First published in the European Intellectual Property Review (E.I.P.R. 2024, 46(7), 470-478).
There is nothing new about people creating content with the assistance of computers. What has changed in recent years is the extent to which computers are able to contribute to the creation process, and the amount of content that is being created in this way. Indeed, for many, 2023 was the year of generative artificial intelligence (AI). ChatGPT and other generative AI models became widely used with an ability to generate the type of content that was historically created by humans, including novels, poetry, scripts, music and images. And some AI-generated content has proven to be valuable e.g. the image of Edmond de Belamy, which sold at auction for U.S. USD432,000 in 2018.1
Generative AI technology is continuing to evolve rapidly with commensurate creative possibilities. Recent releases, such as Google DeepMind’s Gemini,2 are multi-modal rather than single-modal, meaning that they can receive and output a combination of text, code, image, audio and video. In addition, these models are now available on mobile devices,3 further integrating the technology into our everyday lives. It is against this backdrop that interest has been ignited in the copyright protection available for computer generated content.
Traditionally, copyright in a literary, dramatic, musical or artistic (LDMA) work is owned by the human author of that work. However, because AI systems are now undertaking many of the creative tasks typically reserved for human authors, AI developers and users want to know whether content generated by computers can be protected by copyright and, if so, who owns it.
The UK is one of the few countries in the world that has a specific legal framework for copyright in works that are generated by computers.4 When this mechanism was introduced in the Copyright, Designs and Patents Act 1988 (CDPA), it was hailed as “the first copyright legislation anywhere in the world which attempts to deal specifically with the advent of artificial intelligence”.5 The aim was to bring computer-generated works within the scope of UK copyright protection in order to “allow investment in artificial intelligence systems, in the future, to be made with confidence”.6 Whilst this was both laudable and progressive, the manner of implementation was problematic because the legislation didn’t deal with the fundamental originality requirement. It is also not particularly clear how this protection applies in the context of today’s autonomous AI- systems7 and who is deemed to own the copyright in AI-generated content. This article will analyse these difficulties and conclude with some possible improvements that would help to ensure that the UK retains a clear and coherent copyright framework for an AI-enabled world.
Copyright and the originality requirement
LDMA works must be “original” to qualify for copyright protection,8 which corresponds to the independent skill and labour that the author undertakes in creating the work. Historically this has meant that the author had to exert sufficient skill, labour and/ or judgment9 in creation of *E.I.P.R. 471 the work. Today, as the Court of Appeal recently confirmed,10 the UK follows the EU test of “author’s own intellectual creation”, which is often regarded as a higher standard because it introduces a creative element.11 Under this test, in order to constitute an original work, the content needs to reflect “human personality”, result from “free and creative choices” and the “author’s personal touch”,12 and not be dictated by technical considerations, rules or other constraints.13 Nevertheless, the bar for originality is still low and it has been envisioned that an 11-word snippet from a newspaper article could constitute a protectable copyright work.14
By contrast to LDMA works, entrepreneurial works (film, sound recordings, broadcasts and published editions) do not need to be original to qualify for UK copyright. Nevertheless, to prevent straight copies attracting their own protection, they cannot be copied from the same type of entrepreneurial work.15
Authorship
LDMA works
LDMA works reward and provide protection for people who create content. Generally, the first owner of copyright in any original LDMA work will be the “author”16 and this is the person who creates it.17 The author is understood to be the person from whom the protectable elements of the work originates i.e. the person who contributes the skill and labour of a literary, dramatic, musical or artistic nature. In many cases this person will be easy to identify because they will have written the book or play, composed the music, or painted or constructed the artistic work. Additionally, someone may be an author if they originate the expression of the ideas in the work, even though the actual words are written down by someone else (e.g. a person who dictates a letter to a secretary).18 Accordingly, authors can include those who expend skill and effort in creating, selecting, or gathering detailed concepts, data or emotions19 if that person is responsible for the intellectual creativity – the creative choices. On the other hand, someone who merely takes down word for word the text of a work dictated by someone else (a “mere amanuensis”) cannot be an author because they do not play a part in originating the expression of the ideas or information contained in the work20 .
Historically, it was said that an author of an LDMA work had to contribute “skill and labour of the right kind” (reflecting the traditional British test of originality)21 . Accordingly, a reporter taking a shorthand report of another’s speech was held to be the author of the report because they needed to be skilled in shorthand.22 It is, however, debatable whether this would still be the case today if the reporter doesn’t make sufficient free and creative choices to be considered an author.23 In this sense, the framing of the EU test is helpful: sufficient originality results from the author’s own intellectual creation and a person who does not contribute sufficient originality to a work cannot be an author of that work. Or in other words the work must “originate” from an author who must undertake a certain amount of creative labour.
