The IP in AI – Does copyright protect AI-generated works?

While many may have foreseen that computers and AI would assist in the production of useful things, the idea of AI being creative is much more difficult to grapple with and until recently seemed confined to fiction. Rapidly, however, generative AI systems such as ChatGPT, Bard, DALL-E, Midjourney and Stable Diffusion have burst onto the scene, allowing anyone to prompt an AI to produce literary, dramatic, and artistic works. AI systems have created paintings that reproduce the same features and style of well-known painters (such as "Comte de Bellamy", the first AI realised artwork sold by the auction house Christie's, and "Next Rembrandt", created by Microsoft in the style of paintings of Rembrandt). Even music has not been immune, with Dutch public broadcaster VPRO founding an annual AI song contest for “Eurovision-like” songs made using AI, as well as the April 2023 example of the AI-generated Drake and The Weeknd song – Heart on My Sleeve. Industrial and commercial applications are also growing, for example in cancer detection¹ and software development.²

As ever, those investing in systems to generate such works look to intellectual property laws to protect that investment. Traditionally, the patent system has been seen to protect useful inventions, while creative endeavours have been the field of copyright. However, the emergence of generative AI systems has raised the question whether AI-generated works can meet the requirements for copyright protection – and, if so, what the scope and ownership of that protection should be. The current answer to that question, besides the protection afforded in the UK to computer generated works, is likely to depend on whether a human “author” can be identified as having originated the AI-generated work.

Can AI-generated works be “original”?

In each of Australia, China, the EU and UK, a work must be “original” to attract copyright protection. Although there is no express provision requiring the creator be human, the requirement that a work be "original" has been understood as requiring a human’s expression of their free and creative choices, and not copied.

In Australia, while the requirement of originality is not an onerous one, it will be satisfied only if there has been some exertion of “independent intellectual effort” that is directed to the form that the work takes.³

In the UK, to be original, the work must result from the author’s free and creative choices and exhibit their “personal touch”. This is generally expressed in the same terms as the EU harmonised test of the work being the author's "own intellectual creation", which is not expressly stated in the EU law, but has been reiterated by relevant EU case-law,⁴ although there is some debate as to whether the UK and EU tests are exactly aligned.

In China, works in the legal sense of copyright must be "in the fields of literature, art, and science"; be "original"; be able to "be expressed in a certain form"; and be "intellectual achievements". Generally, an AI-generated work is similar in form to works created by natural persons, thus it is easy to judge whether it meets the constituent elements "in the fields of literature, art, and science" and "can be expressed in a certain form".  However, whether an AI-generated work meets the constituent elements of "original" and "intellectual achievements" is still controversial.

So, for example, Courts and Copyright Offices have found that:

• a phone directory, being the mere data output of a computer program without any human input, was not an original work (Australia);⁵
• a completely pirated publication, being a work that has merely been “slavishly copied”, has been denied copyright protection (Australia), although works derived from others’ works can be protected even if they infringe copyright in those earlier works (Australia and UK);⁶
• a selfie taken by a curious macaque was not an original work, since its form owed nothing to the intellectual effort of the camera’s owner (or any other human) (US);⁷
• an artwork autonomously created by a computer algorithm running on Creative Machine could not be considered an "original work of authorship", because authorship is an entirely human endeavour and non-human expression is ineligible for copyright protection (US);⁸
• articles created by AI tools developed by Tencent were copyright-protected works, although this judgment was made by a district court and not a guiding case of the Supreme People's Court (China);⁹
• a photograph of a static three-dimensional object¹⁰ (UK) and a portrait photograph¹¹ (EU) could both attract copyright, despite the need to be realistic images, due to the photographer's discretion in positioning, angle, lighting and focus;
• a newspaper headline could be an original work if there was sufficient creative input (EU),¹² although a purely factual newspaper headline does not have the requisite originality to attract copyright protection (Australia);¹³
• a functional shape cannot be a copyright work where it is dictated by technical considerations leaving no room for creative freedom (UK).¹⁴

Even if a work is sufficiently “original” so as to clear the hurdle for copyright protection to exist, the degree of that originality in turn affects the scope of the protection afforded to the work. As the House of Lords stated in an oft-quoted passage, “The reproduction of a part which by itself has no originality will not normally be a substantial part of the copyright and therefore will not be protected.”¹⁵ This may be particularly relevant where the content used to “train” the AI system remains discernible in the output produced by the AI system.

