Is AI Art Theft? The Moral Foundations of Copyright Law in the Context of AI Image Generation

Abstract

The recent swell of public interest in AI image generating software, such as Midjourney, DALL-E 2, and Stable Diffusion, has led to a great deal of consternation among conventional visual artists. Understanding that the process through which these machines generate images depends ultimately on a machine learning process that involves the use of copyrighted artworks, has led many artists to allege that AI art is theft. There has already been a substantive debate in the literature concerning whether this use of copyrighted material by AI image generating software constitutes copyright infringement in the United States. However, the artists’ complaint appears to be a moral claim, rather than a legal one. In this paper, I will argue that these artists are essentially mistaken. Their copyrights should not protect them from this use of their art by AI image generating software. While I am sympathetic to the claims that this technology will harm visual artists, and that there should be some kind of protection afforded to artists against that harm, that harm is not caused by anything analogous to theft.

The recent swell of public interest in AI image generation software, such as Midjourney, DALL-E 2, and Stable Diffusion, has led to a great deal of consternation among conventional visual artists.¹ Understanding that the process through which these machines generate images depends ultimately on a machine learning process that involves the use of copyrighted artworks, has led many artists to allege that “AI art is theft”.² There are many ways that this allegation could be understood.³ In this paper, I consider an interpretation of this claim rooted in a conception of how copyright law ought to be: that because artists did not give the developers of AI image generation software permission to use their artworks in the development of that software, producing AI generated images with that software wrongs them, and the law ought to protect them from this wrong by treating this behaviour as copyright infringement. I argue that this claim is mistaken: we should not change copyright law to make this kind of use of a copyrighted artwork copyright infringement, because AI art is not analogous to theft.

To motivate this interpretation of the artists’ complaints: while there has already been a substantive debate in the literature concerning whether this use of copyrighted material by AI image generation software constitutes copyright infringement in the United States,⁴ when artists claim that their works are being stolen the claim appears to be a moral claim, rather than a legal one. It would be uncharitable to understand these artists as complaining that this use of their works is legally theft, because that is transparently not the case. The claim instead speaks to the feeling that AI image generation would not have been possible without the work of antecedent artists, and that those artists will be worse off because of this use of their copyrighted works without their permission.

In defence of the intuition that artists ought to have property rights that protect them from this use of their art, it is a fairly commonplace moral intuition that property rights have a moral foundation beyond the positive law. Some political philosophers, such as John Locke, think that the central purpose of the state is to protect property rights, which exist prior to positive law.⁵ H.L.A. Hart noted that if people were invulnerable to violence then we might not need laws against murder and battery.⁶ In a similar vein, artists might reasonably argue that technological change can create ‘loopholes’ in the law, making us newly vulnerable to harms in ways uncontemplated by law, and therefore not forbidden by it. From the perspective of these artists, developments with AI image generation have done just that: artists are entitled morally to a copyright that protects them from this new use of their works, and so the ‘loophole’ must be closed: the law must change to provide them with that protection.

In this paper, I shall argue that this argument is essentially mistaken. Their copyrights should not protect them from this use of their art by AI image generation software. We ought not to revise the law to be in-line with some artists’ moral intuitions about what ought to be a violation of their property rights because, as I shall show, given the nature of the process for generating images with AI, the way that AI art harms artists is not morally different from the way that competition from human artists harms them. As a result, the contemplated remaking of copyright law would have to make parts of the normal artistic process of production for human beings a violation of the property rights of antecedent artists as well.

While I am sympathetic to the claims that this technology will harm artists, and that there should be some kind of protection afforded to artists against that harm, that harm is not caused by anything analogous to theft.

In Section 1, I will give a brief and generalized explanation of how AI image generation software works. In Section 2, I will outline three complaints an artist might have about how AI image generation software harms them: the economic displacement complaint, the free-riding complaint, and the stylistic control complaint. In Section 3, I will review two strains of thought about what the ‘point’ of copyright law is—the Anglo-American conception and the French-Kantian conception. In Section 4, I make a preliminary point relevant to both of these conceptions of the point of copyright law: that the AI image generation process does not involve any use of ‘copies’ of artworks. In Section 5 I will argue that on the Anglo-American conception, treating the AI use of copyrighted images as infringing would essentially involve treating the copyrights in that art like patents, and there are important reasons why this should not be done. In Section 6 I will argue that on the French-Kantian conception as well, we cannot vindicate the claim that an artist’s copyright ought to protect them from the use made of their art by AI. This is because there is no way to do so without adopting a picture of what an author has the right to control that is overly expansive, such that it would forbid a great deal of the artistic production done by human beings.