The author’s original contribution also needs to go to the expression of the LDMA work and not the idea behind it. Accordingly, where a person communicates only a general concept or idea to a writer (e.g. the broad outline of a plot) and the writer clothes the idea in the form of a novel, the copyright will lie with the writer.24 Conversely, a person may be an author25 if they contribute a series of ideas or concepts that are sufficiently detailed, well defined and original, such that they provide the intellectual creativity i.e. the expressive and creative choices made in producing the work and making it “original”.26
Entrepreneurial works
The first owner of any copyright in an entrepreneurial work is also the “author”27 and, again, this is the person who creates it. The CDPA designates the author for each type of work.28 Most relevant for these purposes is the author of a sound recording who is “the producer”,29 defined as the “person by whom the arrangements *E.I.P.R. 472 necessary for the making of [the sound recording] are undertaken.”30 Similarly, when the CDPA was first enacted, the author of a film was also “the person by whom the arrangements necessary for the making of the [film] are undertaken”.31 Consequently, entrepreneurial works provide protection for the people who invest in the content, not necessarily those who contribute creatively to it.
Works that are generated by computers: CDPA section 9(3)
CDPA s.9(3) also allocates authorship (and therefore first ownership) to a specific class of works: LDMA works which are generated by computer “in circumstances such that there is no human author of the work”.32 As there is no human author, s.9(3) states that the author of these works “shall be taken to be the person by whom the arrangements necessary for the creation of the work are undertaken”33. In these circumstances, the copyright expires 50 years from the end of the calendar year in which the work was made34 and the author does not obtain any moral rights.35
Unfortunately, there is an inherent problem in the mechanism set out s.9(3). This is because it covers LDMA works, which are computer-generated without a human author. Yet, an LDMA work, also known as an authorial work, needs to be original to be protectable in the first place, which generally requires a human to provide the necessary skill and labour. The problem appears to be worse now than in 1988 because originality now requires the work to be its “author’s own intellectual creation” to stamp the work created with their “personal touch”, which is obviously challenging when there is no human author. Indeed, some have argued that it is doubtful whether copyright protection (in a European sense) can be available to any computer- generated work.36 The Advocate General in Painer37 took this view, noting that only human creations can be copyright- protected (although the human can employ a “technical aid” like a camera). A similar position has also been taken by the U.S. Copyright Office, which determined that images created using the generative AI model, Midjourney, were not original works of authorship protected by U.S. copyright law because this excludes works produced by non-humans.38 Caselaw from other countries also reflects this understanding.39
Such restrictive approaches cannot be correct as a matter of UK law. Section 9(3) presupposes that there can be protectable works generated by computers with no human author because it provides a mechanism for designating an author for them. The section would be entirely redundant40 if UK copyright protection were restricted to LDMA works that have a human author. Instead, it was the clear intention of the legislature to encourage AI innovation by going further than other jurisdictions, to protect LDMA works generated where it would not be possible to ascribe human authorship,41 and to allocate first ownership to the person who undertook the necessary arrangements to bring the work into existence. Unfortunately, however, the UK legislation doesn’t clarify how the originality of works protected under s.9(3) is to be determined.42 It was noted in the 1981 Green Paper on the Reform of Copyright Protection43 that the general proviso for copyright subsistence of sufficient skill and labour still applied to works created with the aid of a computer but there was no explanation of how this could work in practice. As a result, commentators have canvassed several pragmatic interpretations to try to reconcile the requirement for originality with the aim and purpose of s.9(3).
Guadamuz argues that s.9(3) has been framed as an exception to the originality requirement for the subsistence of copyright44 in these types of works. Alternatively, Bently suggests that there should be a different originality test for these types of works e.g. they are original if they are “novel” i.e. different from previous works or “not copied” i.e. produced as a result of the independent acts of the computer.45 Copinger posits that “the relevant skill and labour is that of the person by whom the arrangements necessary for the creation of the work were undertaken”46 and this finds support in the *E.I.P.R. 473 1986 White Paper on IP and Innovation, which states that the “question of authorship of works created with the aid of a computer will therefore be decided as for other categories of copyright work, i.e. on the basis of who, if anyone, has provided the essential skill and labour in the creation of the work”.47
Laddie, Prescott and Vitoria48 goes further and suggests a theoretical presumption that protection is granted to computer- generated LDMA works that would be original had they been created by a human, i.e. if you would need a human’s intellectual effort in order to create them. We consider this to be the most helpful suggestion. It incorporates a de minimis threshold by excluding extremely simple or banal content (e.g. a simple line drawing of a triangle) or a purely functional creation but provides protection where skill and labour goes into making the arrangements for the creation of the work.
It is unfortunate that there is no express reconciliation of the originality requirement for these types of works in the legislation. Because s.9(3) is expressly stated to apply to computer-generated LDMA works, it would be preferable for the legislation to also set out expressly the requisite originality threshold and how it may differ from other LDMA works. This is certainly an area that would benefit from further clarity in the future. Nevertheless, we can draw an important distinction with how LDMA works under the CDPA are treated differently depending on how they are created. Where a human contributes sufficiently to a LDMA work generated by a computer such that the work can be said to be a result of their free and creative choices, there will be an original (normal) LDMA work and the human will be the author because they have originated the expression of the ideas and “created” the work. They obtain full moral rights, and the copyright expires at the end of the period of 70 years from the end of the calendar year in which the author dies.49 In contrast, it is only in limited circumstances when s.9(3) will apply: where a computer has generated a work and it is not possible to identify a human who has contributed sufficient skill and labour of a literary, dramatic, musical or artistic nature to the work itself. Originality is presumed if the work would be original had it been created by a human, and s.9(3) designates authorship to the person who undertakes the necessary arrangement to create the work. In that case, the copyright lasts only 50 years and there are no moral rights.
