Copyright and Artificial Creation: Does EU Copyright Law Protect AI-Assisted Output?

This article queries whether and to what extent works produced with the aid of AI systems – AI-assisted output – are protected under EU copyright standards. We carry out a doctrinal legal analysis to scrutinise the concepts of “work”, “originality” and “creative freedom”, as well as the notion of authorship, as set forth in the EU copyright acquis and developed in the case-law of the Court of Justice. On this basis, we develop a four-step test to assess whether AI-assisted output qualifies as an original work of authorship under EU law, and how the existing rules on authorship may apply. Our conclusion is that current EU copyright rules are generally suitable and sufficiently flexible to deal with the challenges posed by AI-assisted output.


Introduction
As artificial intelligence (AI) is making inroads into the realm of culture, the alleged creative powers of intelligent machines are reaching almost mythical proportions.Spectacular examples abound, ranging from the impressive Next Rembrandt1 , a Rembrandt-style portrait created with the help of algorithms, to the uncannily well-crafted translations by DeepL. 2 Indeed, AI-assisted creation nowadays encompasses almost the entire spectrum of subject matter enumerated in art.2(1) of the Berne Convention.
The creative powers of advanced AI systems have led some scholars to conclude that the results of artificial creation cannot be protected by copyright, since human beings have lost control of the creative process. 3Some writers, therefore, argue for the introduction of special neighbouring rights to protect 'authorless' AI-generated productions against misappropriation. 4But is this assumption correct?Or can AI-assisted outputs qualify for copyright protection, despite the increasingly important role that machines play in their creation?
To be sure, this is not an entirely new question.As early as the 1960s scholars have grappled with questions related to computer-generated works. 5With the rise of AI, in particular machine learning algorithms, the issue has gained momentum and inspired a vast new body of legal scholarship in recent years. 6his article poses the question whether and to what extent productions created with the aid of AI systems -in short, AI-assisted outputs 7 -are protected under EU copyright standards.Central to our analysis is not the intelligent machine, but the role of human beings in the AI-assisted creative process.Is this role sufficient to elevate the result of this process, the AI-assisted output, to the status of a copyrightprotected work?
This article is based on a study commissioned by the European Commission 8 , the conclusions of which inform the policy position on AI-assisted creation adopted by the Commission in its recent IP Action Plan. 9he analysis presented here is largely doctrinal.Normative questions about the most desirable protection regime for AI-assisted outputs are not addressed.Neither do we discuss alternative regimes available under current law, such as neighbouring rights or unfair competition law. 10 To begin with, in section 2 we present a general overview of the current framework of EU copyright law as it pertains to copyright protected "works".In section 3, we turn to the core question whether and under which conditions AI-assisted outputs qualify as "works" protected under harmonised copyright standards.To this end, we closely scrutinise the notions of "work", "originality" and "creative freedom", as developed in the case law of the CJEU.As we shall demonstrate, the CJEU's decisions in Painer and Football Dataco offer important clues in answering our research question. 11In section 4 we address the issue of authorship of AI-assisted outputs, which is intimately linked to that of the notion of "work".Nevertheless, since this notion has so far remained largely unharmonized, nationally divergent solutions, such as the rules on authorship of computer-generated works in UK and Irish copyright law, have been developed. 12In section 5 we conclude.
Convention". 14In its jurisprudence, the CJEU similarly seeks guidance from art.2(1) of the Berne Convention, which through its incorporation by reference into the TRIPS Agreement and the WCT, 15 has become part of the EU legal order. 16e EU acquis expressly harmonises three -or possibly four -specific categories of copyright-protected subject matter: computer programmes, databases, photographs, and (possibly) works of visual art. 17hese qualify as protected works if they are "original in the sense that they are the author's own intellectual creation".In its 2009 Infopaq judgment, the CJEU extrapolated from this piecemeal harmonisation a general, autonomous concept of EU law of the work as "the author's own intellectual creation". 18This has been confirmed in later judgments, most recently in Levola, Funke Medien, Cofemel and Brompton Bicycle. 19

Production in literary, scientific or artistic domain
From the definition of "work" in art.2(1) of the Berne Convention follows a general requirement that works be produced within the "literary, scientific or artistic domain".Whereas some scholars have given normative meaning to this categoric notion, 20 the CJEU has not clearly embraced this "domain test" as a separate criterion.In Premier League the Court denied copyright to sporting events for reason that they "cannot be regarded as intellectual creations classifiable as works" 21 , which possibly suggests an application of the domain test.However, elsewhere in the judgement it transpires that the Court's exclusion of sporting events is based on the lack of originality. 22Similarly, in Levola Hengelo the Court could have relied on this test to deny work status to the taste of a food product.Instead, it formulated a criterion of "identifiable expression" to achieve the same result. 23It is therefore not clear whether this general test is incorporated into EU copyright law.