1 Background: AI Image Generation

For the purpose of my analysis in this paper, we will only need a very simplified and non-technical picture of how programs such as Midjourney, DALL-E 2, and Stable Diffusion, use generative AI to produce new images. For each of these programs, the end user supplies a prompt describing the kind of image they would like in text, and through a complex process the program produces a new never-before-seen image that attempts to match the description given.⁷

The AI image generation software depends, roughly, on a “text encoder,” and an “image decoder”. The text encoder turns the text of the end user’s prompt into what I will call a ‘transition output’.⁸ This transition output can then be decoded by the image decoder into images. None of the image encoder, the transition output, nor the image decoder, contain images.⁹

The reason AI image creation software works in this way is because of the process through which the image encoder and image decoder are created. These tools are created through a machine learning process that takes as an input vast quantities of “text-image pairs,” which are taken from publicly accessible archives of websites. The text-image pairs pair together an image, and text that is associated with it or describes it, such as the images ‘alt text’, and other text found on the web page in which a given image is embedded. The machine learning process develops pairs of “text encoders” and “image encoders” that create closely matching ‘transition outputs’ from text-image pairs.¹⁰

Approximately, the image decoder used by the AI image generation software attempts to reverse the process performed by the image encoder—starting with the transition output, and creating images that could have produced that transition output if fed into the image encoder. This ‘diffusion’ process involves giving the image decoder a picture of randomized noise, to which the decoder applies techniques for removing noise from the image, guided by the target of creating an image that could generate the transition output which the text encoder has made of the end user’s prompt. With an accurate image decoder, AI image generation software can then ask for a prompt, encode it using the software’s text-encoder into transition output, and then decode the transition output into images.¹¹ Thus, when an end-user uses a tool like DALL-E or Midjourney, the process that the AI image generation software uses to produce an image from their prompt involves no copyrighted images.

The process used to create the AI image generation software depends on the use of vast quantities of copyrighted images scraped from the internet, without the copyright holders’ knowledge or permission.¹² The relationship between the AI image generation software and these images is such that the AI image generation software could never have been produced without them. However, and importantly to my analysis, the AI image generation software itself contains no database of images, and does not depend on access to a database of images for its operation.¹³ While copyrighted images are an essential input in the production of the AI image generation software, they are not a component part of the AI image generation software or used by its process for creating new images.

2 The Nature of An Artist’s Complaints About Generative AI

In order to investigate this issue, we will need to have some idea of the complaints that an artist might make about how AI generated images harm them. In this section I will propose three such complaints. To help frame these complaints, let us imagine the perspective of a reasonably popular digital artist whom I shall call David Davidson. David routinely shares his art on social media, and his notoriety allows him to make a modest living creating art on commission. Recently, people have been posting facsimiles of his work produced by Midjourney in the same social media communities where his fans celebrate his work. These posters are amazed that Midjourney can produce such facsimiles with the right prompt. One fan posts an image produced by the prompt “a Valkyrie riding on the back of a griffon, in the style of David Davidson”, and David notices that it is in many respects similar to a piece of art featuring a Valkyrie riding on the back of a griffon he has already produced and shared on social media. In an unnerving experience, another fan who has recently commissioned David to draw Sherlock Holmes playing chess with Watson, has posted Midjourney’s attempt at producing an image of this scene in David’s style. David finds that Midjourney’s attempt, produced in mere seconds, is in many respects similar to his nearly completed work in progress of the same scene, on which he has spent hours. David, who before this point knew very little about AI generated art, then learns that Midjourney was created using all of the works he had shared on social media up to approximately a year ago (among many other works which he did not create). He of course never gave any AI company permission to use his works in this way. What reasons might David give to explain why he has been wronged?