It follows that, when considering the copyright protection of works generated by today’s AI systems, there are three key considerations: Is it possible to identify a human author of AI-generated content? In which circumstances does s.9(3) apply? And who is the person who undertook the necessary arrangements to create the work?
Is there a human author of AI-generated works?
In Express Newspapers v Liverpool Daily Post, 50 the Court held that grids of 5-letter sequences were protected by copyright as literary works (a table or compilation), despite being generated by a computer. The case was decided before the implementation of s.9(3), but a readiness to find a human author51 meant that the court was able to provide protection. The preparation of the sequences involved the human programmer’s skill and labour, and the computer was “no more than a tool”, which produced the sequences using the programmer’s instructions. Whitford J specifically compared this situation to a pen being used to write a work: “It is unrealistic to suggest that it is the pen which is the author of the work rather than the person who drives the pen”
In Express Newspapers there was a clear nexus between the programmer’s work and his instructions to the computer to create the work. This nexus also existed in many traditional computer systems when the computer programmer played a significant role in devising the logic and rules by which program outputs were generated and/ or exercised significant control over the final output in early “rules-based” AI systems.52 In contrast, today’s complex AI systems can create content using their own training data, logic and learning so the programmer generally has less direction and control over the output. As was explained by the High Court:53
“The ANN [Artificial Neural Network] has learned... It has not done so because any human (programmer) has told it how to do it. It has done it by producing results, being provided with information reflecting its degree of error, adjusting its own internal assessment parameters, reprocessing the files to reduce the error and repeating this process until it gets it sufficiently right sufficiently often.”
Furthermore, expert evidence from the same case explained how:54
“Machine learning eliminates the need to define complex hand-crafted rules that strictly follow a defined specification written by the programmer [as occurs in the development of computer programs] since the abstract machine in ML/AI technology is not processing data on a step-by-step instructional *E.I.P.R. 474 basis, but instead uses training data to learn the logic to solve a specific problem and thereby reconfigures the machine.”
It should be noted that generative-AI includes not only DALLE, ChatGPT, Midjourney, Stable Diffusion and other models with broad-ranging applications, but also systems that deploy machine learning to create limited types of content in response to a user’s input (i.e. systems of “narrow application”). Examples are video games like No Man’s Sky, which generate new virtual environments based on the user’s actions55 or virtual assistants like the Expedia chatbot, which creates trip itineraries based on user inputs.56 All of this combines to complicate the question of whether the programmer’s skill and effort originates the protected elements of today’s AI-generated content.
Developer
Whilst the link between the programmer and the output is less significant in today’s AI systems, the AI developer is still likely to have programmed the features of the model prior to training (sometimes called hyperparameters), to have chosen the training data, fine-tuned the model, prescribed any rules to which the system should adhere in providing output (sometimes called meta- prompts or system prompts) — all of which can have an impact on the expression of the output. At the outset, therefore, it appears that AI developers perform a broad range of acts in the creative process and therefore have many opportunities to contribute towards the original expression of any final output.
This is more likely in a generative AI model of narrow application because the programmer is likely to define more aspects of the output and, in doing so, make individual choices as to the expression that are not entirely functional. This will be particularly so, where the program provides users with a particular experience or an ability to create a particular type of output.
Even in generative AI models of broad application, there is a theoretical possibility that the AI developer might provide sufficient creativity in the final output through the selection of the training data. For example, an image generator that is trained using images of only one breed of dog (e.g. a golden retriever), will necessarily generate images of golden retrievers when generating “dogs” regardless of the user prompt. In this way, the developer could be said to have determined how the user’s prompt will be expressed in the output, at least to some extent. However, on balance, it seems unlikely that this will contribute a sufficient level of creative skill and labour towards the protected work to constitute authorship. Accordingly, the Beijing Internet Court57 has recently determined that the producer of Stable Diffusion, an AI model of broad application, was not involved in the generation process of the images so could not be the author. Furthermore, where the training data is particularly limited, the authors of the initial images (e.g. of the golden retrievers) would arguably have a stronger claim to authorship because the output work is more likely to be an expression of the creative efforts of one or more of those authors. In any event, it is unlikely that this hypothetical scenario can be extrapolated to broader and more diverse datasets. Even though the developer can correct biases caused by training data through fine-tuning and reinforcement learning,58 this will only change the output to a limited degree.59
Overall, from a practical perspective, if the developer wishes to be considered as the author of works generated by its systems, it should choose to define as many of the aspects of the output as possible so that it is considered to have created or originated the expression of the work. Moreover, if it wishes to avoid the user claiming authorship or joint authorship, it should limit its users’ freedom to make creative choices in the final output (see further below). However, this will not always be sufficient and there will be others who may have competing and stronger claims to authorship.
User
The user of a generative-AI system instructs the system to produce content by way of input prompts and parameters (that can dictate how closely the output needs to match the input),60 which can be a significant way of influencing and originating the output of a generative AI model (at least where creative freedom is permitted by the developer).
