for the arrangements necessary for its ‘creation’. However, the workability and benefits of this provision are uncertain and the government has expressed a preference to dispense with it (subject to any further evidence on the positive effects of that provision). 

AI-generated outputs present some big challenges for copyright law. For big challenges, we may need big thinking. A separate right for AI-generated outputs, mentioned above, would be a radical measure and one which would be very difficult to roll out widely. However, if there were such a right, which unlike copyright was not dependent upon intellectual creativity, it could go some way to eradicating any concern that AI-generated outputs ought to be protected and the challenges of ‘shoehorning’ creative human expression into those AI-generated outputs so that they are protected.