Human intellectual effort
Like the Berne Convention, the EU copyright acquis is primarily grounded in the tradition of author's rights (droit d'auteur): copyright protects original expression directly emanating from a human creator.The Berne Convention does not define the "author" of a work, leaving this to the contracting parties, but its text and historical context strongly suggest that "author" and "authorship" for purposes of the Convention refer to the natural person who created the work. 24This implies that copyright protection initially vests in human authors. 25This is confirmed by the Convention's ubiquitous references to the "author" as the originator of works and the beneficiary of protection.The Convention's provision on moral rights (art.6bis) that are expressly granted to "authors" underscores that its minimum standards of copyright protection are triggered only by acts of human creation.
Although EU copyright law nowhere expressly states that copyright requires a human creator, its "anthropocentric" focus (on human authorship) is self-evident in many aspects of the law. 26For one thing, the CJEU's case law on originality, discussed below, completely relies on the notion of a human being engaging in creative acts -reflecting "creative choice".As the Court considered in Painer, "[b]y making those various choices, the author of a portrait photograph can stamp the work created with his 'personal touch'." 27The point is reinforced in Cofemel: "if a subject matter is to be capable of being regarded as original, it is both necessary and sufficient that the subject matter reflects the personality of its author, as an expression of his free and creative choices". 28so, according to the CJEU, the exclusive rights harmonised accorded to the author in the InfoSoc Directive necessarily attach to a human creator, not a legal entity such as a film producer or publisher. 29rhaps the clearest formulation of this principle comes from Advocate General Trstenjak in her opinion in Painer, where she concluded from the wording of art.6 of the Term Directive, that "only human creations are therefore protected, which can also include those for which the person employs a technical aid, such as a camera". 30This conclusion was endored by the Court.
Human rights provide additional arguments in support of the proposition that copyright presupposes human authorship. 31For example, the Universal Declaration on Human Rights (UDHR) protects the moral and material interests of authors resulting from scientific, literary or artistic production.Given that human rights by definition vest in human beings, the concept of authorship under the UDHR necessarily refers to human authorship. 32n sum, the requirement of human intellectual effort excludes from copyright protection outputs that are produced without any human intervention. 33For example, the aesthetically pleasing flowers of a rose or wings of a butterfly cannot be qualified as works.Likewise, a production that is wholly generated by an AI system without any human intellectual effort is excluded from copyright protection.This requirement does not rule out, however, creations by human authors made with the aid of machines, provided that the human contribution to the output meets the legal standard of originality/creativity to which we now turn. 34

Originality/creativity
The EU standard of "the author's own intellectual creation" requires that the subject matter must be (i) "the author's own", that is, not copied, and (ii) an "intellectual creation". 35This twofold requirement usually goes by the name of "originality".As the CJEU, rather circularly, held in Levola and Cofemel: "the subject matter concerned must be original in the sense that it is the author's own intellectual creation". 36n Painer and Funke Medien, the Court clarified that intellectual creation implies originality which in turn implies making personal, 37 creative choices. 38This was more recently confirmed in Cofemel 39 and Brompton Bicycle. 40e Painer decision is particularly instructive, as it concerns subject matter created with the aid of a machine -notably, a photographic portrait.According to CJEU, a portrait photographer "can make free and creative choices in several ways and at various points in its production.[…] By making those various choices, the author of a portrait photograph can stamp the work created with his 'personal touch'.Consequently, as regards a portrait photograph, the freedom available to the author to exercise his creative abilities will not necessarily be minor or even non-existent". 41 Infopaq, a case involving the protection of newspaper articles against unauthorised scanning, the focus of the Court's originality enquiry is on "the form, the manner in which the subject is presented and the linguistic expression." 42The CJEU clarified that, for literary works, the author's "free and creative choices" pertain to the selection, sequence and combination of words. 43While admitting that words in isolation do not amount to intellectual creation, the Court adds that "it is only through the choice, sequence and combination of those words that the author may express his creativity in an original manner and achieve a result which is an intellectual creation." 44iginality or creativity do not, however, imply a requirement of artistic merit or aesthetic quality. 45EU copyright law protects works of high art as much as it protects more mundane intellectual productions, such as simple photographs, industrial design, databases or computer software.Conversely, as clarified in Cofemel, the fact that a production "may generate an aesthetic effect" is no reason to qualify it as subject matter protected under EU copyright law. 46This is an important observation in relation to AI-assisted outputs, many of which are undeniably of aesthetic value.This focus on the act of creation in terms of making free and creative choices implies that economic investment cannot, as such, justify protection.In Football Dataco, the Court squarely rejected "significant skill and labour" on the part of the producer of football fixtures lists as a relevant factor in assessing originality. 47Similarly, in Funke Medien the Court considered "the mere intellectual effort and skill of creating [military status] reports are not relevant in that regard." 48 essence, EU law's requirement of originality is met "if the author was able to express his creative abilities in the production of the work by making free and creative choices." 49In a string of cases the CJEU has had the opportunity to elaborate on the relevant parameters of such creative choices.Importantly, the Court has identified various types of external constraints to creativity: rule-based, 50 technical 51 or functional 52 , and informational 53 .All these may play a role in the legal assessment of AI-assisted outputs in distinct cases.