However, in the context of a generative AI model of broad application, there is a large risk that the user will only contribute a general idea of what is expected and not contribute sufficiently to the expression of the work to be considered its creator. A user who gives a prompt that is insufficiently detailed for its desired output (e.g. “write me a short story about a fairy”), is very similar to the person who only provides a broad outline of a plot. They cannot be the author because the AI is left to determine the expression of the idea almost entirely. Indeed, there is an argument that a prompt can only ever contribute an idea. If 10 painters are asked to produce a painting of “a sheep on a hill eating a daisy”, each painter is likely to produce a different painting and each painter will own the copyright because they have determined how to express the general idea. The person who suggests *E.I.P.R. 475 the idea (the equivalent of an AI prompt) can only own the copyright if the idea/prompt is very specific about defined aspects such as layout, composition and style such that the painter (or the AI model) becomes, in effect, an extension of the prompt-giver’s hand.61
The U.S. Copyright Office has taken a strict approach to this in refusing to protect output from the AI model, Midjourney, even when the human user has undergone a creative, iterative process, inputting “hundreds or thousands of descriptive prompts” until these “iterations [created] as perfect a rendition of her vision as possible”.62 The office did not think that the user could “actually form” the generated images or be the “master mind” behind them because of the significant distance between what the user may direct Midjourney to create and the visual material actually produced.63 A key factor was Midjourney’s unpredictability: the user prompt may influence the generated image, but it cannot dictate a specific result, so the multiple prompts are a process of trial-and-error, rather than control and guidance to reach a desired image. This implies that there can only ever be a human author where the user has complete control over the final output, which may be very difficult to achieve in practice because AI users don’t generally control how the system will interpret a prompt and generate any content. Instead, the office treated the prompt more like instructions to a commissioned artist64 so the prompter (or commissioner) only identifies what it wishes to have depicted, which is an idea and therefore not copyright protected, and it is the machine (or commissioned artist) that determines how those instructions are implemented in its output. In this respect, the office appears to have equated Midjourney prompts with repeatedly hitting a ‘randomise’ button until the user gets something that it wants. While Midjourney clearly has an element of unpredictability, this does appear to disregard the level of influence that a user does have in determining the output of the works, including as to expression, and also the unpredictability that forms a part of some human-authored content. For example, Jackson Pollock is able to protect his works despite this.65
The Beijing Internet Court66 has taken a different approach to this issue in deciding that certain images created using the generative AI model, Stable Diffusion, were protectable by copyright because they reflected the original intellectual investment or personal expression of the human user. The court confirmed that each situation needs to be assessed on a case-by-case basis but this is more likely to occur if the user’s descriptions of the elements in the image and layout composition are clear and specific. Relevant factors in that case included that:
- The image was based on the user’s text description input—the AI replaced the human in drawing lines and colors but presented the user’s creativity and ideas in a tangible way;
- The user had provided certain intellectual input, g. choosing the prompts, arranging the order of the prompts and setting the relevant parameters. In doing so, the user had designed the figure and its presentation style and set the parameters for the layout and composition of the image so these reflected the user’s choice and arrangement; and
- The user’s iterative process of amending and adjusting prompts and parameters to finally obtain an image also reflected the user’s aesthetic choice and personalized judgment.
Most recently, the Municipal Court in Prague67 has held that an image created by Open AI’s Dall-E was not protectable by copyright. The simple prompt inserted by the user in that case was “Create a visual representation of two parties signing a business contract in a formal setting, such as a conference room or a law firm office in Prague. Show only hands.” The claimant failed to prove that they were the author of the image, and the court also observed that the image was not the result of the creative activity of a natural person (i.e. the user).68
It remains to be seen how the UK will approach this question in the context of generative-AI. In Nova v Mazooma,69 it was held that the user of a computer game that simulated a game of pool wasn’t the author of any artistic work in the successive frame images. Whilst the appearance of a particular screen depended to some extent on the way the user played the game (e.g. when the rotary knob was turned the cue rotated around the cue ball), the player wasn’t an author because they hadn’t contributed any input or skill and labour of an artistic nature. All they had done was play the game.
It is therefore unlikely that a user will be considered to be the author of any original work if they only play a game or use an AI-system of narrow application, where their influence over content is constrained by the choices and environment provided by the developer. The output will be significantly dictated by the programmer’s hyperparameters, meta-prompts or system prompts, and the user is unlikely to contribute sufficient skill and labour towards any protectable elements. Conversely, in the context of a more general AI application, much will depend on the particular circumstances, but a detailed and defined prompt stands a better chance of contributing *E.I.P.R. 476 sufficiently to the expression of the output to substantiate authorship. This is analogous to the author who dictates a letter to a secretary70 or provides a detailed and well-defined plot of a novel71 or the person who only uses the computer as a “technical aid”72 to record their creative expression. Furthermore, models that output the same mode of content that they ingest, e.g. text-to-text or picture-to-picture, are likely to use at least some of the user input in the output so it will be easier to find sufficient user originality.73
Joint authorship
Where a number of humans each contribute sufficiently to a LDMA work generated by a computer, there may be claims to joint authorship. However, this requires two or more individuals to work together to create a piece of work such that their contributions are inseparable and indistinct from one another74. This may be difficult to satisfy in relation to most cases of AI-generated content because there is unlikely to be a true collaboration between the developer and user to create content pursuant to a common design.
When does section 9(3) apply?
