[Guest post] Generative AI, originality, and the potential role of contract in protecting unoriginal works

The IPKat has received and is pleased to host the following guest contribution by Katfriends by Adrian Aronsson-Storrier and Oliver Fairhurst (both Lewis Silkin) on protectability of AI-generated outputs under UK copyright. Here’s what they write:

Generative AI, originality, and the potential role of contract in protecting unoriginal works

by Adrian Aronsson-Storrier and Oliver Fairhurst

Artificial Kat

Over the past two years the IPKat has hosted debate on the question of whether the outputs of generative AI tools are protected under copyright law. Since we last wrote for the IPKat on the topic, additional posts have discussed the Science, Innovation and Technology Select Committee’s inquiry into the impact of AI on the creative industries and the UK House of Lords Inquiry into Large Language Models. There have been reports on international developments, including on the US Copyright Office refusals to register AI-generated works [here and here], alongside developments in the EU, China, Czechia and France. Some authors have concluded that the autonomous outputs of AI systems are unprotectable under copyright law, while one Katfriend has gone so far as to wish for the death of copyright as we know it.

This post further explores the argument that the outputs of generative AI tools are unprotected in the UK, despite the language in section 9(3) of the Copyright, Designs and Patents Act 1988 UK (CDPA) relating to computer-generated works. In particular, it considers whether the reasoning in the recent Court of Appeal decision on copyright originality in THJ Systems v Sheridan (discussed previously on the IPKat here, and by us here) might imply that AI generated works are not protectable. The post then discusses the surprising consequence that following the CJEU’s Ryanair decision, if AI generated content is unprotected by copyright law, in some circumstances contract may provide more effective protection. In particular, where outputs are unprotected by copyright, there may be more scope for enforcing provisions in generative AI model licence agreements which restrict the use of outputs to train other models. Such ‘synthetic’ training data can be very valuable, as it can be used where real-world training data is scarce or where the use of real-world data is difficult due to privacy or security concerns.

Does UK copyright law protection extend to computer-generated works which are not original?

Ordinarily, for a literary, dramatic, musical or artistic work to be protected under the CDPA it must be original. Traditionally, the English Courts applied a test of ‘sufficient skill, labour or effort’ to determine originality. However, since the Court of Justice decision in Infopaq, the harmonized test of originality applied in the English Courts is now the “author’s own intellectual creation” for subject matter falling within the scope of the Information Society Directive.

As discussed in our previous post, there has been ambiguity and debate about whether the UK’s unusual provision on computer generated works, s 9(3), operates as an exception to that originality requirement. Section 9(3) of the CDPA provides that “In the case of a literary, dramatic, musical or artistic work which is computer-generated, the author shall be taken to be the person by whom the arrangements necessary for the creation of the work are undertaken”. “Computer generated” is defined in the Act as meaning “the work is generated by computer in circumstances such that there is no human author of the work”. Much of the debate about the operation of s 9(3) in the context of generative AI has concerned whether the user of an AI programme, its programmer, the person responsible for model training, or the company providing users with access to the programme is the person who has made the arrangements necessary for the creation of any resulting work. While some commentators have looked to see who has made creative contributions to the creation of a work to determine who made the “arrangements necessary”, case law interpreting similar statutory language relating to sound recordings or films has tended to instead centre on those financing or organising the activity necessary for its making.

In the 1988 debates in the House of Lords discussing the computer-generated work provisions, Lord Beaverbrook commented that the “the person by whom the arrangements necessary for the creation of a computer-generated work are undertaken will not himself have made any personal, creative effort”. This comment has been taken by academics including Dr Andres Guadamuz as suggesting s9(3) is an exception to the creativity and originality requirements for the subsistence of copyright. In contrast, Dr Patrick Goold has argued that the most plausible interpretation of the provisions is that the computer-generated works must be original to be protected, but that if computer-generated works are subject to an originality requirement, then 9(3) is meaningless and inherently contradictory. This, he says, is because if computer-generated works must be original to be protected, the person who supplies the necessary originality would be entitled to claim copyright in the normal manner, separately of the question of who made the ‘arrangements necessary’ for its creation.

