Human Creative Contribution to AI-Based Output – One Just Can(’t) Get Enough

Abstract

The article attempts to clarify the issue of the copyrightability of AI-based output by streamlining the assessment of human creative contribution to the creation process assisted by generative AI systems. It starts with briefly outlining the state of the art of modern generative AI systems contributing to a better understanding of AI. Then the article presents a five-part test to distinguish between sufficient and insufficient human creative participation and contemplates the following authorship scenarios: sole authorship of AI designers or users; their joint authorship; and non-authorship. The article proceeds with applying the test to output produced using Midjourney, a generative AI tool. It concludes that in many cases of using AI to create output human creative participation remains sufficient for copyright protection eligibility. However, there are also developments and circumstances that drive the increase in ‘authorless’ AI-based output. Moreover, depending on the specific circumstances, using even the same AI system may lead to different test results.

I. Introduction

Artificial intelligence (AI) has caused quite a stir in creative industries over the last few years. Popular depictions of AI as omnipotent, independent, unpredictable and autonomous have set the stage for fears that existing copyright rules will both fail to cope with the challenges posed by using AI to create output and hinder innovation. Today, when the veil is being gradually lifted and expectations of AI performance do not always match reality, this article goes back to the beginning to find the end.¹ In copyright law, such a beginning is the argument that AI can be more than a ‘mere tool’ in human hands. While views are divided over this issue, those who support this belief often take it as a factual premise without elaborating on where the tool/non-tool boundary lies.

This article attempts to clarify the issue of copyrightability of AI-based output, given that it is the starting point in a chain of subsequent legal challenges (such as justification for protection and search for an alternative regime). Section II briefly discusses the state of the art of modern generative AI systems. Section III presents a five-part test to distinguish between sufficient and insufficient human contribution to the creative process when assisted by generative AI systems and, thus, between copyrightable and non-copyrightable AI-based output. Section IV contemplates the following authorship scenarios that may arise after the application of the test: sole authorship of AI designers or users; their joint authorship; and non-authorship. Section V applies the test to output produced using Midjourney, a generative AI tool. Section VI concludes that while modern generative AI systems remain tools in human hands there are developments and circumstances that increasingly entail a decrease in human creative contribution to the process of generating output and an increase in a class of ‘authorless’ output.

II. Deromanticising AI

When it comes to using AI in the creative industries, headlines about AI ‘killing’ human creativity² may pop into our minds. This media terminology seems to have its counterpart in legal academia, which is AI that ceases to be a ‘mere tool’.³ Even a magic wand appears to require more from its master than an AI does from a human. AI is attributed with the ability to programme itself⁴ and to take autonomous and independent steps,⁵ particularly to make its own creative ‘largely independent decisions that surmount human input and creativity’⁶ and result in ‘producing different creative works without any human assistance’.⁷ All one needs to do is simply push a button. This perception of AI stems from an initial lack of understanding of its technical aspects and a following anthropomorphic framing of AI in legal and policy discussions.

However, to paraphrase a famous expression, AI is as AI does.⁸ The trend towards deromanticising AI has already resulted in the debunking of many myths about its capabilities and shown that reality is more mundane. This article focuses on generative AI. While the definition of generative AI is yet to be formalised, it is common to use this term to refer to AI systems that make possible the generation of new output.⁹ Although the generative AI boom has fuelled debate on the approach or even the advent of the artificial general intelligence era,¹⁰ this article takes a cautious stance on categorising generative AI systems as such. It acknowledges significant advances in this field; however, it also argues that modern generative AI systems remain constrained to the tasks they were designed to perform.

This article also argues that the following characteristics hold true for modern generative AI systems. The functionalities of AI systems are based on design choices made by humans¹¹ and the conditions of systems’ deployment. Humans continue to define areas of application and objectives for AI,¹² and ‘pave the way’ to achieving the set goals.¹³ Creating AI systems capable of transcending constraints on their own remains at the level of aspirations.¹⁴ In other words, modern generative AI systems play chess and many other games even better while the room is still on fire.¹⁵ This leads the author of this article to consider that, as before,¹⁶ such systems continue to fall into the category of artificial narrow intelligence.¹⁷

Then how can AI-based outcomes cause a ‘surprising’ effect?¹⁸ AI’s characteristic of being a black box and giving the impression of unpredictability, randomness and an independent mind¹⁹ stems from the complexity and magnitude of computations performed by modern AI systems and their resulting limited explainability.²⁰ How about autonomy, a property that is widely attributed to AI systems? Did this myth stand up to scrutiny? The commonly reported AI autonomy has turned out to be no more than a high level of automation.²¹ Thus, if being a tool presupposes ‘assist[ing] the expression of ideas’,²² given the current state of the art, generative AI systems continue to fall into this category,²³ though as sophisticated²⁴ and especially efficient computational²⁵ tools.

The author of this article admits that she was also tempted by the use of anthropomorphic language at the outset of her research on the subject. However, she takes this opportunity to contribute to reversing this trend. In view of the criticism of the anthropomorphisation of AI,²⁶ this article considers it more beneficial for the research to refer to:

• i) automation instead of autonomy;

• ii) using AI for creating/generating output instead of AI able to create on its own; and

• iii) sufficient/insufficient human creative contribution instead of AI as a tool/non-tool.

At the same time, one cannot ignore the trend of AI systems continuing to grow in complexity. The aforementioned computational capacity and high level of automation distinguish modern generative AI systems from ‘traditional’ computer programmes. These properties do not endow AI with creative abilities, but they do reduce the creative input of humans.²⁷ A reduction in participation to the creative process alone does not translate into non-copyrightability, as long as the human creative contribution remains sufficient. Given the multiple ways in which AI systems can be used, different levels of automation and other technical aspects of AI systems, the determination of a transition point where human creative contribution fades is highly case-dependent.²⁸ The desire to make the issue of copyright protection eligibility of AI-based output clearer necessitates the development of a test for distinguishing between sufficient and insufficient human creative contribution when using AI systems.²⁹ That is the goal this article seeks to achieve.

III. Delineating sufficient and insufficient human creative contribution

Compared to the number of mentions that using AI to create output may make it non-copyrightable, relatively few attempts have been made to address the issue of determining sufficient and insufficient human creative participation in this regard. Without claiming to list all of them, this article will mention the tests and criteria proposed by Bruce E. Boyden,³⁰ Tim W. Dornis,³¹ Jane C. Ginsburg and Luke Ali Budiardjo,³² Daniel J. Gervais,³³ P. Bernt Hugenholtz and João Pedro Quintais,³⁴ and Ana Ramalho.³⁵ Below this article presents its contribution to turning this case-specific issue³⁶ into a streamlined task. Building on existing advances in the matter, it offers a five-part test. The test is centred on the process of creating and therefore primarily targets jurisdictions with a subjective standard of originality. The parts of the test have been arranged in such a way as to ensure a smooth flow through the process of assessment, but they do not present a strict sequence. It is possible (and even encouraged) to return to the relevant phase when there is a need to clarify any issue.³⁷

1. AI systems, the technical aspect

Part one of the test involves exploring the technical aspects of the AI system in question. Determining whether a painting qualifies for copyright protection does not involve an enquiry into the process of manufacturing brushes and paints. The same applies to photos: their eligibility does not depend on technical aspects of cameras. Even reliance on ‘traditional’ computer programmes for creating output does not presuppose knowledge of their design and operation when validating such output against copyrightability requirements. We can think of similar examples covering other copyright subject matter. Then what makes modern generative AI systems so ‘special’ that we have to reconsider our standard practices? There are two main reasons. First, as indicated earlier,³⁸ the fact that the use of AI for creating output ‘may blur the human contribution’³⁹ to the creative process emerges from the technical properties of such systems. This distinguishes AI from other tools. In keeping with the previous analogies, ‘the paintbrush does not have the capacity to change, it does not make decisions based on past painting experiences, and it is not trained to learn from data’,⁴⁰ and in the case of computer programmes ‘the software developer will anticipate the desired response for all possible inputs to the computer’.⁴¹ Exploring technical aspects of the AI system in question can assist in determining whether AI designers or users retain control over the output’s expression. Second, where AI-based output proves to fall under copyright protection, an understanding of the design and operation of the AI system in question will assist in attributing creative decisions to the human author(s).