The CJEU does not however seem to require that the author's creativity or personality ("personal stamp") be objectively discernible in the resulting expression (the output).What appears to be sufficient is that prospectives author exercise their "free and creative choices" and thereby express their personality. 54But is that enough -or does the law additionally require that the "creative space" be creatively used as the ideas are being expressed in the final production?On the face of it, the case law does indeed suggest the latter.The Court speaks of choices that must be "creative", and that "by making those various choices, the author of a portrait photograph can stamp the work created with their 'personal touch'". 55This language suggests that exercising creative freedom in a non-creative way, e.g. by making only obvious choices, would not result in a protected work.On the other hand, as we have seen before, the requirement of originality or creativity does not entail a test of artistic merit or aesthetic quality, or that the work be novel (new).
National courts have dealt with this problem in different ways.For example, the Dutch Supreme Court expressly denies copyright protection to "trivial" or "banal" expression, even under conditions of broad creative freedom. 56The copyright cases so far decided by the CJEU do not give much guidance on how to assess the "creativeness" of the act of creating, if at all, nor do they define a minimum standard of creativity.
Early CJEU decisions suggest that if the external constraints allow an author sufficient creative freedom, then the level of creativity actually required by the Court is fairly low. 57In Infopaq the Court suggested that even a short, 11-word text fragment might qualify. 58Judging from the reasoning in Painer, the originality of a photographic work is practically a given. 59Even in a case concerning run-of-the-mill school portrait photographs, "the freedom available to the author to exercise his creative abilities will not necessarily be minor or even non-existent". 60This suggests that even a combination of fairly obvious choices in the design, execution and editing of an AI-assisted output could suffice.By contrast, in Funke Medien, the Court (in line with Advocate General Szpunar) expressed serious doubts over whether the military status reports at issue could qualify as "works", since the standard format of these reports and their purely informational purpose left (too) little room for creative choices. 61Even though the Court's reasoning in Painer and Funke Medien point in different directions, the focus of the CJEU's originality analysis in both cases is on the availability of creative choice.

Expression
A fourth prerequisite for copyright protection is that the human creator's creativity be "expressed" in the final production.The use by the author of their creative freedom must be somehow perceptible in the author's expression.Ideas that are not given shape or form cannot qualify as "works".The CJEU has on several occasions confirmed that expression is a sine qua non for copyright protection.Both in Infopaq and in BSA, the Court states that the author must have "express[ed] his creativity in an original manner". 62n Painer, the CJEU observes that, for a work to be original, the author must be able to "express his creative abilities in the production of the work by making free and creative choices." 63Similarly, in Funke Medien the Court opines that "only something which is the expression of the author's own intellectual creation may be classified as a 'work' within the meaning of Directive 2001/29". 64And in Levola Hengelo the Court has underscored that the author's creative choices must be sufficiently clearly expressed in the interest of legal certainty. 65is requirement of expression implies a causal link between an author's creative act (the exercising of their creative freedom) and the expression thereof in the form of the work produced.But it remains unclear whether and to what extent the original features of the work should (all) be preconceived or premeditated by the author.Indeed, it is fair to assume that the concept of a work as "the author's own intellectual creation" requires not only human agency or intervention, but also some degree of authorial intent. 66If a work must be "created" by a human "author" and subsequently "expressed", then this notion clearly cannot encompass wholly haphazard acts of nature, such as the shape of a flower or a solidified lava stream.
But does copyright law require specific intent of every original feature of the work, or does overall authorial intent suffice?Assuming that human authorship goes hand in hand with -and often partly relies on -fortuitous expression, such as slapdash paint drippings in a work of art, a requirement that all expressive features of the work be preconceived would be too strict -and not supported by existing law and practice.Instead, general authorial intent is probably enough.That is to say, it is sufficient that the author has a general conception of the work before it is expressed, while leaving room for unintended expressive features. 67 the end, the CJEU's focus on creative choice as the hallmark of intellectual creation suggests that it is the process of creating rather than the ensuing act of expression that is truly decisive for copyright protection, provided there is an attributable connection between the creative process and the expression. 68This conclusion is in line with EU law's rejection of artistic or qualitative merit as a relevant criterion for protection.
In sum, current EU copyright law, as interpreted by the CJEU, leaves room for the protection of AI-assisted outputs in a wide range of creative fields.As long as the output reflects creative choices by a human being at any stage of the production process, an AI-assisted output is likely to qualify for copyright protection as a "work".