First, the appearance of this new way of producing art in David’s distinctive style is threatening to his business. Immediately, some of those fans posting these AI facsimiles might have otherwise commissioned him to create these images. That is money that he has directly lost the opportunity to earn.¹⁴ Less immediately, the ability of a machine to produce works that, at least to most of David’s fans, are hardly distinguishable from his artworks, is likely to cheapen his art in the view of his customers. A reasonably good substitute for what he produces is now available essentially for free to anyone. In addition to the lower price, Midjourny produces the image instantaneously. Midjourney might not be able to offer a superior product in the market for David-Davidson-Style-Art, but nevertheless David can anticipate being outcompeted on price and turnaround time. It is as though he is selling hand-sewn clothes, and a competitor has arrived with clothes made by a sewing machine. If this happens, he may have to stop producing art professionally and take up another line of work. This seems unfair, as he has invested a lot of time and effort in developing the skills of a professional artist. Let us call this the economic displacement complaint.

Second, David appreciates that Midjourney would never have developed the ability to produce artworks in the David Davidson style without the use of David Davidson’s art in the process through which Midjourney was created. Midjourney is free-riding off of the work that David Davidson has put in.¹⁵ He invested in the creation of his distinctive artistic style, and he built a community of admirers for this style of artwork. Midjourney did not pay-in to the development of this product or the creation of this market for it, but it stands to capture a significant portion of the pay-out that David Davidson was expecting. Let us call this the free-riding complaint.¹⁶

Third, David Davidson has lost the control that he once enjoyed over the body of David-Davidson-style works of art. In the short term, David is vulnerable to impersonation. Let us suppose that some of his fans have already used Midjourney to create posters advertising political events for causes that David Davidson is vehemently opposed to. Now some others might be confused into believing he supports these causes, and the idea that others might think this of him is deeply upsetting to him.¹⁷ In the longer term as Midjourney facsimiles of his artwork proliferate he will not be vulnerable to impersonation in this way, but he still will have lost creative control over the body of works of art in the David Davidson style. David does not want the corpus of David Davidson-style artworks to include artwork promoting certain messages or depicting certain subjects or themes. He used to be in a position to control this, and now he is no longer. Let us call this the stylistic control complaint.

Now that we have these complaints in mind, we can settle whether the moral intuitions behind these complaints demonstrate that the use of antecedent art that AI image generation software makes is tantamount to theft. If they are, we ought to remake our copyright law to redress this. As I shall ultimately show, these moral intuitions do not call for that. In the following sections, I shall explain two conceptions of what copyright law is for, and how those conceptions speak to the artists’ complaints described in this section.

2.1 The Economic Displacement Complaint & Copyright Law

But before describing these two conceptions of copyright law, it will be valuable to discuss the economic displacement complaint. It is not credible that this is a complaint that copyright law is meant to redress, because economic displacement seems to affect all sorts of workers irrespective of whether their work produces copyrightable material. The artists’ complaint about economic displacement is more or less the same complaint about technological changes leading to the devaluing of previously valuable skills made by various groups throughout history, going back at least as far as the Luddites of the 19th century. Arguably, this sort of complaint about economic displacement resulting from technological change is just one species of the more general complaint frequently made by socialists, such as G. A. Cohen, about market competition.¹⁸ It would be beyond the scope of this article to pass judgement on the merits of this complaint in and of itself. However, it is evident that the purpose of copyright law is not to redress this broader complaint about market competition or the narrower complaint about having one’s livelihood undermined by technological change. On any plausible theory of what copyright law is about, copyright law ought not protect weavers whose skills are made less economically valuable by the introduction of power looms, because this change does not involve taking or using their ideas or property. Additionally, it is just not morally plausible that workers who do, in the course of their work, produce copyrightable material (such as visual artists), are uniquely or particularly deserving of protection from this kind of disruption arising from technological change.