The first place to start the ‘technical exploration’ is identifying the type of AI system. One common way of classifying AI systems is to divide them into symbolic AI and statistical or sub-symbolic AI.⁴² The reason why this is important for the test is found in the fact that ‘[a]s a rule, symbolic AI systems reflect a higher degree of human contribution or intervention – including a closer connection to an output – than statistical AI systems’.⁴³ The group of sub-symbolic AI encompasses many techniques, with machine learning (ML)⁴⁴ being among them.⁴⁵ Since ML is currently the most commonly used subfield of AI, we will focus on it in this article.

For the test presented in this article the key factors are the way in which the given AI system is built (methods and models) and the way in which it is trained (data, process and techniques). The overall performance of a particular AI system depends on the combination of these factors. Nonetheless, this article further considers them in isolation to justify the relevance of each for the test. It is believed that methods of ML influence the model robustness and the variability of the output.⁴⁶ The choice of methods is dictated by the objectives pursued in creating the given AI system. In this article we examine generative AI.⁴⁷ The history of generative AI, which dates back more than 50 years,⁴⁸ reveals a shift in preference from evolutionary computation to neural networks.⁴⁹ Nowadays there are even specific ML models that have become popular for generative AI purposes.⁵⁰ Furthermore, it is believed that currently ‘there [is] no such thing as a single best ML model’⁵¹ fit for all problems. For the purposes of our test, it is important that deciding on a method and specific model can affect the complexity and unpredictability of the output,⁵² as well as its quality and style.⁵³

Since ML models are data-driven, input data and training process play an essential role too. As far as training data is concerned, both quality and quantity matter. Training a model using bad data will most likely result in a bad model, ‘a process known as garbage in garbage out’.⁵⁴ In turn, lack of training data can prevent a model from generalising⁵⁵ and lead to its overfitting. Under such circumstances a model ‘will produce output that is very similar to the training data’.⁵⁶ Training techniques are also among the key factors.⁵⁷ Examining training techniques can be useful in clarifying whether the training process was laborious and whether it allowed for creativity.

Other technical factors may be relevant for the test. Such additional technical aspects of an AI system may manifest themselves in the course of exploring the main ones. One example might be AI’s explainability: AI systems may be ‘too complex and multi-dimensional for human [designers] to comprehend’.⁵⁸ This can translate into a lack of understanding as to the output’s expression.⁵⁹ Such lack of human comprehension can be an additional argument in attesting to designers’ ignorance regarding future outcome.

2. Human-made decisions, the quantitative aspect

Part two of the test involves identifying all human-made decisions in the creation process (the quantitative aspect). It requires one to ‘work back through the creative process until a human was involved’.⁶⁰ As reasoned in the previous section, in our case the search for human-made decisions extends to the creation of a generative AI system. Thus all human-made decisions in the creation process can be grouped into two categories:

• i) those that are taken to create a generative AI system; and

• ii) those that are made to generate output.

3. Human-made decisions, the qualitative aspect

Part three of the test calls for evaluating the human-made decisions and singling out the free and creative ones (the qualitative aspect). That is easier said than done. First, creative choices may spring from a combination of unprotected ideas.⁶¹ Second, there may be ‘chameleon’ choices with a nature that depends on the specific circumstances. Thus both decisions to create a generative AI system and to generate an outcome can be further divided into three sub-categories: (i) technical;⁶² (ii) free and creative; and (iii) ‘chameleon’.

On the part of AI designers,⁶³ for example, a technical choice would be deciding on the number of an output’s variations provided to a user to select from, and a creative choice would be modifying ML model parameters to influence the final output. Selection of training data would be an example of a ‘chameleon’ choice. For example, the selection of data which is dictated by the desire to train an ML model to distinguish between dogs and cats requires effort rather than creativity. In turn, training an ML model on certain data with a view to generating outcomes in a particular style may involve creativity.

An example of a technical choice by users would be the one about the number of generation attempts. Selecting a genre, tempo and rhythm, tonality, loudness, and timbre, etc. to generate music would fall under creative decisions. Opting for a particular AI system is a ‘chameleon’ decision: it may turn into a creative one when it affects the output, e.g. an AI system that generates output in a certain style and quality.

4. The ‘creative distance’ between decisions and expression

Part four of the test requires assessing the ‘creative distance’⁶⁴ between the free and creative decisions identified in the previous phase and the output’s expression, namely whether the relevant choices ‘lead to the manifestation of the key expressive elements of the [outcome]’.⁶⁵ This is due to the fact that heavy reliance on AI in the creative process ‘stretches the causation bond between the human author and the final creative output’⁶⁶ ‘to the point that it may no longer be possible to trace the creative elements directly to the mind of a particular human author’.⁶⁷ For the purposes of this part of the test it is also essential to establish general authorial intent.⁶⁸ Without excluding unintended expressive features,⁶⁹ it allows for determining whether the final expression falls within the ambit of the author, and whether the author’s free and creative decisions allowed for it.

5. The originality requirement

Part five of the test necessitates determining whether the free and creative decisions that led to the final output are sufficient to meet the originality requirement. This part of the test may encounter resistance due to the difficulties that establishing originality entails.⁷⁰ However, this presents no novel challenges. It has been with us since the establishment of copyright law. Moreover, there is no new subject matter: what is different is the process of creation. Another thing that remains unchanged is the difference in the originality thresholds of national copyright laws. Thus the final decision on the copyrightability of AI-based output is national law specific. One may also assume that the use of AI systems to generate output may prompt revisiting the originality requirement, but that is a subject for another enquiry.

IV. (Non-)authorship scenarios

Having laid out the test, this article proceeds with discussing authorship scenarios that may arise after the application thereof.

1. Authorship of AI designers

Where designers create an AI system intended to generate

• i) a particular outcome with no users involved,

• ii) an output that remains the same ‘no matter which human user caused the output to be generated’⁷¹ or

• iii) a result where a user is bound by and has a limited role in the ‘completed creative plan’ of the AI designers,⁷²

the free and creative choices made to develop the AI system qualify as the ones that determined the outcome.⁷³ Even where AI designers do not have a clear and precise idea of the future output, the constraints they set in the course of building a system may result in them being authors.⁷⁴ However, taking into account the trend ‘that more and more AI capability is being offered “as a service” rather than as “bespoke” tailored AI systems’,⁷⁵ in many cases AI designers and investors are not so much interested in creating particular content as in enabling others to create their output. This is the point where users take over the act of output’s creation.

2. Authorship of users of AI

When a generative AI system is designed to require the participation of users and does not limit their ‘creative autonomy’,⁷⁶ the user is considered to be ‘the person who most immediately cause[s] the work to be brought into being’⁷⁷ and whose ‘creative contribution interrupts the designer’s authorship claim’.⁷⁸ This is due to the fact that it is the user who switches on a device, runs an AI programme, decides to create output⁷⁹ and, thus, ‘supplies [the work’s] but-for cause’.⁸⁰ An important requirement here is that the user maintains sufficient creative control over the process.⁸¹ Out of three phases (conception, execution and redaction) of a creative process, AI has largely ‘taken over’ the execution stage, but humans maintain their role in the remaining phases.⁸²

Thus the more free and creative choices a user makes during both or one of these phases, the greater the chance of obtaining copyright.⁸³ Under such circumstances, the utilising of AI will be no more than ‘the controlled outsourcing of tasks to a sophisticated machine tool’.⁸⁴ This does not pose a novel risk to copyright law and will not result in non-authorship.