Are AI-assisted outputs "works"? A four-step test
In light of the preceding analysis, we shall now examine whether AI-assisted productions can qualify as "works" protected under EU copyright law.Our focus is on outputs produced by or with the aid of an AI system.This is in line with a clear trend towards the use of general-purpose ("off-the-shelf") AI software or services for the production of creative content. 69In the following we will generally assume a "user" of an AI system not involved in its development, who produces an artefact with the aid of the system -the AI-assisted output.It is this user, and this artefact, that will be central to our copyright analysis.
As our inquiry into EU copyright law reveals a four-step test must be met for an AI production to qualify as a "work": • Step 1 -Production in literary, scientific or artistic domain; • Step 2 -Human intellectual effort; • Step 3 -Originality/creativity (creative choice);

Step 1: Production in literary, scientific or artistic domain
As noted, many AI productions resemble archetypal works, and belong to "the literary, scientific or artistic domain" without any difficulty.AI systems are capable of generating almost the entire spectrum of work types mentioned in art.2(1) of the Berne Convention, including news articles, poems, musical compositions, paintings, maps, industrial designs, geographical maps, photographs, films, et cetera.For these kinds of outputs, passing this initial test will therefore be unproblematic, assuming the domain requirement is a material prerequisite under EU law at all.

Step 2: Human intellectual effort
In addition, to qualify as a "work", the AI-assisted output must be the result of human intellectual effort.The criterion of human intervention does not however rule out AI productions as a matter of course.As the EU Court has clarified in Painer, it is entirely possible to create works of authorship with the aid of a machine or device.
Moreover, leaving aside the futuristic scenario of a completely autonomous creative robot, AI-assisted outputs will always go hand in hand with some form of human intervention, be it the development of the AI software, the gathering and choice of training data, the drawing up of functional specifications, supervising the creative process, editing, curation, post-production, etc.Even if the connection between the human intervention and the AI-assisted output is increasingly remote, at this point in time it is hard to conceive of content that is generated through AI that involves no human agency whatsoever.What is problematic today and for the immediate future is whether, and to what extent, a natural person's involvement with the AI-assisted output -however remote -is sufficient for it to qualify as an intellectual creation.This brings us to the third criterion.

Step 3: Originality or creativity (creative choice)
The third and most crucial criterion is originality or creativity.In the words of the CJEU, this test is met "if the author was able to express his creative abilities in the production of the work by making free and creative choices." 70As we have seen, the emphasis here is on the existence (a priori) of sufficient creative space, rather than on the creativity of the production as such.
As the Painer case illustrates, creative choices may occur at various levels and in different phases of the creative process: preparation, execution, and finalisation. 71Consequently, a creative combination of ideas at distinct stages in the creative process might be enough to qualify the result as a "work" protected under EU copyright.Inspired by the Painer decision, it is useful to have a closer look at the process of creating works with the aid of AI systems.As the Painer court has well understood, creativity in machine-aided production may occur at three distinct phases of the creative process, which we propose to label "conception", "execution" and "redaction". 72The figure below provides a simplified diagram of this iterative creative process.

Diagram of an iterative creative process
The conception phase involves creating and elaborating the design or plan of a work.This phase goes beyond merely formulating the general idea for a work. 73It requires a series of fairly detailed design choices on the part of the creator: choice of genre, style, technique, materials, medium, format, et cetera.It also involves conceptual choices relating to the substance of the work: subject matter (news article, portrait), plot (novel, film), melodic idea (musical work), functional specifications (software, databases), 70 Case C-469/17 -Funke Medien, para.19 ; Case C-145/10 -Painer, paras.87-8. 71Ramalho (n 26) 7. 72 See Ginsburg and Budiardjo (n 6).(discussing "detailed conception" and controlled execution).See also Ramalho (n 26) 7. (distinguishing "preparation", execution" and "final' phases in analysis of Painer judgment). 73Ginsburg and Budiardjo (n 6) 347-348.etc. 74 As the CJEU has clarified in Painer, creative choices at this pre-production stage are important factors in a finding of originality of the final production.
In the case of productions created with the aid of machine learning (ML) algorithms, above and beyond the design choices identified above, additional conceptual choices may involve the choice of AI system (e.g., the type and characteristics of the models used), as well as the selection and "curation" of input data (e.g., in the labelling of training data) and other parameters. 75With AI-assisted outputs most of these conceptual choices will be exercised by human actors.The AI system at this stage has no role in the creative process other than acting as an external constraint limiting the designer's creative possibilities.
The execution phase involves, in simple terms, converting the design or plan into what could be considered (rough) draft versions of the final work.This phase involves the producing of text, the painting of art work, the notation or first recording of music, the "shooting" of photographs or video, the "coding" of software, etc.With traditional forms of creation, the role of a human author at this execution stage is crucial.The novelist converts the plot for a novel into words, the composer translates musical ideas into notes.From the 19 th century onwards, machines have played an increasingly important auxiliary role in this creative phase.Photographs and films cannot be made without cameras, music not recorded without recording devices, etc.Nevertheless, the human author has always stayed in full control of the execution phase.That is to say, the role of the machine was essentially that of a tool in the creative process.