It is possible for each of these three artists’ complaints to be morally legitimate even if they turn out to have nothing to do with copyright law or theft. The economic displacement complaint is essentially the problem that Trystan Goetze, in his paper “AI Art Is Theft”, identifies with AI art.¹⁹ Goetze argues that the difference between the use that AI makes of antecedent art and the use that humans do is essentially one of rapidity and scale, and notes that this creates a problem related to distributive justice for the artists that are displaced by this practice.²⁰ Goetze, in explaining how this relates to the theft allegation, notes that artists may feel that they have a right to access and make profitable use of this common pool of antecedent artworks, and that the value of that right is being taken away from them (indeed, they may feel that it is being stolen) by the overuse of this cultural resource that AI art generation makes.²¹ It seems very plausible to me that artists have a complaint from the perspective of distributive justice, but the over-use of this cultural resource is not meaningfully analogous to the ‘theft’ of particular works. The economic displacement complaint that artists have is not morally different from other workers whose livelihoods are threatened by new technologies.

The primary claims I aim to make in this paper are about whether or not copyright protections ought to be extended such that the artists’ complaints can be redressed by them. The concerns of human visual artists about economic displacement perhaps ought to be redressed by law, but copyright law is not the right tool for this job.

3 The Point of Copyright Law: Two Conceptions

Now that we have a picture in mind of the kinds of grievances visual artists have with generative AI, we are in a position to answer questions about how the law should relate to them. Morally speaking, should these complaints be redressed by copyright law? In this section, I will review two conceptions of the point of copyright law, the Anglo-American conception, and the French-Kantian conception, to later analyze in detail how well these conceptions answer to the artists’ grievances developed in the previous section.

First, the Anglo-American conception. On this conception, the justification for copyright law is instrumental: it conceives of these new property rights as invented and disseminated for the sake of promoting this useful form of production that would otherwise, in the absence of state protection, be underproduced. Copyright law in this tradition is concerned with balancing the commercial incentives for authors to produce against the public interest in access to useful information—after all, the ultimate end of this production is the dissemination of useful information to the public. This conception is very influential over the copyright law of the United Kingdom, and the United States.²²

This justification seems to line up reasonably well with the free-riding complaint that I have identified in the previous section. Some scholars of copyright law view judges as highly sympathetic to plaintiff claims of free-riding in the context of interpreting copyright law.²³ If the point of copyright law and patent law is to promote inventive and artistic work by allowing creators to recover the up-front costs of their creation through a limited monopoly, then free-riding off of the creations of artists and inventors would seem to be exactly what copyright law ought to forbid. This claim will be investigated in more detail in the following section, where I shall argue that on the Anglo-American conception, construing AI image generation as necessarily copyright infringing would involve treating copyrights like patents in an objectionable way.

Next, the French-Kantian tradition: on this conception, authors have an intrinsic morally relevant relationship to their works that makes it possible for others to wrong them by failing to respect their rights as an author. It is because of the possibility of these moral wrongs that there ought to be legal copyright protection.²⁴ It is not merely a matter of public policy, but moral necessity, that the author of a copyrighted work is the person who has made the creative choices that defined the work.²⁵ It is for this reason that some scholars argue that it would be immoral for humans to own the copyright to AI-generated images: because they are not the authors of those works in the above sense.²⁶

Immanuel Kant develops a defence of this conception of copyright law, first and primarily in his 1785 essay “On the Wrongfulness of Unauthorized Publication of Books”.²⁷ Kant conceived of a published book as a vehicle for authorial speech. Because publishing a book is an act of communication by its author, publishing a book without its author’s permission is disseminating the author’s speech under their name and is thus a kind of compelled speech.²⁸ Kant claims that the speech which the book represents “belong[s] exclusively to the person of the author, and the author has in them an inalienable right always himself to speak through anyone else, the right, that is, that no one may deliver the same speech to the public other than in his (the author’s) name”.²⁹ We can see here that Kant conceives of copying books in this way as immoral without reference to its relation to the positive law.

The stylistic control complaint articulated in the previous section seems like it ought to be redressable by copyright law on the French-Kantian conception. Plausibly David Davidson, as the author of the David-Davidson style of artwork, ought to have some kind of power over how this style is used, in a way that plausibly precludes the use of AI tools to mimic it. This claim will be investigated in more detail in Section 5.

4 Preliminary Argument: AI Image Generation Does Not Involve ‘Copying’

Before I address how the above two conceptions of the purpose of copyright law apply to the artists’ complaints, I shall make a preliminary point that will be relevant to my analysis in both subsequent sections. Because AI image generation software does not contain copies of any copyrighted images, or access copyrighted images in the process of creating an AI generated image, the production of any particular AI generated image does not involve ‘copying.’