3. Joint authorship

It has been argued that if creative choices belong to both AI designers and users, the joint authorship option may come into play.⁸⁵ Its main advantage consists in ‘avoid[ing] the necessity of having to elect from two AI authorship frontrunners’.⁸⁶ However, upon closer examination, this approach seems to have little prospect from both doctrinal and practical standpoints.⁸⁷

The doctrinal part of the reflection on this issue should begin by acknowledging ‘different national characterizations of joint works and joint authorship’.⁸⁸ The source of these differences is the legal uncertainty that the Berne Convention for the Protection of Literary and Artistic leaves regarding multiple authorship.⁸⁹ Notwithstanding the lack of a common approach to joint authorship, it is believed that ‘the [Berne] Convention concept of joint works implies active collaboration between the participants’.⁹⁰ National copyright laws commonly support and develop the idea of cooperation.⁹¹ This requirement calls into question the viability of the joint authorship option as regards AI-based output.⁹² When AI is used for creating output, generally there is a cooperation gap⁹³ between designers and users in terms of knowledge,⁹⁴ time and space.⁹⁵ The mere fact that participants collaborate at different stages is not in itself a barrier to co-authorship. The obstacle lies in the lack of cooperation between the designers and the users of AI in a material way on generating specific outputs.⁹⁶

On a practical side, considering AI-based output through the lens of joint authorship ‘may result in an endless loop of sub-divided rights’.⁹⁷ The fact that ‘[j]oint authorship fractionates ownership rights, rather than consolidating them’⁹⁸ reinforces the above suggestion. Moreover, another obstacle may lie in the exploitation of the output, as national copyright laws may provide for a requirement to reach an agreement between all co-authors or to account to the other co-authors.⁹⁹ Such complexities result in ‘little basis for harmony of interest between these parties’.¹⁰⁰ This has shaped the following trend: when AI systems are offered as a service, AI designers usually seek to avoid burdening users with downstream copyright claims.¹⁰¹ Terms of services commonly have provisions under which the rights to AI-based output remain with users. The latter grant licences for such output to service providers.¹⁰²

4. Non-authorship

There may be cases when applying the test to AI-based output reveals a lack of sufficient human creative participation. Admittedly, AI-based output itself and computations implemented on AI systems result from precise and explicit steps¹⁰³ conceived by the designers of the AI rather than ‘magic’.¹⁰⁴ Nonetheless, AI designers cannot control or anticipate¹⁰⁵ outcomes that will be generated with the utilisation of the AI system, ‘since these depend on further interactions with a user’.¹⁰⁶ Thus they provide ‘the potentiality for the creation of the output, but not its actuality’.¹⁰⁷ This ‘[AI designers’] inevitable ignorance’¹⁰⁸ of future output prevents them from receiving authorship status.¹⁰⁹

When it comes to employing AI systems for generating output by users, their impact on the creative process becomes more noticeable. Though using generative AI still leaves room for human participation in the conception and redaction phases,¹¹⁰ recent developments in generative AI seem to shift the part of humans from that of actors to extras. Particularly in the conception phase, users’ involvement in the creation process may be limited to preconceiving a general idea, specifying a theme, setting the main preferences and goals or making other abstract choices.¹¹¹ Indeed, such choices affect the final output. For instance, if at the outset a user opts for calm relaxing music background accompanied by the gentle sound of water, a programme will not generate an energising melody with rhythmic beats and a quick tempo from these specifications, provided that it works properly. Given that copyright law does not require the author’s conception of the work to ‘reflect a complete or even an accurate prediction of the resulting work’s contents’,¹¹² a glimmer of hope for copyrightability may appear on the horizon. However, in this case there are many dissimilar ways to express the aforementioned parameters. This fact makes the boundaries of the conception blurry, constraints on its execution inadequate¹¹³ and user’s authorship ‘increasingly difficult to defend’.¹¹⁴ After receiving preferences from a user, AI will proceed with executing them and offering one or more options of expressions. ‘Ratifying’ the expression, ‘even if [some elements of the outcome] were not consciously planned’ can compensate for the vague conception and add creativity.¹¹⁵ However, if the user’s role is limited solely to selecting from the provided alternatives, copyright protection does not arise.¹¹⁶ This conclusion can be reversed when a user modifies the output after generating it.¹¹⁷ In view of recent trends, the hope for copyright eligibility evaporates for the final time if such changes are not sufficiently creative.¹¹⁸

Now, let us imagine copyright law – more specifically, its idea/expression dichotomy principle¹¹⁹ – as a rectangular balance board with a centre ball under it. In this analogy an idea becomes an expression and copyright can arise (provided that the copyrightability requirements are met) when both sides of the board get off the ground. In turn, one is confronted with an idea when weight is shifted too much towards either side of the board and the latter hits the ground. One side of this board would be represented by the merger doctrine,¹²⁰ the scènes à faire doctrine¹²¹ and various types of external constraints.¹²² The challenges brought by using AI systems to create output have prompted the author of this article to consider the other side of the board. This article would frame it under the term multiplicity doctrine. Where the first side refers to an idea that (due to specific circumstances) can be expressed in one or a very limited number of ways, the multiplicity doctrine covers abstract and general choices that fail in fully constraining execution¹²³ and leave room for a wide variety of expressions. Thus such choices are too broad to define the output and result in its eligibility for copyright protection. This article does not go further into detail, but, considering the above reflections, this author believes that employment of AI for generating output might reinvigorate the age-old debate on the idea/expression dichotomy.

V. Applying the test to Midjourney-based output

Midjourney is an AI-powered text-to-image tool provided by Midjourney, Inc. ‘to augment human creativity and foster social connection’.¹²⁴ The company frames itself as an independent research lab,¹²⁵ and its founder and CEO David Holz prefers to be referred to only as the founder to tone down the ‘businessy’ connotation.¹²⁶

After completing the registration process and selecting a subscription plan, a user can start generating outcomes with Midjourney.¹²⁷ To do this a user needs to enter a prompt with a command. Prompts are instructions given to AI systems to generate the desired output.¹²⁸ Midjourney’s website provides for two categories of prompts, i.e. text and image prompts. Either of them can be used alone or together with another category. The choice of a command (/imagine or/blend) will depend on the category of a prompt.¹²⁹ This article will further focus on the output generated based on text prompts. As for text prompts, Midjourney’s website specifies that ‘the words and phrases in a prompt [are broken down] into smaller pieces, called tokens, that can be compared to its training data and then used to generate an image’.¹³⁰ After running the prompt with the command, a user will be presented with a grid of four images. Then the user can re-run the same prompt for generating a new grid, create variations of or upscale¹³¹ a particular image. The user can continue experimenting with variating and upscaling the output or choose the preferred image.

Technical aspects of Midjourney are covered by a veil of mystery. The fact that Midjourney is not open source¹³² serves as a contributory factor to this mystery. One may find speculations derived from social media and online forums that the platform uses a combination of ML algorithms (namely convolutional neural networks and generative adversarial networks), computer vision technology and ‘proprietary’ image enhancement techniques.¹³³ However, in a U.S. class-action lawsuit filed against Stability AI Ltd., Stability AI, Inc., Midjourney, Inc. and DeviantArt, Inc., it is mentioned that ‘[o]n information and belief, Stable Diffusion was used in iterations of the Midjourney Product’.¹³⁴ In turn, Stable Diffusion is a latent text-to-image diffusion model.¹³⁵ Thus it remains unclear which methods and model Midjourney is based on. David Holz’s responses regarding the data used for training are not specific either. They range from ‘a big scrape of the Internet’, ‘the open data sets’¹³⁶ and the data based on the users’ preferences regarding the use of Midjourney to ‘12 training datasets’.¹³⁷ The founder describes the default art style of Midjourney as ‘a bit whimsical and abstract and weird’.¹³⁸ In the meantime, users can specify a style they want an image to be generated in. He also mentions that the tool has its ‘favourite’ colours and faces, which apply in the case of vague prompts from users but the Midjourney team pursues the goal of providing a greater variety.¹³⁹ As David Holz puts it, ‘[Midjourney] tends to blend things in ways you might not ask, in ways that are surprising and beautiful’.¹⁴⁰

The lack of information on the technical aspects of Midjourney prevents us from fully assessing the choices made by its designers. Nevertheless, the available information enables one to deduce the designers’ ignorance regarding future outcomes generated by users. Moreover, the utilising of Midjourney demonstrates that different results are generated for the same prompt. This seems to reinforce the said findings. The exception to it may be presented by the output generated based on short prompts that ‘rely heavily on Midjourney’s default style’.¹⁴¹ In the latter case one could assume that authorship may be attributed to the designers. However, a study of technical aspects may reveal that the said default style does not result from ‘deliberate’ choices¹⁴² by the designers.¹⁴³ This may happen when AI designers lack understanding as to why the AI system arrived at a particular result. Such a factor can be an additional element in proving the designers’ ignorance regarding future outcomes that diminishes the chances of authorship for the designers and leads to the ‘authorless’ status of the output.

On the part of Midjourney’s users, the choices that are available to them include the following:

• creating an outcome using Midjourney;

• using a free trial¹⁴⁴/ choosing and subscribing for a paid plan;¹⁴⁵

• creating a prompt/ generating a prompt using other services (e.g. ChatGPT)/using a ready-made prompt;

• using the/imagine or/blend command and running the prompt;

• re-running the same prompt for generating a new grid of outcomes;

• creating variations of the generated output;

• upscaling the generated output;

• after upscaling the output, creating variations of the upscaled result and creating a new grid of four options;

• after upscaling the output, opting for a particular upscaler;

• choosing the preferred generated output;

• saving the generated output;

• opening any upscaled output in the gallery on Midjourney’s website; and

• rating any upscaled outcome.