With AI-assisted creation this has arguably changed, in degree if not in nature.ML systems can be instructed and trained to perform complex tasks and produce sophisticated output in ways that the user of the system will not be able to (precisely) preconceive, understand or explain.From the user's perspective, this creates the impression of an autonomously operating system; one that the user does not fully control or comprehend, and that strains the classification of an AI system as a "tool".This is particularly true for Deep Learning systems, where the architecture based on several layers of neural networks greatly increases the distance between the user and the machine during the execution phase. 76ereas some AI systems are capable of generating highly sophisticated, "work-like" content at this stage of the creative process, the quality of the output should not be mistaken for proof of "creativity".What is essential for our copyright analysis is human creativity.
Finally, the redaction phase involves processing and reworking the draft versions produced in the execution phase into a finalised cultural production ready to be delivered to a publisher or other intermediary, or directly to the market.This final phase may involve a wide range of activities, depending on the genre and medium of the production.These include extensive rewriting, editing, correction, formatting, framing, cropping, colour correction, refinement and all sorts of (other) "post-production" activities that are necessary to give the final touch to the production before it is published and marketed.
Redaction is an underestimated but important , final stage in the creative process, allowing the human author many additional creative choices.As the Painer court has explained, this final phase of the creative 74  process may involve a variety of creative choices. 77Indeed, depending on the circumstances, creative choice at the redaction phase may even suffice for a finding of originality of the entire production.For example, in a case involving geographical maps directly created on the basis of unprotected satellite photographs, the French Court of Cassation accepted that the maps qualified for copyright protection because they were "the result of a personalised implementation of a complex technology by a process of transformation and improvement of choices, in particular colours, contrasts and of luminosity". 78th AI-assisted outputs this has not fundamentally changed.Even a largely autonomously operating AI system will normally not deliver output that is immediately ready for publication or commercial use.More likely, the output produced by the AI system in the execution phase will require redaction by human actors, especially when it is intended for commercial exploitation.For example, a professional musician using an AI music composer such as AIVA or MuseNet would probably rework and edit output generated by the AI system before finalising the composition. 79en not all AI-assisted productions will call for extensive redaction.For example, translation machines such as DeepL and Google Translate generate output that is almost ready to use.Nevertheless, here too some human redaction will be required to convert the output into a useful and potentially marketable professional translation.Indeed, DeepL allows its users virtually endless creative freedom in selecting and rephrasing the wording and ordering of each (part of) the translated text.
In some cases, the redaction role of the human user will be reduced to that of selecting or refusing readymade output generated by the AI system.This raises an interesting question from a copyright perspective.Clearly, the mere act of selecting may be one of many factors contributing to a finding of originality.But what if selecting one from multiple AI outputs is the only choice left to the user?Like many other questions raised by AI, this is not a novel issue.
In the past, the emergence of non-traditional art forms such as the ready-mades created by conceptualist artists, have triggered similar questions.What is it that elevates a pre-existing artefact such as a prefabricated urinal 80 or a bicycle wheel 81 to a work of art -and, by implication, to a work of authorship?According to Swiss copyright scholar Kummer, the decisive creative act here is converting the (in itself unprotectable) idea of a "ready-made" into copyright protected expression by presenting the artefact (the "objet trouvé") as a work of art. 82Kummer's "presentation theory" implies that the mere act of selecting a pre-existing object suffices to convert the object into a work.While Kummer's theory has been embraced by some copyright scholars, it remains controversial. 83In any case, personal selection undoubtedly contributes to a finding of originality of an AI-assisted output. 77Case C-145/10 -Painer. 78Supreme Court, Civil Chamber I (France) 8 January 2002 [R.I.D. A. 2002, no.193, 321] (Père-Lachaise cemetery map). 79See AIVA, https://www.aiva.ai/and OpenAI, Musenet, https://openai.com/blog/musenet/. 80Tate, Marcel Duchamp, Fountain, https://www.tate.org.uk/art/artworks/duchamp-fountain-t07573. 81 As the preceding discussion reveals, the use of highly advanced AI systems in the production of cultural goods does not imply that human beings have totally surrendered their vital role in the creative process to machines.Whereas the human creator has been partly or largely replaced by the machine in the execution stage of creative production, the human role in the conception stage remains essential, while his role in the redaction stage may have become even more important than before -given that many AIassisted outputs will probably require more redactional work than rough drafts produced by human beings.This leaves both the design choices in the conception phase and the editing and post-production choices at the redaction phase for human authors.