There has been relatively limited UK case law directly considering section 9(3). Until recently, the leading decision addressing the provision was the High Court decision in Nova Productions v Mazooma Games. In that case it was argued that the composite visual frames of a coin-operated pool game called ‘Pocket Money’ were artistic works, but Kitchin J left open whether they were created by the relevant artist and programmer Mr Jones, or were created by the computer program which he wrote (paragraph 104). In either event Mr Jones would be considered the author, either in the usual manner as the artist or if it was computer-generated as the person who made the arrangements necessary for the creation of the work. Similarly, in dismissing the potential that a player of the video game may be an author of any of the artistic works in the composite frame images, Kitchin J denied that the player had either made contributions of an artistic kind, nor had they undertaken any of the arrangements necessary for the creation of the images (paragraph 106). While the High Court judgment was subsequently appealed, the Court of Appeal did not address section 9(3) or whether the composite frames were computer-generated. The High Court judgment did not therefore resolve the question of whether s 9(3) is an exception to the originality requirement.

A second occasion where the English Courts have considered 9(3) arose last year, in THJ Systems v Sheridan, a copyright dispute arising in the context of the breakdown in a business partnership. Here the High Court and Court of Appeal considered copyright protection and ownership of a graphic user interface for software used in analysing option trades.

As one of 29 approved issues to be determined at the first instance trial, the High Court was asked to determine who was the author of the graphic user interface “under section 9(3) of the 1988 Act”. The court discussed the Nova case, before concluding that the arrangement of tables and graphs making up the software graphical user interface “did involve the exercise of sufficient skill and labour for the result to amount to an artistic work”. The High Court did not however expressly analyse whether the works were computer-generated in the sense that there was no human author of the work, nor was there explicit analysis of who had made the arrangements necessary for the creation of the work, beyond the conclusion that the relevant software developer was the author (paragraph 216).

The High Court decision was appealed to the Court of Appeal, including on whether the High Court applied the correct test of originality when determining whether copyright subsisted in the graphical user interface. Arnold LJ held in his leading judgment that the High Court erred when it applied “skill and labour” as a test of originality, and that since the CJEU decision in Infopaq, copyright “is liable to apply only in relation to a subject-matter which is original in the sense that it is its author's own intellectual creation". The Court of Appeal went on to re-assess the originality of the graphic user interface (or more specifically, the “risk and price charts” identified in the claimant’s pleadings) applying the correct test of the “author’s own intellectual creation”, before concluding that they were original. The Court of Appeal did not however need to go on and explicitly consider the authorship of the risk and price charts, as the defendants’ conceded that if they were original then the relevant software developer was the author. As a consequence, the Court did not explicitly discuss the interpretation of s 9(3). Nor did the Court determine whether the programmer was the author of the work on the basis that he had made the arrangements necessary for the creation of a work where there was no human author, or whether he was the author in the ordinary sense of having supplied the necessary originality. In addition to the defendants’ concession on authorship, the Court of Appeal’s choice not to address s 9(3) is also explained by the parties’ decision not to address the issue in their submissions to the Court.

Despite its silence on s 9(3), the Court of Appeal decision could be interpreted to support the view that s 9(3) will rarely apply or may even be meaningless. The factors that the Court of Appeal considered when determining that the graphic user interface was original included the free and unconstrained choices of the programmer, such as the choice of fonts and colours and his choices as to the layout of user interface elements. These were human choices, suggesting that the programmer would have been entitled to copyright protection for the “risk and price charts” as the artist and author of an artistic work, had the pleadings and approved issues for the High Court trial had not framed the dispute through the lens of s 9(3). More fundamentally, the Court of Appeal’s approach of considering the originality of the work could be taken to suggest that s 9(3) does not operate as an exception to the originality requirement in the sense discussed by Dr Guadamuz.

Alternatively, this may be reading too much into the decision in circumstances where the Court of Appeal’s analysis was limited by the Defendants’ concession on authorship, and where s 9(3) was not explicitly addressed (as the claimant did not need to rely on section 9(3) in the circumstances). In a public event discussing the judgment, Arnold LJ was also reluctant to pre-empt the question of whether generative AI outputs were protectable in the UK, noting that it was an issue that was likely to be the subject of future litigation.

If computer-generated works do require originality to be protected, and where the correct test for originality requires an “author's own intellectual creation” involving the author being able to “express their creative abilities in the production of the work by making free and creative choices so as to stamp the work created with their personal touch”, it may be questioned whether much of the output of generative AI tools such as ChatGPT or Dalle-3 are protected in the UK.