Among these choices, the following acts may qualify for requiring creativity: deciding to generate an outcome, creating a prompt, choosing a particular upsclaler¹⁴⁶ and opting for the preferred output. One may encounter difficulties in characterising such acts as re-running the same prompt, creating a new grid and variations, and upscaling. At the outset, one may assume that such are indicative of the users’ dissatisfaction with the outcome failing to comply with their plans. Accordingly, such acts may lead ultimately to creativity, as users will continue to run these commands until they ‘arrive at’ and ‘approve’ the output meeting their authorial intent. However, the act of opting for a particular result is singled out in a separate category, and, as argued earlier,¹⁴⁷ this act alone does not suffice for establishing creativity. Examination of the referred acts as such reveals that they constitute a process of iterating the result, which is rather a laborious act. At the end of the day, no matter how resource-intensive such acts are, they do not require creativity. The remaining decisions from the list represent technical steps.

Decisions on generating an outcome and choosing a particular upscaler and the preferred output directly influence the result. However, even if all of them occur, it is unlikely to be sufficient to meet the originality requirement. Therefore, one may assume that the chance for Midjourney-based output to comply with the originality criterion ‘resides’ with the act of creating a prompt.¹⁴⁸

Below we invite you to compare three text prompts by Midjourney’s users.¹⁴⁹

	Prompt	User
1.	Spicy east European soup	by Susi
2.	Men-with-monkey-on-shoulder 35-years-old fantasy	by Merten
3.	White round shiny dinner table with wood chairs: one red, one yellow, one orange, one turquoise, one green and one blue. Colourful bowls and plates on the table. Furniture design, innovative and creative, different, unusual model. Contemporary, extreme, modern, avant garde style. Isolated on white. Photorealistic, like in an interior design magazine. --ar 4:3 --v 5 --s 750	by PixelCat

Among these three prompts, the first one is the shortest and the most general. Since such prompts rely largely on Midjourney’s default options, the chances of authorship for Midjourney’s designers can increase, particularly if such default options come from the informed choices by the designers. However, such interest in authorship on the part of Midjourney’s designers is hardly noticeable.¹⁵⁰ The second prompt is more specific; nonetheless, it is still too broad to define a particular expression or at least its main frames and characteristics. This is the space where the multiplicity doctrine applies. It deprives users of claims for authorship. At the same time, any hope for authorship of Midjourney’s designers slips away as well, as they could not foresee all possible combinations and variations of instructions for generating output. Under such circumstances ‘authorlessness’ may emerge. It can be still compensated if a user continues to edit the output after selecting the preferred alternative and ‘invests’ enough creativity in the redaction phase. The third prompt is the most elaborate. Although there may be differences in executing its details, in general such a prompt provides a clear understanding of the future expression. If the output represents these instructions, the link between the authorial intent, instructions and expression remains unbroken. Under such circumstances a user can be regarded as a candidate for the authorship status. These findings reinforce the modern-sounding speculations about the significance of prompts’ complexity for determining the (in)sufficiency of human creative contribution and, ultimately, the copyrightability of AI-based output.¹⁵¹

At the very end, whenever claims for authorship arise – either for AI designers or users – the final verdict will depend on the particular jurisdiction, its national copyright law and the threshold of originality.

VI. Concluding remarks

Although modern generative AI systems continue to be tools, their properties reduce the human creative participation in producing output using AI. Since the degree of human creative contribution to the creation process assisted by a generative AI system has proven to be a case-by-case issue, this article offers a five-part test to streamline the respective assessment and enhance legal certainty and discusses four authorship scenarios resulting from the application of such test. Inferences on copyrightability and authorship of AI-based output also depend on clarifying the relationship between AI inputs and outputs, and ascertaining whether there are copyright infringements in this respect. These issues deserve special attention¹⁵² and, for this reason, are not covered in this article.

In many cases, human creative participation remains sufficient for copyright protection eligibility. At the same time, there are circumstances that give rise to a class of ‘authorless’ output¹⁵³ produced by distributed collective intelligence¹⁵⁴ where no individual creative contribution is sufficient for copyright protection eligibility. They include the following factors:

• developments in AI;¹⁵⁵

• increasing popularity of offering AI systems ‘as a service’ and designers’ ignorance regarding future outcomes;

• lack of ‘genuine’ collaboration between the designers and users of AI;

• users’ reliance on generic instructions and mere ‘ratification’ of the output.

This article also finds that, depending on the specific circumstances, using even the same generative AI system may lead to different test results. The said ‘authorlessness’ follows, though, from the ‘lack of any author’,¹⁵⁶ rather than from the ability of AI systems to go beyond design on their own or any other myths shrouding AI. This brings us to the emergence of work-like output. It is the outcome that ‘bear[s] the external hallmarks of creativity’,¹⁵⁷ makes an impression of being human-authored and, accordingly, resembles works within the meaning of copyright law.¹⁵⁸ However, examination of the process of its creation does not make it possible to identify a human author. This makes the information on AI utilisation in the creation process a core asset for establishing copyrightability and authorship, and countering its concealment is a major challenge we face today.

ACKNOWLEDGMENTS

The author would like to thank the Max Planck Institute for Innovation and Competition for the tremendous enduring support. This article particularly benefited from the feedback of Dr Daria Kim and Giulio Matarazzi. Any mistakes remain those of the author.

Footnotes

*

Doctoral Student at the Max Planck Institute for Innovation and Competition, Munich, Germany; Taras Shevchenko National University of Kyiv, Ukraine; e-mail: [email protected].

1

This is an allusion to Stephen J Patterson and James M Robinson (trs), ‘The Gospel of Thomas’ (Biblical Archaeology Society, 16 March 2023) 18 <https://www.biblicalarchaeology.org/daily/biblical-topics/bible-versions-and-translations/the-gospel-of-thomas-114-sayings-of-jesus/> accessed 8 May 2023.

2

See eg Nelson Granados, ‘Human Borgs: How Artificial Intelligence Can Kill Creativity and Make Us Dumber’ (Forbes, 31 January 2022) <https://www.forbes.com/sites/nelsongranados/2022/01/31/human-borgs-how-artificial-intelligence-can-kill-creativity-and-make-us-dumber/> accessed 5 April 2023.

3

See eg Florian De Rouck, ‘Moral Rights & AI Environments: The Unique Bond between Intelligent Agents and Their Creations’ (2019) 14(4) JIPLP 299, 301; Paolo Guarda and Laura Trevisanello, ‘Robots as Artists, Robots as Inventors: Is the Intellectual Property Rights World Ready?’ (2021) 43(11) EIPR 740, 740; Yang Xiao, ‘Decoding Authorship: Is There Really no Place for an Algorithmic Author under Copyright Law?’ (2023) 54 IIC 5, 11.

4

See Bram Van Wiele, ‘The Human-Machine Synergy: Boundaries of Human Authorship in AI-Assisted Creations’ (2021) 43(3) EIPR 164, 171.

5

See Martin Senftleben and Laurens Buijtelaar, ‘Robot Creativity: An Incentive-Based Neighbouring Rights Approach’ (2020) 42(12) EIPR 797, 800-01.

6

Van Wiele (n 4) 167.

7

Pratap Devarapalli, ‘Machine Learning to Machine Owning: Redefining the Copyright Ownership from the Perspective of Australian, US, UK and EU Law’ (2018) 40(11) EIPR 722, 727. See also Aviv H Gaon, The Future of Copyright in the Age of Artificial Intelligence (Edward Elgar 2021) 234-41 claiming that AI systems can be creative.

8

This is a paraphrase of the expression ‘Handsome is as handsome does’.

9

See eg Jiao Sun and others, ‘Investigating Explainability of Generative AI for Code through Scenario-Based Design’ (IUI’22: 27th International Conference on Intelligent User Interfaces, Helsinki, 22-25 March 2022) <https://doi.org/10.1145/3490099.3511119> accessed 14 May 2023; ‘Decoding the Magic of Generative AI and How it Works’ (Techfunnel, 22 December 2022) <https://www.techfunnel.com/information-technology/generative-ai/> accessed 10 May 2023.