Moreover, it is important to realise that the three-phase creative process described above is simply a model to analyse and explain the authorial choices that contribute to a finding of originality.In reality, the creative process will be iterative.The execution phase will often yield unexpected results that inspire conceptual changes.Redaction as well may inspire new ideas that feedback to the conceptual level.In light of the Painer court's reasoning regarding machine-aided creation, which designates both conceptual choices and post-production decisions as relevant factors in the originality analysis, these choices should in many cases be sufficient for a finding of originality in AI-assisted outputs. 84is conclusion is in line with copyright rules in many national laws that allocate authorship to the person that "masterminds" (conceives) and closely supervises the execution of a work by others, without that person materially contributing to the execution phase of creation. 85In the words of Professor Ginsburg, "authorship places mind over muscle: the person who conceptualises and directs the development of the work is the author, rather than the person who simply follows orders to execute the work.Most national copyright laws agree that mere execution does not make one an author.An 'author' conceives of the work and supervises or otherwise exercises control over its execution." 86ile the CJEU has not pronounced itself on the issue of computer-generated productions, there is some case law at the national level that supports our general conclusion.For example, the Paris Court of First Instance has held that "computer-assisted musical composition, when it involves human intervention, the choice of the author [...] leads to the creation of original works". 87In the same vein, the Bordeaux Court of Appeal opined "that a work of the mind created by a computer system can benefit from the rules protecting copyright, provided that it reveals even in a minimal way the originality that its creator wanted to bring." 88

Step 4: Expression
The fourth part of our four-step test of copyright protection is that the human creator's creativity be "expressed" in the final production.As previously discussed, we derive from this criterion a prerequisite 84 See also Dreier, p. 882. 85Ginsburg & Budiardjo , p. 360.See also Ginsburg (n 24) 1072.For example, art.6 Dutch Copyright Act provides: "Where a work has been made according to the design by and under the direction and supervision of another person, that person is considered to be the author of the work. of general authorial intent: the human author must have a general conception of the work before it is expressed, while leaving room for unintended expressive features.Prima facie, this requirement might present an obstacle for AI-assisted outputs.Due to the "black box" characteristic of ML systems, the human author in charge of designing the production at the conception phase will not be able to precisely predict or explain the outcome of the execution phase.This, however, need not rule out "work" status of the final output, if such output stays within the ambit of the author's general authorial intent.Moreover, even completely unpredicted, non-explainable, quasi-random AIassisted output might still be converted into a protected "work" at the redaction phase.
What "expression" does not require is that courts engage in an assessment of the production's creative merit, aesthetic value or cultural importance.As the case law of the CJEU suggests, it is sufficient for a production to be the expression of free creative choices.

Borderline cases
The preceding analysis does not imply that all AI-assisted outputs will unconditionally qualify as copyrightprotected works under EU copyright standards.Much will depend on the facts and circumstances of a case.While sophisticated art created with the aid of an AI system, such as The Next Rembrandt portrait, will necessarily involve important human creative input at several stages of the creative process, this may not be the case for more mundane AI-assisted output such as weather forecasts and news reports.
In extreme cases, the AI system will leave its users no meaningful choice beyond pushing a few buttons.Such cases are evident in the domain of natural language generation, such as the GP-T2 and GP-T3 text generators developed by OpenAI. 89One spectacular illustration is Talk to Transformer (now InferKit), which automatically completes a text based on a text fragment (prompt) supplied by the user. 90omewhat similar tools are Deep AI's Text Generation API 91 and StoryAI 92 .Recently, OpenAI has begun to experiment with applying the "transformer" model previously used on text to images, by training it with pixels.In such cases, however, except for the user-generated prompt it will be difficult to identify any creative choice by the human user at the conception, execution or redaction phases.Consequently, the AI-assisted output generated by such systems would probably not qualify as "works". 89See: Open AI, GPT-2: 1.5B Release (5 November 2019), https://openai.com/blog/gpt-2-1-5b-release/;Tom B Brown and others, 'Language Models Are Few-Shot Learners' [2020] arXiv:2005.14165[cs] <http://arxiv.org/abs/2005.14165>accessed 28 September 2020. 90See Talk to Transformer, https://talktotransformer.com.Talk to Transformer has recently been transformed into the paid product InferKit.According to its website, "InferKit offers an interface and API for custom AI-based text generators.Whether you're a novelist looking for inspiration, or an app developer, there's something for you."See InferKit, https://inferkit.com/.For a basic explanation of how InferKit's text generation tool works, see InferKit, Docs, Text generation https://inferkit.com/docs/generation("InferKit's text generation tool creates continuations of any text you give it, using a state-of-the-art neural network.It's configurable and can produce any length of text on practically any topic.You can also create custom generators for specific kinds of content.").