While a user of such tools can express their creative abilities in writing an input prompt (which if sufficiently original may be protected itself as a literary work), the resulting image or text output from a generative AI tool may not reflect that user’s personal touch. This is because users may have no knowledge or control over the training corpus or the way the model operates, and even when using lengthy text prompts have somewhat limited control over the expression of the output. For example, consider the prompt and steps taken by the user whose AI generated image ‘The Spring Breeze Brings Tenderness’ was protected in a widely discussed Chinese case in November 2023. While the Chinese Court accepted that the user had used prompts to ‘set the parameters for the layout and composition of the image’, it is notable that in describing the image creation process, they gave examples (shown below) where changing the random seed used to generate an image in response to the same prompts created images with entirely different compositions. If this dispute was analysed under a UK or EU originality standard, it may be difficult for an English Court to conclude that the user had control over or was responsible for the type of free and creative choices giving rise to copyright originality in a portrait as identified in the Painer case, such as the “background, the subject’s pose and the lighting” or “the framing, the angle of view and the atmosphere created”. The user’s post-creation labour in selecting between a multitude of AI generated outputs may have been protected under the traditional “skill, labour and judgement” originality standard, but is less likely to be sufficiently original under the standard required by the Court of Appeal in THJ Systems.

The AI-generated images at issue in the Chinese case Lee v Liu
were produced with identical detailed prompts, but with different ‘random seeds’

Similarly, if originality is required, the developer or person responsible for training such AI tools may not be able to establish that they have expressed their creative abilities in the production of any specific generative AI output, as they may find it difficult to establish that they have stamped their personal touches upon all of the unlimited number of potential images that could be generated from the latent space within the model. This suggests that if originality is required, the output of generative AI tools may not always be protected by copyright in the UK despite s 9(3).

Are there alternatives if copyright protection is unavailable?

While contractual terms are often used to restrict behaviour of users in relation to copyright content, contract can also be used in relation to AI output that may not be protected by copyright. In some contexts, contractual protection for AI content will be preferable to copyright, as the scope for contractual restriction on user entitlements may be expanded where a work is not within the scope of copyright law.

For example, in the Ryanair case, the CJEU held that the Database Directive applies only to databases protected by copyright or the sui generis right. The Database Directive includes certain exceptions to the rights of a database owner in Articles 6(1) and 8, which are binding and cannot be set aside by contract. The Court of Justice held however that if a database is not protected by copyright or the sui generis right, then the owner of the database may use contract to prevent a user taking advantage of those exceptions. If this reasoning was followed more generally outside the context of databases, where content is unprotected by copyright but subject to restrictive contractual terms, even the s 29A ‘Copies for text and data analysis for non-commercial research’, s 30(1ZA) ‘Quotation’ and s 30A ‘Caricature, parody or pastiche’ copyright exceptions in the CDPA could not be relied upon by a user. This would be despite the explicit contractual override provision in each exception which provides that “to the extent that a term of a contract purports to prevent or restrict the making of a copy which, by virtue of this section, would not infringe copyright, that term is unenforceable”. This could have particular impacts on those seeking to undertake non-commercial text and data analysis in accordance with the UK exception, who could rely on s 29A to disregard contractual terms prohibiting computational analysis of a copyright protected work, but may breach contractual terms and conditions if the content is not protected by copyright (for example, if it was an unoriginal AI generated output).

In relation to AI outputs, the enforceability of restrictive contractual terms is an important issue. Many AI tools have terms and conditions which restrict the use of model outputs. For example, the Meta licence terms for the Llama 2 text model restricts the use of Llama 2 outputs to train any other large language model. Similarly, the Google Generative AI Additional Terms of Service (which apply to tools such as Gemini, a generative AI model for text and, controversially, images) prohibit the use of the Google generative AI tools to develop machine-learning models or related technology. These restrictions on using AI outputs exist to protect valuable markets for synthetic data for use as an input in the development of machine learning tools. If the AI outputs of the Llama 2 or Gemini models were protected by copyright under s 9(3) of the CDPA, then the restrictive contractual terms would be ineffective to prevent the use of the outputs for non-commercial research under the s 29A exception in the CDPA. In contrast, if as suggested above, generative AI outputs may be unprotected by copyright due to lacking originality, then these contractual restrictions will be effective to prevent the use of the outputs in developing machine learning models. As a consequence, AI model owners such as Meta and Google may prefer that AI outputs are unprotected under UK copyright law, as that may provide them with stronger contractual control over the use of outputs from their models and allow them to prevent the use of their models to create synthetic data for use in developing competing models.