10

See eg Tambiama Madiega, ‘General-Purpose Artificial Intelligence’ (European Parliamentary Research Service, 31 March 2023) <https://epthinktank.eu/2023/03/31/general-purpose-artificial-intelligence/> accessed 8 May 2023.

11

See Pascal D König and others, ‘Essence of AI: What is AI’ in Larry A DiMatteo, Cristina Poncibò and Michel Cannarsa (eds), The Cambridge Handbook of Artificial Intelligence: Global Perspectives on Law and Ethics (CUP 2022) 28.

12

See Tim W Dornis, ‘Of “Authorless Works” and “Inventions Without Inventor” – The Muddy Waters of “AI Autonomy” in Intellectual Property Doctrine’ (2021) 43(9) EIPR 570, 577-78; Gerald Spindler, ‘AI and Copyright Law: The European Perspective’ in Larry A DiMatteo, Cristina Poncibò and Michel Cannarsa (eds), The Cambridge Handbook of Artificial Intelligence: Global Perspectives on Law and Ethics (CUP 2022) 258.

13

See Jean-Marc Deltorn and Franck Macrez, ‘Authorship in the Age of Machine Learning and Artificial Intelligence’ (2018) CEIPI Research Papers No 10, 14, who, drawing on machine learning systems, list steps humans take to create such systems and bring them into action. These human decisions range ‘from the choice of the training set to the training protocol, from the internal architecture in which the inference model is expressed to the objective function that will decide of its fate’.

14

See Daria Kim and others, ‘Ten Assumptions about Artificial Intelligence that Can Mislead Patent Law Analysis’ (2021) Max Planck Institute for Innovation and Competition Research Paper No 21-18, 25-30.

15

This is a paraphrase of the quote ‘The definition of today’s AI is a machine that can make a perfect chess move while the room is on fire’. Fei-Fei Li referred to it as a phrase written in the 1970s. See Will Knight, ‘Put Humans at the Center of AI’ (MIT Technology Review, 9 October 2017) <https://www.technologyreview.com/2017/10/09/3988/put-humans-at-the-center-of-ai/> accessed 8 May 2023.

16

See eg European Commission, Directorate-General for Communications Networks, Content and Technology, and others, ‘Trends and Developments in Artificial Intelligence Challenges to the Intellectual Property Rights Framework. Final report’ (European Commission 2020) 21; Mauritz Kop, ‘AI & Intellectual Property: Towards an Articulated Public Domain’ (2020) 28 TIPLJ 297, 333; Gaon (n 7) 124; Ana Ramalho, Intellectual Property Protection for AI-generated Creations: Europe, United States, Australia and Japan (Routledge 2021) 10.

17

‘Artificial narrow intelligence comprises systems that are capable of solving specific tasks like playing video games, driving cars, recognition of speech et cetera’. See Noah Klarmann, ‘Artificial Intelligence Narratives: An Objective Perspective on Current Developments’ (ArXiv, 18 March 2021) 7 <https://doi.org/10.48550/arXiv.2103.11961> accessed 20 May 2023.

18

See Kim and others (n 14) 39.

19

See Xiao (n 3) 15-17.

20

See Daria Kim, ‘“ AI-Generated Inventions”: Time to Get the Record Straight?’ [2020] GRUR International 443, 454; See Kim and others (n 14) 41, 44-49.

21

See Kim, ‘AI-Generated Inventions’ (n 20) 446-47; See Kim and others (n 14) 41, 63-71; Eliza Mik, ‘AI as a Legal Person’ in Jyh-An Lee, Reto M Hilty and Kung-Chung Liu (eds), Artificial Intelligence and Intellectual Property (OUP 2021) 422-27.

22

Van Wiele (n 4) 167.

23

See Ryan Abbott and Elizabeth Rothman, ‘AI-Generated Output and Intellectual Property Rights: Takeaways from the Artificial Inventor Project’ (2023) 45(4) EIPR 215, 215.

24

See European Commission, Directorate-General for Communications Networks, Content and Technology, and others (n 16) 116.

25

See Benjamin Williams, ‘Painting by Numbers: Copyright Protection and AI-generated Art’ (2021) 43(12) EIPR 786, 792.

26

Anthropomorphisation may affect human perception of AI and, ultimately, lead humans astray in AI conceptualisation and policy discussions. See David Watson, ‘The Rhetoric and Reality of Anthropomorphism in Artificial Intelligence’ (2019) 29 Minds and Machines 417, 432-35; Kim, ‘AI-Generated Inventions’ (n 20) 455; Carys Craig and Ian Kerr, ‘The Death of the AI Author’ (2021) 52(1) Revue de Droit d’Ottawa 31, 63-65; Kim and others (n 14) 4; Murray Shanahan, ‘Talking About Large Language Models’ (ArXiv, 16 February 2022) <https://doi.org/10.48550/arXiv.2212.03551> accessed 27 June 2023; Isabella Hermann, ‘Artificial Intelligence in Fiction: Between Narratives and Metaphors’ (2023) 38 AI & Society 319; Daria Kim, ‘On Words That Come Easy’ [2023] GRUR International 433, 433.

27

See Dornis, ‘Of Authorless Works’ (n 12) 578; Enrico Bonadio and others, ‘Implications of Artificial Intelligence in Action – a Jamaican Perspective’ (2022) 44(10) European Intellectual Property Review 611, 614; Barry Scannell, ‘When Irish AIs are Smiling: Could Ireland’s Legislative Approach be a Model for Resolving AI Authorship for EU Member States’ (2022) 17(9) Journal of Intellectual Property Law & Practice 727, 729; Spindler (n 12) 258.

28

See Bruce E Boyden, ‘Emergent Works’ (2016) 39 Colum JL & Arts 377, 383; Mikko Antikainen, ‘Copyright Protection and AI-Generated Works - A Fight We Have Already Lost?’ [2018] AIDA 243, 259; Deltorn and Macrez (n 13) 22.

29

See Deltorn and Macrez (n 13) 7; Christina Varytimidou, ‘The New A(I)rt Movement and Its Copyright Protection: Immoral or E-Moral?’ [2023] GRUR International 357, 361.

30

See Boyden (n 28) 385-94.

31

See Dornis, ‘Of Authorless Works’ (n 12) 578-81.

32

See Jane C Ginsburg and Luke Ali Budiardjo, ‘Authors and Machines’ (2019) 34 Berkeley Tech LJ 343, 424-26.

33

See Daniel J Gervais, ‘The Machine as Author’ (2020) 105 Iowa L Rev 2053, 2098-105.

34

See P Bernt Hugenholtz and João Pedro Quintais, ‘Copyright and Artificial Creation: Does EU Copyright Law Protect AI-Assisted Output?’ (2021) 52 IIC 1190, 1200-07.

35

See Ramalho (n 16) 54-57.

36

Text to n 28.

37

eg, during the fourth part of the test (the assessment of ‘creative distance’ between the relevant decisions and the output’s expression) one may find that it is necessary to revisit the first part (the exploration of the technical aspects of the AI system in question) to clarify the impact of the technical aspects on the output’s expression.

38

Text to n 27.

39

Deltorn and Macrez (n 13) 16.

40

Marian Mazzone and Ahmed Elgammal, ‘Art, Creativity, and the Potential of Artificial Intelligence’ (2019) 8(1) Arts <https://doi.org/10.3390/arts8010026> accessed 17 May 2023.

41

Begoña Gonzalez Otero, ‘Machine Learning Models under the Copyright Microscope: Is EU Copyright Fit for Purpose?’ (2020) Max Planck Institute for Innovation and Competition Research Paper No 21-02.

42

See European Commission, Directorate-General for Communications Networks, Content and Technology, and others (n 16) 24. There are also in-between methods that combine symbolic and sub-symbolic AI approaches. See Eleni Ilkou and Maria Koutraki, ‘Symbolic Vs Sub-symbolic AI Methods: Friends or Enemies?’ (CIKM Workshops, Galway, October 2020).

43

European Commission, Directorate-General for Communications Networks, Content and Technology, and others (n 16) 25.

44

Machine learning is a subset of artificial intelligence ‘which involves learning from data. The purpose of machine learning is to facilitate a computer to achieve a specific task without explicit instruction by an external party’. See Kishor Bharti and others, ‘Machine Learning meets Quantum Foundations: A Brief Survey’ (ArXiv, 12 June 2020) 2 <https://arxiv.org/abs/2003.11224> accessed 20 May 2023.