91 DeepAI, Text Generation API, https://deepai.org/machine-learning-model/text-generator("The text generation API is backed by a large-scale unsupervised language model that can generate paragraphs of text.This transformerbased language model, based on the GPT-2 model by OpenAI, intakes a sentence or partial sentence and predicts subsequent text from that input") (our emphasis). 92See Storyai, About -Q&A, https://storyai.botsociety.io/about("Write the story you couldn't quite find the words to complete with this easy to use OpenAI model.Input 40 words to start, and watch what the model comes up with.It's powered by the GPT-2 774M model released on August 20th 2019 by OpenAI.") of the AI system and the users of the system.For our authorship analysis, we follow the three-phase model of creativity previously developed: conception, execution and redaction. 101As we have seen, in the case of artefacts produced with the aid of AI systems, the conception phase and -in many cases -the redaction phase will entail creative choices by human beings that justify a finding of copyright protection of the AIassisted output.Authorship in such cases is to be attributed to the person or persons individually or collectively engaging in these creative choices.If more than a single author is involved in the process, and the authors collaborate, this will lead to co-authorship, even if the creative contributions occur at different stages of the creative process.For example, the Dutch Supreme Court has held that a stylist who creatively arranged needlework pieces to be photographed for a magazine was a co-author with the photographer of the resulting photographs. 102ch of the literature on AI and copyright focuses on the scenario of the AI system producing content with only limited input on the part of the user of the system. 103If the role of the user of the system is indeed so constricted that he cannot exercise free choices at any stage of the creative process, the user will not qualify as author of the ensuing production.In such cases, the role of the user is essentially reduced to initiating a prompt (e.g., writing an initial sentence) or "pushing buttons", as in the case of the AI text generation tools discussed above.Here the user's role is somewhat comparable to that of a person playing a computer game. 104For example, authorship of film footage generated by a person playing the popular video game Grand Theft Auto most likely vests in the developers and animators of the video game -not in the player of the game.Even if the player feels empowered and "in control" of whatever transpires on the computer screen, he lacks control over the creative process, and his choices do not amount to creative acts justifying a claim of authorship. 105 regards such AI systems, where users are effectively no more than passive "players", the user clearly does not have a valid claim to authorship in the AI-assisted output (i.e., in anything beyond its initial prompt) -leaving the developer of the AI system as the only candidate for authorship of the AI-assisted output. 106Note, however, that a valid authorship claim may only arise if it is established that the output qualifies as a "work" in the first place.In the case of AI text generation tools such a finding, however, seems unlikely.The text generated by the AI system was not preconceived by the designer of the system, nor is it creatively redacted.At best one could argue that the output text is an adaptation (transformation) of the text the user input, of which the user (not the developer) is the author.between AI developers, engineers, art historians, jointly creating a work of authorship. 107If the AI system developer played a creative role in the process, he clearly deserves co-authorship status.
In many if not most cases, however, the developers of AI systems will not collaborate in a material way with the users on generating specific outputs. 108For example, providers of general-purpose AI systems or services will provide users access to their general AI capabilities, without being involved or having knowledge of the specific productions to be created with the aid of their systems.In such cases, instances of (co-)authorship by AI systems developers are unlikely to materialise, since under prevailing copyright doctrine co-authorship can only arise if the work is the result of a "concerted creative effort", i.e., if multiple authors collaborate according to a common plan to create a specific work.
Moreover, co-authorship claims by AI developers will also be unlikely for obvious commercial reasons.A developer claiming authorship (or co-ownership) of outputs generated with the aid of its system would probably not attract many customers.Assuming that AI systems will eventually become standard services or tools in the hands of business or commercial users and individual creators (similar to, e.g., Photoshop or Garage Band), the contractual terms of use of the AI system will probably resolve and preclude any such (co)authorship claims. 109Indeed, the terms of use of the popular DeepL AI-powered translation service expressly disclaim any authorship or copyright in relation to content produced by its users with the aid of DeepL. 110 the EU, allocating authorship to developers of AI systems may be further complicated by the divergent treatment of computer programmes, databases and other creative content. 111Like computer games, AI systems that generate audio-visual content are a mix of computer software, databases and (in some cases) audio-visual works.The authorship of the component parts (software, databases, other works) will rarely coalesce in a single author.It may therefore be problematic to establish (co-)authorship of the output generated by the system in those cases where the AI developer has a valid claim to (co-)authorship, that is, when the developer and the user of the AI system collaborate in producing creative output. 107See Microsoft reporter, 'The Next Rembrandt: Recreating the Work of a Master with AI' (Blurring the lines between art, technology and emotion: The Next Rembrandt, 04 2016) <https://news.microsoft.com/europe/features/nextrembrandt/>accessed 15 July 2020. 108See Samuelson (n 5) 1223-1224. 109For a brief overview of the many legal and practical complexities that authorship/ownership claims by AI developers would entail, see CLSPA, Alexandra Bensamoun and Joëlle Farchy, 'Mission du CSPLA sur les enjeux juridiques et économiques de l'intelligence artificielle dans les secteurs de la création culturelle' (CLSPA -Conseil Supérieur de la Propriété Littéraire et Artistique 2020) 39 <https://www.culture.gouv.fr/Sites-thematiques/Propriete-litteraire-etartistique/Conseil-superieur-de-la-propriete-litteraire-et-artistique/Travaux/Missions/Mission-du-CSPLA-sur-lesenjeux-juridiques-et-economiques-de-l-intelligence-artificielle-dans-les-secteurs-de-la-creation-culturelle>.