If, as suggested above, the user or developer contributes originality to an AI-generated work, they will be the author of an original (normal) LDMA work and s.9(3) will not apply. And if the courts continue to show the same readiness to find a human author as in Express Newspapers,75 s.9(3) may, somewhat surprisingly, end up being of limited application to AI-generated works. On the other hand, as AI continues to develop, and its autonomy continues to increase, it may become more common for the developer to have only contributed functional coding and the user to have only prompted the system with a general idea, neither of which provide sufficient skill and labour towards the protectable elements of the work. In those cases, the AI will have created the work without any original contribution from the developer or user and there will be no human author. Nevertheless, the work will be protectable if it would be original had it been created by a human, (i.e. you would need a human’s intellectual effort to create it) and s.9(3) will allocate authorship. There can be no claim to joint authorship - the sole author is the person (or company)76 who undertook the arrangements necessary.
Identifying the author under section 9(3)
Both the 1977 Whitford Report on Copyright and Designs Law77 and the 1981 Green Paper on Reform of Copyright Protection78 investigated the possibility of a general answer to the question of who should be regarded as the author of a work created with the aid of a computer. Three candidates, alone or in combination, were considered: the creator of the program that controls the computer, the originator of the data upon which the computer operates to create the new work, and the person responsible for running the computer to produce the work. However, in the following White Paper,79 the government concluded that no specific provisions should be made because the circumstances can vary so much, and the question of authorship (and therefore first ownership) needs to be determined on a case-by-case basis in order to avoid unfairness. The question is who, if anyone, has provided the essential skill and labour in the creation of the work.
What is clear is that assessing the identity of the person who undertook the arrangements necessary is a different test to that of human authorship. As was stated during the legislative progress of the CDPA: “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.”80 As mentioned above, if they do contribute sufficient creative effort, they would be an author themselves and s.9(3) would not apply.
To date, the protection for computer-generated LDMA works has been rarely relied upon81 and there is very little court guidance on the identity of “the person by whom the arrangements necessary for the creation of the work are undertaken” under s.9(3). We are aware of only one decision that expressly applies s.9(3): Nova v Mazooma Games.82 The judge held that, in so far as the composite frames were computer generated works, then the arrangements necessary for the creation of the work were undertaken by the programmer because he devised the appearance of the various elements of the game and the rules and logic by which each frame was generated, and he wrote the relevant computer program. As such he was deemed to be the author by virtue of s.9(3). The user had not undertaken any of the arrangements necessary for the creation of the frame images because he had only played the game.*E.I.P.R. 477
The court didn’t think it was necessary to decide whether the programmer originated the artistic skill or labour essential to the creation of the frames (in which case they were a human-authored LDMA) or had made the necessary arrangements in instructing the computer to create them (in which case they were a computer-generated LDMA). Either way, the rights were owned by the programmer. There will, however, be situations where this distinction will be significant, for example, it will affect the term of the copyright and whether moral rights apply to the work.
Whilst Nova should not necessarily be considered prescriptive in relation to today’s technology, it does show that the factors considered by the court (who wrote the computer program, who devised the appearance of the work and the rules and logic by which the copyright work was generated) all point -towards the programmer being the author. Consequently, some commentators have stated that the author of works generated by AI will “normally be the computer programmer or software engineer who created the algorithm(s) which in turn created the work”.83 However, as set out in the White Paper84, this isn’t inevitable and others have stated that it would be inconvenient and misleading to treat the programmer as the owner of copyright in all works produced by these systems. Accordingly, they argue that the author will “normally be the operator or the person directing the operation of the machine”85 (i.e. the user). Others agree and state that, “if a user acquires a program capable of producing computer-generated works, and uses it to generate a new work, then ownership would go to the user.”86 However, this was not the case in Nova: even though the user owned and powered the computer and played the game that generated the artistic works, they were not entitled to own them because they hadn’t made the necessary arrangements for their creation.
For completeness, we should note that there is some further (albeit limited) guidance that can be drawn from the case law regarding entrepreneurial works where the designated author is also “the person by whom the arrangements necessary for the [creation/making] of the work are undertaken”. The idea in these cases is to identify the “moving force” who initiated the making of the work i.e. the person who instigated it and organized the activities necessary for its making, the person without which there would never have been a work87. Being the person “closest” to the actual making of the work or the one who “presses the button”88 does not qualify someone to be the author. On the other hand, the designated author should not remain too removed from the creation of the work e.g. by only providing finance and not exercising any degree of organizational control.89
Following these principles, the user may struggle to claim authorship simply because they are the person who provides the final instruction and presses the button to create any work—they will need to do more. Conversely, AI developers may claim to be the driving force behind the work because it wouldn’t exist without the system that they developed, but they may become too far removed from its creation, particularly when they don’t retain control over how their system is used.
Overall, it is clear from this analysis that, as was Parliament’s intention, there is no general rule and authorship will be allocated depending on the particular circumstance.90 This has its advantages: it is flexible and can deliver a fair result in each case. However, it also leads to inevitable uncertainty in this hugely important and ever-expanding industry.