45

See European Commission, Directorate-General for Communications Networks, Content and Technology, and others (n 16) 25 mentioning that the group of statistical techniques ‘includes machine learning, neural networks, deep learning, decision trees, and many other sorts of learning techniques’.

46

See Kim and others (n 14) 61.

47

For the definition, see text to n 9.

48

See Mazzone and Elgammal (n 40).

49

See Dipankar Dasgupta, Deepak Venugopal and Kishor Datta Gupta, ‘A Review of Generative AI from Historical Perspectives’ (TechRxiv, 17 February 2023) <https://doi.org/10.36227/techrxiv.22097942.v1> accessed 8 April 2023. For the factors that enabled the leap in generative AI, see also Eva Cetinic and James She, ‘Understanding and Creating Art with AI: Review and Outlook’ (2021) <https://doi.org/10.48550/arXiv.2102.09109> accessed 14 May 2023.

50

They include generative adversarial networks, variational autoencoders, transformer models and diffusion models. See eg Abhilash Krishnan, ‘Generative AI and Its Possibilities in Creating Art Forms’ (Medium, 5 January 2023) <https://medium.com/@abhilashkrish/generative-ai-and-its-possibilities-in-creating-art-forms-ba989e3a329d> accessed 8 April 2023. See also Miguel Civit and others, ‘A Systematic Review of Artificial Intelligence-Based Music Generation: Scope, Applications, and Future Trends’ (2022) 209 Expert Systems with Applications 1, 9-14; Dasgupta, Venugopal and Gupta (n 49).

51

See Michael A Lones, ‘How to Avoid Machine Learning Pitfalls: A Guide for Academic Researchers’ (ArXiv, 9 February 2023) 6 <https://doi.org/10.48550/arXiv.2108.02497> accessed 16 May 2023.

52

See eg Civit and others (n 50) 12. See also Nicholas Carlini and others, ‘Quantifying Memorization Across Neural Language Models’ (ArXiv, 6 March 2023) <https://doi.org/10.48550/arXiv.2202.07646> accessed 27 June 2023, who in their article on large language models and memorisation identified three properties that contribute to large language models memorising parts of their training data. This results in the fact that ‘when prompted appropriately, [large language models] will emit the memorized training data verbatim’. One of the properties that ‘significantly impact memorization’ is model scale: ‘larger models memorize significantly more than smaller models do’.

53

See eg Dasgupta, Venugopal and Gupta (n 49).

54

See Lones (n 51) 3.

55

ibid.

56

Kim, ‘On Words That Come Easy’ (n 26) 434. Another factor that may have a negative impact on the ability of AI to generalise can be data duplication. See Carlini and others (n 52).

57

Basically, they are classified into four categories, which cover supervised, unsupervised, semi-supervised and reinforcement learning. See SAS Institute, ‘The Machine Learning Landscape. A Quick Guide to the Different Types of Learning’ (White Paper, 2018) 1-2.

58

Ginsburg and Budiardjo (n 32) 402.

59

eg AI designers can lack understanding of why an AI system started ‘favouring’ certain colours.

60

Scannell (n 27) 739.

61

See European Commission, Directorate-General for Communications Networks, Content and Technology, and others (n 16) 73.

62

By ‘technical’ this article means decisions that are driven by external constraints and considerations devoid of creative freedom. For the external constraints in the EU copyright framework see European Commission, Directorate-General for Communications Networks, Content and Technology, and others (n 16) 72-74.

63

As development of AI systems is not limited to ‘traditional programming’ and requires expertise that does not involve programming as such (eg expertise in the domain-specific area or data-related tasks), this article uses the term ‘AI designers’ to collectively refer to human participants involved in the different stages of building AI systems.

64

See Ramalho (n 16) 55.

65

Ginsburg and Budiardjo (n 32) 431 (original emphasis removed).

66

Alina Trapova, ‘Copyright for AI-Generated Works: A Task for the Internal Market?’ (Kluwer Copyright Blog, 8 February 2023) <https://copyrightblog.kluweriplaw.com/2023/02/08/copyright-for-ai-generated-works-a-task-for-the-internal-market/> accessed 8 April 2023. See also Vincenzo Iaia, ‘To Be, or Not to Be … Original Under Copyright Law, That Is (One of) the Main Questions Concerning AI-Produced Works’ [2022] GRUR International 793, 799; Silke von Lewinski, ‘L’Intelligence Artificielle et le Droit d’Auteur’ in Entre Art et Technique: Les Dynamiques du Droit. Mélanges en l’Honneur de Pierre Sirinelli (Dalloz 2022) 149.

67

Craig and Kerr (n 26) 72.

68

For more details on the significance of general authorial intent for copyright protection eligibility, see European Commission, Directorate-General for Communications Networks, Content and Technology, and others (n 16) 75, 82-83.

69

See European Commission, Directorate-General for Communications Networks, Content and Technology, and others (n 16) 75.

70

See Gaon (n 7) 249, 255.

71

Pamela Samuelson, ‘Allocating Ownership Rights in Computer-Generated Works’ (1986) 47 U Pitt L Rev 1185, 1206-07.

72

Ginsburg and Budiardjo (n 32) 378.

73

See Samuelson (n 71) 1206-07; Ginsburg and Budiardjo (n 32) 378; Spindler (n 12) 261.

74

See Tim W Dornis, ‘Artificial Creativity: Emergent Works and the Void in Current Copyright Doctrine’ (2020) 22 Yale JL & Tech 1, 50.

75

European Commission, Directorate-General for Communications Networks, Content and Technology, and others (n 16) 6.

76

Ginsburg and Budiardjo (n 32) 424.

77

Samuelson (n 71) 1202.

78

Ginsburg and Budiardjo (n 32) 419.

79

See Senftleben and Buijtelaar (n 5) 807; Williams (n 25) 790.

80

Sam Ricketson and Jane C Ginsburg, International Copyright and Neighbouring Rights: The Berne Convention and Beyond (3rd edn, OUP 2022) 376.

81

See Ginsburg and Budiardjo (n 32) 426-28.

82

See European Commission, Directorate-General for Communications Networks, Content and Technology, and others (n 16) 116; Hugenholtz and Quintais (34) 1204.

83

See Antikainen (n 28) 259.

84

Ricketson and Ginsburg (n 80) 374.

85

See Spindler (n 12) 262.

86

Samtani Anil and Abigail Lim Chiu Mei, ‘Who is Thy Author? Recommendations to Integrate “Machine-Authored” Works into the Copyright Domain in Singapore’ (2022) 44(1) EIPR 5, 10.

87

See Samuelson (n 71) 1221-24.

88

Ricketson and Ginsburg (n 80) 365.

89

ibid 365.

90

ibid 376.

91

See European Commission, Directorate-General for Communications Networks, Content and Technology, and others (n 16) 76.

92

See Anil and Mei (n 86) 10-12.

93

See Samuelson (n 71) 1223 pointing to the lack of ‘direct dealings’ between users and designers. See also Robert Yu, ‘The Machine Author: What Level of Copyright Protection is Appropriate for Fully Independent Computer-Generated Works?’ (2017) 165 U Pa L Rev 1245, 1259-60; Ginsburg and Budiardjo (n 32) 416.

94

See Ricketson and Ginsburg (n 80) 376 stating that ‘in many cases of computer-enabled outputs, the programmer and user would have no knowledge of one another’.

95

See Ricketson and Ginsburg (n 80) 376 considering that ‘contributions [of designers and users] will likely be a-synchronous’. See also Samuelson (n 71) 1223 specifying that ‘the user typically will use the generator program at a site remote from the programmer, and at a time when the programmer has no involvement in the work done by the program’s user’.

96

See European Commission, Directorate-General for Communications Networks, Content and Technology, and others (n 16) 84-85.

97

Anil and Mei (n 86) 10.

98

Samuelson (n 71) 1222.

99

See Ricketson and Ginsburg (n 80) 376-77.

100

Samuelson (n 71) 1223.

101

See European Commission, Directorate-General for Communications Networks, Content and Technology, and others (n 16) 85.

102

See eg ‘DeepL Pro – Terms and Conditions’ (DeepL) <https://www.deepl.com/ru/pro-license?tab=pro> accessed 8 April 2023; ‘Terms of use (OpenAI)’ <https://openai.com/policies/terms-of-use> accessed 8 April 2023.

103

See James Grimmelmann, ‘There’s No Such Thing as a Computer-Authored Work – And It’s a Good Thing, Too’ (2016) 39 Columbia JL & Arts 403, 408. See also Ginsburg and Budiardjo (n 32) 398.