In summary, there are interesting arguments that suggest that following the Court of Appeal decision in THJ Systems v Sheridan, the requirement for originality in the CDPA may mean that many AI outputs are unprotected under UK copyright law. In some circumstances however contractual control over these outputs can partially fill the void left in the potential absence of copyright protection, and in relation to protecting markets for synthetic data the owners of existing AI models may even prefer that their outputs are not protected by copyright.

[Guest post] Generative AI, originality, and the potential role of contract in protecting unoriginal works  [Guest post] Generative AI, originality, and the potential role of contract in protecting unoriginal works Reviewed by Eleonora Rosati on Thursday, April 25, 2024 Rating: 5

3 comments:

  1. An interesting post. Presumably the extent of contract protection would be narrow, depending on how users generating outputs can further communicate those outputs online (and accepting we're discussing UK law only as well)? The contract is with the system user, not a subsequent third party, so absent copyright protection the outputs - if placed online by a generating user - would then be fair game for anyone to copy and use in TDM training? Similar situation to public domain works behind paywall "liberated" by someone paying the fee to use (thinking along the lines of A History of the World According to Getty Images).

    ReplyDelete
  2. Hi AJS, thanks for the insightful comment (and for drawing my attention to 'A History of the World According to Getty Images' - I will have to go and watch it now!).

    You are absolutely right that the protection available under contract law is limited, given restrictions will only bind parties to the contract. There might however be ways of expanding a contractual restriction into a quasi-property entitlements (consider attempts to do so via viral open source or copyleft licences, or the possibility of embedding 'browsewrap' contractual restrictions into the metadata of an AI generated file).

    While individual AI generated works might be generated by users and placed online (and therefore downstream users might not be bound by a contractual restriction between the initial user and platform on use of those individual images for training), I suspect that the real value that Google and Meta will want to protect will be in licensing large scale synthetic training data - the creation of millions of images on demand for training the next generation of generative AI tools. Those large scale uses of generative AI tools are much more likely to involve situations where contract can bite.

    While this piece is only focused on the position under UK law, it is worth noting that the Ryanair case is from the CJEU, and similar issues may play out in EU member states (for an EU perspective on contract and the EU TDM exceptions where this is touched upon, see Kretschmer et al in Copyright Law and the Lifecycle of Machine Learning Models at section 4.4 - https://link.springer.com/article/10.1007/s40319-023-01419-3). I am still reading up on the US position, which might turn on the relevant Circuit's approach to the pre-emption doctrine in s 301(a) of the US Copyright Act (compare ProCD, Inc. v. Zeidenberg, 86 F.3d 1447 with Universal Instruments Corp. v. Micro Sys. Eng’g, Inc., 924 F.3d 32 - it is worth tracking down some of Guy A. Rub's scholarship on the issue).

    ReplyDelete
  3. I'm not sure the contractual restriction point is significant as it may at first appear. A lot of the big gen AI operators' terms & conditions provide that the user will own any copyright in the output, not the AI operator itself (though some operators have less commercial terms on this front).

    ReplyDelete

All comments must be moderated by a member of the IPKat team before they appear on the blog. Comments will not be allowed if the contravene the IPKat policy that readers' comments should not be obscene or defamatory; they should not consist of ad hominem attacks on members of the blog team or other comment-posters and they should make a constructive contribution to the discussion of the post on which they purport to comment.

It is also the IPKat policy that comments should not be made completely anonymously, and users should use a consistent name or pseudonym (which should not itself be defamatory or obscene, or that of another real person), either in the "identity" field, or at the beginning of the comment. Current practice is to, however, allow a limited number of comments that contravene this policy, provided that the comment has a high degree of relevance and the comment chain does not become too difficult to follow.

Learn more here: http://ipkitten.blogspot.com/p/want-to-complain.html

Powered by Blogger.