Proposals for change
Despite the concerns described above, there are unlikely to be any changes to s.9(3) in the foreseeable future. In June 2022, the UK government acknowledged91 that stakeholders had raised concerns about the originality criterion but saw no need for immediate change: the provision is rarely used in practice, there is little evidence that it is causing any harm, and AI development is still in its early stages. Instead, the government promised to continue to evaluate options as AI systems advance. This is not an entirely satisfactory answer given today’s exploding generative AI market—we are already witnessing a huge surge in AI-generated content, so there is a real and urgent need, either to clarify the s.9(3) protection mechanism or to devise a sensible alternative.
A sui generis right? The recognition of machine authorship
To date, the UK government has rejected the idea of protecting AI-generated content by way of a separate sui generis or related right rather than as a LDMA work, although Bently argues that UK law already achieves this, in substance, because of the peculiarities of s.9(3) protection92. Protection is as a LDMA work but is granted to the person undertaking the necessary arrangements for their creation (usually reserved for entrepreneurial works); *E.I.P.R. 478 the duration is a fixed shorter term (equivalent to broadcasts and unpublished sound recordings); and the designated author obtains economic rights but no moral rights (again, similar to most entrepreneurial works). Nevertheless, setting this out in the legislation as a separate category of copyright protection (e.g. a “machine creation”) would be far easier to understand. The legislation could clarify the originality requirement (if any) for this type of protection, could introduce a shorter duration and provide further guidance on the identity of the author in light of AI technology.93 This “lesser” right could still incentivize investment in AI, whilst respecting human creativity, which would continue to attract broader economic and full moral rights.
The downside to this would be the added layer of complexity introduced by parallel systems of human and AI-derived works. If there were different levels of originality, people might be tempted to “game the system” and claim rights over machine-generated content when they can’t meet the higher threshold of originality required for a human-authored work. Alternatively, they may try to claim the higher level of copyright protection for a work generated by a machine. To avoid this, there would need to be, either a presumption that works are generated by humans unless there is proof to the contrary (or vice versa), or a suitable method for distinguishing works that are created by humans from those created by AI. This might be something for the future because present discussions concerning labelling of AI-works are still at an early stage.94
Preferred approach: Improvements to section 9(3)
Even without a new category of work, legislative changes could still be made to s.9(3) to clarify the issues identified in this article. The originality requirement for LDMA works should be dealt with expressly and there would be more certainty (but less flexibility) if the section identified more clearly the author of works generated by AI. This would then allow those who disagreed to adjust the allocation contractually. This would be a proportionate way of remedying the issues rather than abolishing the s.9(3) protection altogether.95
The UK was progressive and forward-thinking in 1987 when it decided to encourage AI innovation and investment by providing copyright protection for computer-generated works. The conservative government also stated that it wanted to establish a pro- innovation environment for AI.96 However, in order to achieve this, it is imperative that copyright laws keep up with the pace of technological developments, and provide a coherent, transparent and fair mechanism for protecting (potentially very valuable) computer-generated works.
Footnotes
1. The Guardian, "A portrait created by AI just sold for USD432,000. But is it really art?" (26 October 2018).
2. Pichai and D. Hassabis, "Introducing Gemini: our largest and most capable AI model" (06 December 2023), Google The Keyword
3. TechCrunch, Samsung’s Galaxy S24 line arrives with camera improvements and generative AI tricks (17 January 2024).
4. Other jurisdictions with similar provisions include: New Zealand (Copyright Act 1994 s.9(3)), Ireland (Copyright and Related Rights Act 2000 21(f)), India (Copyright Act 1957 s.5(2)(a)), Hong Kong (Copyright Ordinance s.11(3)).
5. Copyright, Designs and Patents Bill, Hansard, HL, 489, col. 1477 (12 November 1987).
6. Copyright, Designs and Patents Bill, Hansard, HL, 489, col. 1521 (12 November 1987).
7. UK Intellectual Property Office, "Consultation outcome - Artificial Intelligence and Intellectual Property: copyright and patents: Government response to consultation" (GOV.UK, updated 28 June 2022)
8. Copyright, Designs & Patents Act 1988 (CDPA) 1(1)(a).
9. Laddie, Prescott and Vitoria, "The Modern Law of Copyright" 5th at [3.40].
10. THJ Systems Ltd v Sheridan [2023] EWCA Civ 1354 (THJ Systems) at [23].
11. See, for example, SAS Institute Inc v World Programming Limited [2013] EWCA Civ 1482 at [36]–[37]: "If the Information Society Directive has changed the traditional domestic test, it seems to me that it has raised rather than lowered the hurdle to obtaining copyright protection".
12. Eva-Maria Painer v Standard Verlags GmbH (C-145/10) C:2011:798 at [89]–[94]; Football Dataco Ltd v Yahoo! UK Ltd (C-604/10) EU:C:2012:115 (Football Dataco) at [38]; Funke Medien NRW v Germany (C-469/17) EU:C:2019:623 (Funke Medien) at [19], [23]-[25]; Cofemel—Sociedade de Vestuário SA v G- Star Raw CV (C-683/17) EU:C:2019:721 (Cofemel) at [30]; and SI, Brompton Bicycle Ltd v Chedech/Get2Get (Brompton Bicycle) EU:C:2020:461 (C-833/18) at [23] and [26].