104

See Kim and others (n 14) 67 arguing that ‘any computational operations implemented on a computer can be done by a human on paper’. However, calculations performed by AI systems would take a lifetime or even more for a human who would try to do them manually without a computer. See Williams (n 25) 788, who reached the same conclusion. See also Grimmelmann (n 103) 407 stating that ‘[computationally complicated operations] are really just (much) faster versions of operations that could be carried out by hand’.

105

See Samuelson (n 71) 1208; Jane C Ginsburg, Burrow-Giles v. Sarony, (U.S. 1884): Copyright Protection for Photographs, and Concepts of Authorship in an Age of Machines (Twelve Tables Press 2020) 273.

106

Deltorn and Macrez (n 13) 21.

107

Samuelson (n 71) 1209.

108

Ginsburg, Burrow-Giles V. Sarony (n 105) 273.

109

See Jane C Ginsburg, ‘People Not Machines: Authorship and What It Means in the Berne Convention’ in Graeme W Austin and others (eds), Across Intellectual Property: Essays in Honour of Sam Ricketson (CUP 2020) 87; Ricketson and Ginsburg (n 80) 375-76.

110

See European Commission, Directorate-General for Communications Networks, Content and Technology, and others (n 16) 116; Hugenholtz and Quintais (34) 1204.

111

See Samuelson (n 71) 1201; Gabriele Spina Alì, ‘The Times They are AI-Changin’: Copyright and Computer Generated Works’ [2018] AIDA 367, 391; Baptiste Lecharny, Eve Schenberg and Benoît Seville, ‘La Propriété Intellectuelle à l’Aune de l’Avènement de l’Intelligence Artificielle’ (2022) 86 Revue du Droit des Technologies de l’Information 5, 16; Spindler (n 12) 260.

112

Ginsburg and Budiardjo (n 32) 352. See also Pablo Fernández Carballo-Calero, 25 Things You Should Know about Artificial Intelligence Art and Copyright (Thomson Reuters Aranzadi 2022) 102 stating that ‘the conception does not imply a correlation between the expectations and the result’.

113

See Ginsburg and Budiardjo (n 32) 360; Fernández Carballo-Calero (n 112) 101; von Lewinski (n 66) 149.

114

Samuelson (n 71) 1201. See Ginsburg and Budiardjo (n 32) 360 pointing out that ‘the less formed the initial ideas and the less influence the putative principal author exercises over the process of execution, the less likely will sole, or even any, authorship be attributed to the person claiming to have conceived the work’. See also Ginsburg, ‘People Not Machines’ (n 109) 86-87; Ricketson and Ginsburg (n 80) 375.

115

See Boyden (n 28) 391. See also European Commission, Directorate-General for Communications Networks, Content and Technology, and others (n 16) 81 summarising that ‘personal selection undoubtedly contributes to a finding of originality of an AI-assisted output’.

116

See Ginsburg and Budiardjo (n 32) 368-69, 426. See also Boyden (n 28) 391; Spina Alì (n 111) 374-75; Anne Lauber-Rönsberg and Sven Hetmank, ‘The concept of authorship and inventorship under pressure: Does artificial intelligence shift paradigms?’ (2019) 14(7) JIPLP 570, 574; Dornis, ‘Of Authorless Works’ (n 12) 581; von Lewinski (n 66) 149. For an alternative view, see Williams (n 25) 790, who considers that judgement over the final output is a creative decision which law should acknowledge. See also European Commission, Directorate-General for Communications Networks, Content and Technology, and others (n 16) 81 sketching Kummer’s ‘presentation theory’ according to which ‘the mere act of selecting a pre-existing object suffices to convert the object into a work’.

117

See Hugenholtz and Quintais (34) 1203, who, after distinguishing three phases (conception, execution and redaction) of the creative process assisted by AI, arrive at the conclusion that, ‘depending on the circumstances, creative choice in the redaction phase may even suffice for a finding of originality of the entire product’.

118

See United States Copyright Office, ‘Cancellation Decision re: Zarya of the Dawn’ (VAu001480196, 2021) 1, 10. After examining the images which the applicant claimed to have edited, the United States Copyright Office concluded that ‘the editing alterations [were] not sufficiently creative to be entitled to copyright’.

119

The idea/expression dichotomy is a fundamental principle of copyright law. According to it, copyright law ‘withholds protection from ideas and other building blocks of literary and artistic expression: words and basic plots and themes in the case of literary works, individual notes and common themes in music, color and perspective in visual art’. See Paul Goldstein and P Bernt Hugenholtz, International Copyright: Principles, Law, and Practice (4th edn, OUP 2019) 203. This principle has been a subject of debate for a number of years. Its opponents would point to its lack of clarity and the difficulties in applying the principle consistently. In turn, its supporters would consider the principle itself as a ‘safeguard’ against overprotection, and its flexible nature as an advantage. For the critique, see eg Eleonora Rosati, ‘The Idea/Expression Dichotomy: Friend or Foe?’ in Richard Watt (ed), Handbook on the Economics of Copyright: A Guide for Students and Teachers (Edward Elgar 2014). For the principle’s justification, see eg Noam Shemtov, Beyond the Code: Protection of Non-Textual Features of Software (OUP 2017) 104-108.

120

The doctrine refers to a merger ‘when an idea and its particular expression become inseparable’. See Rosati (n 119) 67.

121

‘The doctrine of scènes à faire applies to those expressions which are so associated with a particular genre, motif, or idea that one is compelled to use them’. See Rosati (n 119) 67.

122

For the various types of external constraints in the EU law, see European Commission, Directorate-General for Communications Networks, Content and Technology, and others (n 16) 72-74.

123

See Fernández Carballo-Calero (n 112) 101.

124

‘Privacy Policy’ (Midjourney) <https://docs.midjourney.com/docs/privacy-policy> accessed 6 April 2023. On the social role of Midjounney, see Jonas Oppenlaender, ‘The Creativity of Text-to-Image Generation’ (Academic Mindtrek, 16-18 November 2022) <https://doi.org/10.1145/3569219.3569352> accessed 8 April 2023 describing ‘the [Midjourney] community [as] a vast social learning resource, unlocking the social creativity of what previously was a lone-creator practice’.

125

See ‘Midjourney’ (Midjourney) <https://www.midjourney.com/home/?callbackUrl=%2Fapp%2F> accessed 6 April 2023.

126

See Rob Salkowitz, ‘Midjourney Founder David Holz on the Impact of AI on Art, Imagination and the Creative Economy’ (Forbes, 16 September 2022) <https://www.forbes.com/sites/robsalkowitz/2022/09/16/midjourney-founder-david-holz-on-the-impact-of-ai-on-art-imagination-and-the-creative-economy/> accessed 6 April 2023.

127

This article provides a simplified account of using Midjourney. For a more detailed guide, see ‘Quick Start’ (Midjourney) <https://docs.midjourney.com/docs/quick-start> accessed 20 May 2023.

128

See Eray Eliaçık, ‘AI Prompt Engineering is the Key to Limitless Worlds’ (Dataconomy, 24 February 2023) <https://dataconomy.com/2023/01/what-is-ai-prompt-engineering-examples-how/?utm_content=cmp-true> accessed 8 April 2023.

129

If a prompt consists of image and text parts, the user has to use/imagine command. If a prompt includes solely images, the user can use either/imagine or/blend command. See ‘Blend’ (Midjourney) <https://docs.midjourney.com/docs/blend> accessed 12 May 2023.

130

‘Quick Start’ (n 127).

131

For the purposes of using Midjourney, upscaling is referred to as ‘the process of increasing the resolution and/or size of one of the images [that a user has] generated’. It is also argued that upscaling can ‘add lots of new details that can alter the overall look and feel’. See Christian Heidorn, ‘Explained: Midjourney Upscale Methods’ (Tokenized, 10 May 2023) <https://tokenizedhq.com/midjourney-upscale-commands/> accessed 12 May 2023.

132

A search performed by the author of this article could not identify authoritative information in this respect.

133

See Ahfaz Malik, ‘Let’s Talk Midjourney’ (LinkedIn, 6 March 2023) <https://www.linkedin.com/pulse/lets-talk-midjourney-ahfaz-malik> accessed 6 April 2023. See also Finnn, ‘I Love and Hate Midjourney’ (Polycount, September 2022) <https://polycount.com/discussion/231283/i-love-and-hate-midjourney> accessed 6 April 2023, where a polycount platform user mentions that Midjourney uses a convolutional neural network.