110 DeepL Pro Terms and Conditions, available at https://www.deepl.com/pro-license/.Article 7.5 of DeepL Pro's terms and conditions provides: "DeepL does not assume any copyrights to the translations made by Customer using the Products.In the event that the translations made by Customer using the Products are deemed to be protected under copyright laws to the benefit of DeepL, DeepL grants to Customer, upon creation of such translations, all excusive, transferable, sublicensable, worldwide perpetual rights to use the translations without limitation and for any existing or future types of use, including without limitation the right to modify the translations and to create derivative works." 111 The UK provisions that inspired the Irish regime are similar, but not identical. 122If the existence of a "work" is conditional upon human authorship, this statutory language seems to suggest that the Irish and British regimes allocate authorship to productions that would not qualify as original "works" according to EU copyright law standards.Whether that is, indeed, the correct reading of these provisions, is however still unclear.Since the introduction of the regime on computer-generated works in UK law in 1988, it has led to just a single court decision, which has not clarified this issue. 123 the British regime indeed protects "authorless" computer-generated works, this would imply that an AIassisted output that does not meet the standard of originality (and therefore is without an "author") could nonetheless be accorded copyright protection under Irish and UK law, with the producer ("the person by whom the arrangements necessary for the creation of the work are undertaken") as its author and copyright owner.Not surprisingly, the British and Irish regimes have been criticised as being incompatible with EU copyright standards. 124Indeed, a national rule that accords copyright protection to subject matter that does not meet the standard of "the author's own intellectual creation" is hard to reconcile with the CJEU's case law that implies that the notion of a "work" is fully harmonised and therefore does not allow national laws to accord copyright protection under more lenient conditions. 125

Conclusions
As our inquiry into EU copyright law reveals, four interrelated criteria are to be met for AI-assisted output to qualify as a protected "work": the output is (1) a "production in the literary, scientific or artistic domain"; (2) the product of human intellectual effort; and (3) the result of creative choices that are (4) "expressed" in the output.Whether the first step is established EU law is uncertain.Since most AI artefacts belong to the "literary, scientific or artistic domain" anyway, and are the result of at least some "human intellectual effort", however remote, in practice the focus of the copyright analysis is on steps 3 and 4.
Based on an analysis of the CJEU's case law, we conclude that the core issue is whether the AI-assisted output is the result of human creative choices that are "expressed" in the output.In line with the Court's reasoning in the Painer case, we distinguish three distinct phases of the creative process in AI-assisted production: "conception" (design and specifications), "execution" (producing draft versions) and "redaction" (selecting, editing, refinement, finalisation).
While AI systems play a dominant role at the execution phase, the role of human authors at the conception stage often remains essential.Moreover, in many instances human beings will also oversee the redaction stage.Depending on the facts of the case, this will allow human beings sufficient overall creative choice.Assuming these choices are expressed in the final AI-assisted output, the output will then qualify as a copyright-protected work.By contrast, if an AI system is programmed to automatically execute content without the output being conceived or redacted by a person exercising creative choices, there will be no "work".
Due to the "black box" nature of some AI systems, persons in charge of the conception phase will sometimes not be able to precisely predict or explain the outcome of the execution phase.But this does not present an obstacle to the "work" status of the final output if that output stays within the ambit of the person's general authorial intent.Authorship status will be accorded to the person or persons that have creatively contributed to the output.In most cases this will be the user of the AI system, not the AI system developer, unless the developer and user collaborate on a specific AI production, in which case there will be co-authorship.If "off-the-shelf" AI systems are used to create content, co-authorship claims by AI developers will also be unlikely for merely commercial reasons, since AI developers will normally not want to burden customers with downstream copyright claims.We therefore expect this issue to be clarified in the contractual terms of service of providers of such systems.
In conclusion, we believe that the EU copyright framework is generally suitable and sufficiently flexible to deal with the current challenges posed by AI-assisted creation.Producers of AI-assisted outputs will in many cases enjoy copyright protection.Moreover, 'authorless' AI productions might still qualify for protection against misappropriation under less demanding IP regimes, such as neighbouring rights and sui generis database protection, 126 or other doctrines such as trade secrets and unfair competition. 127In this light, regulatory intervention to extend copyright protection beyond the current EU rules would be justified only if solid empirical economic analysis were to reveal that the absence of protection harms overall economic welfare in the EU.