13. Bezpecnostní softwarová asociace – Svaz softwarové ochrany v Ministerstvo kultury (C-393/09) EU:C:2010:816 at [48]-[49]; Football Association Premier League Ltd v QC Leisure (C-403/08) EU:C:2011:631 at [98]; Football Dataco at [39]; Funke Medien at [24]; Cofemel at [31]; and Brompton Bicycle at [24] and [27].
14. Infopaq International A/S v Danske Dagblades Forening (C-5/08) EU:C:2009:465 (Infopaq). Arnold LJ in THJ Systems at [27] citing Infopaq at [45]-[48], stated that the consequence of a low degree of creativity is that the scope of protection conferred by copyright is correspondingly narrow, so that only a close copy will infringe.
15. CDPA 5A(2), 5B(4), 8 and 56(6).
16. CDPA 11, which includes exceptions for employees, crown and parliamentary copyright and copyright of certain international organizations.
17. CDPA 9(1).
18. Donoghue v Allied Newspapers Ltd [1938] 106 at [109].
19. Cala Homes v Alfred McAlpine Homes [1995] FSR 818; Kogan v Martin & Ors [2019] EWCA Civ 1645.
20. Donoghue v Allied Newspapers Ltd [1938] 106 at [109].
21. Bently et al, Intellectual Property Law, 6th edn, at [137].
22. Walter v Lane [1899] 2 749.
23. Bently et al, Intellectual Property Law, 6th edn, at [125].
24. Donoghue v Allied Newspapers Ltd [1938] 106; Tate v Thomas [1921] 1 Ch. 503.
25. Copinger and Skone James on Copyright 18th at [4-22] and [4-23]; Ibcos Computers Ltd v Barclays Mercantile Highland Finance Ltd [1994] F.S.R. 275 at [302].
26. Copinger and Skone James on Copyright 18th (Copinger) at [4-23].
27. CDPA 9(1).
28. CDPA 9(2).
29. CDPA 9(2)(aa).
30. CDPA 178.
31. CDPA 9(2)(a) (now amended). Since 1 July 1994, s.9(1)(2)(ab) CDPA has recognized the creative contribution of the film director and the authors of a film are the producer and principal director.
32. CDPA 178.
33. The CDPA also amended 2 of the Registered Designs Act 1949 to provide for designs generated by computer in circumstances where there is no human, such that the person by whom the arrangements necessary for the creation of the design are made will be the author. Similarly, in relation to UK unregistered design right, CDPA s.214(2) states that, in the case of a computer-generated design, the person by whom the arrangements necessary for the creation of the design are undertaken shall be taken to be the designer.
34. CDPA 12(7).
35. CDPA 79(2)(c) and 81(2).A
36. Bently et al, Intellectual Property Law, 6th Edn, at [127].
37. Eva-Maria Painer v Standard Verlags GmbH (C-145/10) C:2011:798 at [89]–[94] at [121].
38. Second Request for Reconsideration for Refusal to Register Théâtre D’opéra Spatial (Copyright Review Board September 5, 2023). U.S. Copyright Office, Library of Congress. Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence, 16 March 2023 88 FR 16190.
39. Australia: it is necessary to identify a human author in order for there to be an original literary work (Telstra Corporation Limited v Phone Directories Company Pty Ltd (2010) FCA 44); Singapore: copyright only arises when a work is created by a human author (Asia Pacific Publishing Pte Ltd v Pioneers & Leaders (Publishers) Pte Ltd [2011] SGCA 37.
40. Goold, ORCID: 0000-0003-1097-8291 (2021). The Curious Case of Computer-Generated Works under the Copyright, Designs and Patents Act 1988 (City Law School Research Paper 2021/03). London, UK: The City Law School.
41. Copyright, Designs and Patents Bill, Hansard, HL, 489, col. 1477 (12 November 1987).
42. Bently et al, Intellectual Property Law, 6th edn, at [127].
43. Department of Trade Reform of the Law relating to Copyright, Designs and Performers’ Protection: A Consultative Document (Cmnd 8302), July 1981 at [35].
44. Guadamuz, Do androids dream of electric copyright? Comparative analysis of originality in artificial intelligence generated works, IPQ 2017, 2 169–186 at [176].
45. Bently et al, Intellectual Property Law, 6th edn, at [127].
46. Copinger and Skone James on Copyright 18th edn, [3-238].
47. Department of Trade and Industry, IP and Innovation (White Paper, Cmnd 9712, 1986).
48. Laddie, Prescott and Vitoria, The Modern Law of Copyright 5th edn, at [36.44].
49. CDPA 12(1) and (2).
50. Express Newspapers Plc v Liverpool Daily Post & Echo Plc [1985] 1 L.R. 1089; [1985] F.S.R. 306.
51. Copinger and Skone James on Copyright 18th edn, at [3-237].
52. It was not until 2017 that Google launched the first transformer model, which is the foundation of many generative AI
53. Emotional Perception AI Ltd v Comptroller-General of Patents, Designs and Trade Marks [2023] EWHC 2948 (Ch) at [11].
54. Emotional Perception AI Ltd v Comptroller-General of Patents, Designs and Trade Marks [2023] EWHC 2948 (Ch) at [37].
55. Guadamuz, Do androids dream of electric copyright? Comparative analysis of originality in artificial intelligence generated works, IPQ 2017, 2 169-186.