134

Andersen and others v Stability AI Ltd and others (ND Cal 2023) 3:23-cv-00201, Document 1, para 34. In its Notice of Motion and Motion to Dismiss Plaintiffs’ Complaint and to Strike Class Claims. Midjourney, Inc. points out that such allegations are without actual or stated basis. See Andersen and others v Stability AI Ltd and others (ND Cal 2023) 3:23-cv-00201, Document 52, para II(B).

135

‘Stable Diffusion’ (Stable Diffusion) <https://stablediffusionweb.com/> accessed 21 May 2023. The assumption about the use of Stable Diffusion may have arisen from an interview given by David Holz. When briefly outlining the company’s history he describes the developments against the background of which ‘working on the imagination part’ of the company started. Breakthroughs regarding diffusion models are among such events. See Salkowitz (n 126), where David Holz mentions that ‘[t]here were some breakthroughs on diffusion models, people understanding clip, openAI, that sort of thing’.

136

Salkowitz (n 126).

137

James Vincent, ‘An Engine for the Imagination’: The Rise of AI Image Generators. An Interview with Midjourney Founder David Holz’ (The Verge, 2 August 2022) <https://www.theverge.com/2022/8/2/23287173/ai-image-generation-art-midjourney-multiverse-interview-david-holz> accessed 6 April 2023.

138

Vincent, ‘An Engine for the Imagination’ (n 137).

139

ibid.

140

ibid.

141

See ‘Prompts’ (Midjourney) <https://docs.midjourney.com/docs/prompts> accessed 6 April 2023, according to which ‘very short prompts will rely heavily on Midjourney’s default style, so a more descriptive prompt is better for a unique look’.

142

In the meaning that the designers did not intend to make any particular style a default one.

143

See eg Vincent, ‘An Engine for the Imagination’ (n 137), where David Holz mentions that Midjourney favourites a face, and the designers do not know ‘why it happens’.

144

However, free trials were halted at the end of March 2023. See James Vincent, ‘AI Image Generator Midjourney Stops Free Trials but Says Influx of New Users to Blame’ (The Verge, 30 March 2023) <https://www.theverge.com/2023/3/30/23662940/deepfake-viral-ai-misinformation-midjourney-stops-free-trials> accessed 6 April 2023 reporting that ‘Midjourney has halted free trials of its service’ due to ‘extraordinary demand and trial abuse’. The Midjourney team turned on free trials to celebrate the release of a new version of Midjourney at the beginning of May 2023. However, it was a temporary offer, which lasted for just one weekend. See Matt Novak, ‘AI Image Creator Midjourney Reopens Free Trials After Month-Long Pause’ (Forbes, 5 May 2023) <https://www.forbes.com/sites/mattnovak/2023/05/05/ai-image-creator-midjourney-reopens-free-trials-after-month-long-pause/> accessed 12 May 2023.

145

Free trials are supposed to provide a general idea of how Midjourney functions and can be used. While free trial users are not limited in terms of availability of tools, they are largely restricted in terms of attempts to produce an outcome. These attempts depend on Graphics Processing Unit (GPU) time. Free trial users are provided with a very limited amount of GPU time. On Midjourney’s GPU time, see ‘Fast and Relax Modes’ (Midjourney) <https://docs.midjourney.com/docs/fast-relax#:~:text=The%20Average%20Job%20the%20Midjourney,values%20will%20take%20less%20time.> accessed 12 May 2023. For a comparison of Midjourney’s subscription plans, see ‘Subscription Plans’ (Midjourney) <https://docs.midjourney.com/docs/plans> accessed 12 May 2023. Moreover, another major difference lies in the allocation of rights on the output. As a general rule, Midjourney, Inc. recognises users’ ownership of the output. However, the output generated by users who are not subscribed to a paying plan presents one of the exceptions to this rule. Midjourney, Inc. reserves ownership of such output and grants a licence to it under the Creative Commons Noncommercial 4.0 Attribution International License. See ‘Terms of Service’ (Midjourney) <https://docs.midjourney.com/docs/terms-of-service> accessed 12 May 2023.

146

Midjourney offers several upscaling models. The choice of an upscaler may be conditioned upon the desired resolution, details and textures. See ‘Upscalers’ (Midjourney) <https://docs.midjourney.com/docs/upscalers> accessed 6 April 2023.

147

Text to nn 115-116.

148

When faced with images generated using Midjourney and used in work submitted for registration, the United States Copyright Office contended that prompts did not determine a specific result but rather provided general directions for the future outcome. This translated into the lack of users’ control ‘over the initial image generated, or any final image’. Moreover, according to the Office the fact that it took the applicant ‘significant time and effort working with Midjourney’ did not make her the ‘author’. These findings were further elaborated on by the United States Copyright Office in the statement of policy on works containing material generated by AI. While these conclusions mark a big step towards clarifying the legal uncertainties around the copyrightability of AI-based output, this author would like to explore prompting as a (non-)creative act in more detail. The article considers that creating a prompt does not necessarily imply the lack of creativity ‘invested’ in the output. For the respective findings of the United States Copyright Office, see United States Copyright Office, Cancellation Decision (n 118) 7-10; United States Copyright Office, 'Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence' (Statement of policy, 37 CFR Part 202, 2023). For an opinion challenging prompting as a creative practice, see Jon McCormack and others, ‘Is Writing Prompts Really Making Art?’ (ArXiv, 2 February 2023) <https://doi.org/10.48550/arXiv.2301.13049> accessed 27 June 2023.

149

While this article focuses on the examples of text prompts, it also acknowledges the role that image prompts can play for the purposes of assessing originality. In this respect, it might be particularly interesting to follow another application to the U.S. Copyright Office for registering AI-based output: ‘Rose Enigma’. It is claimed that, to generate the output, the applicant provided text instructions as well as an image. It is also stated that by including an image, along with choosing other settings, the applicant was able ‘to generate [using Stable Diffusion, an AI-powered tool,] an image that visually realized [the applicant’s] mental conception’. See Franklin Graves, ‘Copyright Office Listening Sessions Underway’ (LinkedIn, 4 May 2023) <https://www.linkedin.com/pulse/ai-issue-platform-disclosures-here-laion-fights-back-against-graves/> accessed 12 May 2023.

150

According to Midjourney’s Terms of Service, ‘[Midjourney’s customers] own all Assets [they] create with the Services, to the extent possible under current law’. For the exceptions to this general rule, see ‘Terms of Service’ (n 145). Although this rule refers to ownership and not authorship, it demonstrates an overall lack of a strong willingness to assert rights.

151

See ‘Shira Perlmutter Discusses Generative AI, Prompt Engineering and Copyright’ (Digital Media Licensing Association, 2 November 2022) <https://www.youtube.com/watch?v=1ZdOI2inQ4A> accessed 4 May 2023. Shira Perlmutter, the Register of Copyrights and Director of the U.S. Copyright Office, also mentions that another question may be ‘whether there would be copyright protection in the prompt itself or in the output of what the machine does with the prompt’, and ‘the suggestion is that it could be a registration [of an] eligible derivative work’. However, it is a ‘fact-dependent question’.

152

See eg Andrés Guadamuz, ‘A Scanner Darkly: Copyright Infringement in Artificial Intelligence Inputs and Outputs’ (2023) <https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4371204> accessed 8 April 2023. Moreover, copyright concerns about generative AI are among the reasons behind certain amendments to the proposal for the AI Act in the European Union. In terms of generative AI, such amendments (among other things) include an obligation to ‘document and make publicly available a sufficiently detailed summary of the use of training data protected under copyright law’. See Parliament, ‘Amendments adopted by the European Parliament on 14 June 2023 on the proposal for a regulation of the European Parliament and of the Council on laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) and amending certain Union legislative acts’ COM(2021)0206 – C9-0146/2021 – 2021/0106(COD), art 28b(4).

153

See Ginsburg and Budiardjo (n 32) 434; Ginsburg, Burrow-Giles V. Sarony (n 105) 273.

154

See Kim, ‘On Words That Come Easy’ (n 25) 434, who coins and elaborates on the concept of distributed collective intelligence.

155

The increasing complexity of AI systems is coupled with an increasing simplification of their operation.

156

Ginsburg and Budiardjo (n 32) 434.

157

Craig and Kerr (n 26) 41.

158

See Boyden (n 28) 378.