The effect of those decisions is that, for copyright protection to exist, it is likely to be necessary for a human to have at least exercised some control over the creative process that resulted in the work in question.

As a consequence, the relevant question is likely to be to what extent the form of the AI-generated work was a result of human intellectual effort, as opposed to the mechanical output of an (admittedly complex) computer program. This may require a case-by-case exploration of the creative process that led to the work’s creation. At one end of the spectrum are AI outputs generated from simple user prompts. At the other end are works generated by a process in which human authors worked in conjunction with an AI system, using dynamic prompts and iterative refinement.

Who is the “author” of an AI-generated work?

A further requirement for works to be copyright-protected is that they originate from one or more identified “authors”. The identification of the author or authors of a copyright work serves both formal and substantive purposes. As purely formal matters, it is necessary to establish a connection between the author and a relevant jurisdiction (namely a signatory to the Berne Convention) for copyright protection to be afforded, and it identifies the original owner of that copyright. More substantively, the identity of the author is inextricably linked to the assessment of originality. For this reason, the Australian High Court has remarked that to proceed without identifying the author can cause a court “to go awry”.16

In the context of AI-generated works, there are a number of possible solutions to the problem of identifying the author, including:

1. there is no author;
2. the author is the creator of the AI system, having enabled the creation of the work by building and training the AI system;
3. the author is the user of the AI system, having instructed the AI system to create the work;
4. some combination of (1)-(3) above.

The question of authorship is linked to the assessment of originality and creative input, as discussed above.

Some guidance may be drawn, by analogy, from the position with respect to other technologically-created works. For example, in Australia, copyright in a film or sound recording is originally owned by a “maker”, rather than an author, the maker for a film being the director, producer and/or the screenwriter of the film.¹⁷ A similar approach has been adopted in the UK expressly in relation to computer-generated works, which is discussed further below.

The EU approach is also very similar, despite the fact that in contrast to the UK position, EU law does not provide protection for computer-generated works. The strong interest in this subject is shown by the report "Trends and developments in artificial intelligence - Challenges to the intellectual property rights framework" published by the EU Commission.¹⁸ The Commission analysed the different types of players, i.e. developers, users, AI systems and/or co-authors to try to give some guidance of how to establish who is the author of the specific AI work.

The UK position: computer generated works

Conclusion

While the boundaries of copyright in, and in particular authorship of, AI generated works is not yet clear, what is apparent is that the question of what involvement a human had in the “creative” process will be critical. Even where AI system output is original and creative and therefore worthy of copyright protection under current legal frameworks, the key question of authorship, and therefore ownership, will depend on which human exercised the requisite creative input or intellectual effort. In this regard, the UK provision which deems the author of a computer-generated work to be the “person by whom the arrangements necessary for the creation of the [creative] work are undertaken” is a helpful but incomplete guide. It correctly draws focus to the question of creation but does not address who/what is the necessary creative link in that chain.

As AI develops, a clearer model of authorship may emerge. There are a number of policy grounds for this, including to provide economic incentives to create and maintain AI models/software.

In short, questions of authorship and originality remain the main barrier to copyright subsistence in AI-created output. Overcoming the issue of “whether our creations can create”, through legislative or other means, will be key in protecting such works.

For more on the developing area of intellectual property protection and risks for AI and ML systems, follow the AI in IP series on our IP blog.