56. Expedia Group, Chatgpt Wrote This Press Release—No, It Didn’t, But It Can Now Assist With Travel Planning In The Expedia App (04 April 2023), available at: https://www.expediagroup.com/investors/news-and-events/financial-releases/news/news-details/2023/Chatgpt-Wrote-This-Press-Release--No-It-Didnt-But-It-Can-Now- Assist-With-Travel-Planning-In-The-Expedia-App/default.aspx/.
57. Li Yunkai v Liu Yuanchun (2023) Jing 0491 Min Chu 11279, 27 November 2023.
58. OpenAI, How should AI systems behave, and who should decide? (16 February 2023)
59. OpenAI, Reducing bias and improving safety in DALL·E 2 (18 July 2022)
60. Morais, Parameters: The Secret Power To Have More Control of ChatGPT (07 June 2023)
61. Copinger and Skone James on Copyright, 18th edn, at [4-22].
62. U.S. Copyright office letter, 21 February 2023 at [7]
63. U.S. Copyright office letter, 21 February 2023 at [9]
64. U.S. Copyright office letter, 21 February 2023 at [10]
65. Hedrick, I Think, Therefore I Create: Claiming Copyright in the Outputs of Algorithms (2019) 8 NYU J. Intell. Prop. & Ent. L. 324.
66. Li Yunkai v Liu Yuanchun (2023) Jing 0491 Min Chu 11279, 27 November 2023.
67. Š. v Taubel Ref. No. 10 C 13/2023-16.
68. Article 5(1) of the Czech Copyright Act (Act No. 121/2000 on Copyright and Rights Related to Copyright and on Amendment to Certain Acts (the Copyright Act), as amended by Act 81/2005, Act No. 61/2006 and Act No. 216/2006) states that the author of a work is "the natural person who created the work".
69. Nova Productions Ltd v Mazooma Games Ltd [2006] EWHC 24 (Ch) at [105]-[106].
70. Donoghue v Allied Newspapers Ltd [1938] 106 at [109].
71. Donoghue v Allied Newspapers Ltd [1938] 106; Tate v Thomas [1921] 1 Ch. 503.
72. Eva-Maria Painer v Standard Verlags GmbH (C-145/10) C:2011:798 at [89]–[94].
73. For example, ControlNet
74. CDPA 10(1): Kogan v Martin [2019] EWCA Civ 1645 at [31].
75. Kogan v Martin & Ors [2019] EWCA Civ 1645 at [31].
76. Copinger and Skone James on Copyright 18th at [3-238].
77. The Whitford Committee Report on Copyright and Designs Law (Cmnd 6732), March
78. Reform of the Law relating to Copyright, Designs and Performers’ Protection (Cmnd 8302), July
79. Department of Trade and Industry, Intellectual Property and Innovation (Cmnd 9712), April
80. Lord Beaverbrook, HL Deb vol 493 col 1305 25 February
81.UK Intellectual Property Office, "Consultation outcome—Artificial Intelligence and Intellectual Property: copyright and patents: Government response to consultation" (GOV.UK, updated 28 June 2022)
82. Nova Productions Ltd v Mazooma Games Ltd [2006] EWHC 24 (Ch) at [105]–[106].
83. Dickenson et al., Creative machines: ownership of copyright in content created by artificial intelligence applications, E.I.P.R. 2017, 39(8), 457–460 at [458].
84. Department of Trade and Industry, Intellectual Property and Innovation (Cmnd 9712), April
85. Dworkin and R. Taylor, Blackstone’s Guide to the Copyright, Designs & Patents Act 1988 (London: Blackstone Press, 1989) at p.47.
86. Guadamuz, Andres, "Do Androids Dream of Electric Copyright? Comparative Analysis of Originality in Artificial Intelligence Generated Works" (June 5, 2020). Intellectual Property Quarterly, 2017 (2).
87. Bamgboye & Anor v Reed and Ors [2002] EWHC 2922 (QB); Century Communications Ltd v Mayfair Entertainment UK Ltd [1993] E.M.L.R. 335.
88. Adventure Film Productions SA v Tully [1993] M.L.R. 376.
89. Henry Hadaway Organization Ltd v Pickwick Group Ltd [2015] EWHC 3407 (IPEC); Adventure Film Productions SA v Tully [1993] E.M.L.R. 376; Slater v Wimmer [2012] EWPCC 7.
90. Bently et al, Intellectual Property Law, 6th Edn at [138].
91. UK Intellectual Property Office, "Consultation outcome—Artificial Intelligence and Intellectual Property: copyright and patents: Government response to consultation" (GOV.UK, updated 28 June 2022)
92. Bently et al, Intellectual Property Law, 6th edn, at [128].
93. UK Intellectual Property Office, "Consultation outcome—Artificial Intelligence and Intellectual Property: copyright and patents: Government response to consultation" (GOV.UK, updated 28 June 2022)
94. Labeling AI-Generated Images on Facebook, Instagram and Threads (6 February 2024)
95. The Software Directive: A UK Perspective in Lehmann M and Tapper C (1993): A handbook of European Software Law, Clarendon Press at [143–161].
96. HM Government, National AI Strategy, September 2021