As Michael Murray argues in “Generative AI Art: Copyright Infringement and Fair Use,” AI image generation does not necessarily involve copyright infringement, at least under US law, because the process of creating AI generated images does not involve copying. In order for copyright infringement to take place in the United States, the alleged infringer “must have made an unauthorized copy of original elements of the valid, copyrightable work owned by the [copyright holder]”.³⁰ The core of Murray’s argument is that AI image generation software does not contain any copyrighted works or make use of copyrighted works in its production of new images. Murray notes that the word “copy” is not defined by the relevant US statute, as he puts it,

“[w]hat is relevant about this requirement is that for copyright infringement to be established, the defendant must have created a work that is expressive and fixed that can be compared side-by-side with the allegedly infringed work. And with that observation, we are now so far removed from the actual workings of visual generative AI training and the process of image generation that it strains the imagination to discuss these requirements.”³¹

If the AI image generation software’s text encoder, image decoder, or any transition output it produces, were visually compared to any given copyrighted image on which it allegedly infringes, it would not resemble that image at all. This is because these components are not copyrighted images or sets of such images, nor do they contain or represent them.

By analogy, imagine that I went to the sculptures section of the Louvre, inspected each sculpture carefully, and wrote into a journal a very detailed general description of the properties of the sculptures which I saw. I plan to sell this journal to my friend, an aspiring sculptor, who thinks it would be useful to her in the production of her own works. My journal notes the range of heights, materials, motifs, evoked emotions, curves, and shadows cast by the sculptures, among other things, and how these properties are correlated with each other. We can see that this journal that I have produced is in no sense a copy of any of the sculptures I have observed, nor could any particular excerpt of what I have written be construed to be a copy of any of the sculptures I have observed. What this journal communicates is different from what any of these sculptures communicates in many substantial ways including: the content of the communication, the form of the communication, and the purpose of the communication. In these respects, my journal is related to the sculptures from which it is derived in the same way that the text encoder and the image decoder, relate to the copyrighted images which were involved in the process through which they were created.

If Murray is correct that AI image generation software does not contain or make use of copies of any copyrighted works to produce new images, then the AI image generation software does not necessarily infringe on the copyright of those whose images the creation of the AI image generation software depends on. Likewise, by producing images using this AI image generation software, the purveyors of image-generating software such as Midjourney and Stable Diffusion do not necessarily infringe on anyone’s copyright by fulfilling an end user’s request for an image.³²

Murray argues that this is the case in the context of US law, but given the fundamentalness of the proposition that ‘copyright infringement’ necessarily involves some kind of copying or publicizing of a protected work to copyright law, his argument likely applies far more widely than the law of the United States. For example, in the statute governing German copyright law, the exclusive rights that a copyright holder has to exploit the copyrighted work are each defined as to require either copying or publicly communicating the work.³³ The language in the German statute closely corresponds to the language concerning the same rights in the Berne Convention, an international treaty governing copyright law, to which 181 states, including all European Union member states, are parties.³⁴

While this is not the only plausible argument demonstrating that AI image generation is not copyright infringing,³⁵ as I shall demonstrate in the subsequent two sections, whether or not AI generated images are produced through a process that involves copying is not merely legally relevant in some jurisdictions, but represents a morally relevant difference as far as the ‘AI art is theft’ allegation is concerned.

5 On the Anglo-American Conception, Should Copyright Law Redress the Artists’ Complaints?

On the Anglo-American conception, the purpose of copyright law is to promote the creation and dissemination of socially valuable works by incentivizing artists and others to create them. At the surface level, it is clear from my example in Section 2 that human artists, especially established ones, would benefit greatly if it were to be determined that the production of AI generated images involves copyright infringement in virtue of its use of their copyrighted works in the development of AI image generation software. Assuming that this finding would make the production of AI generated images at their present level of quality impossible (at least in the short term) artists would not have to worry about free-riding robots undermining the value of their productive endeavours. Some scholars have described copyright law as useful and important precisely for the sake of addressing these sorts of concerns about free–riding.³⁶

However, in this section, I will conclude that on the Anglo-American conception, we should not think that copyright law is the right tool to address the artists’ complaints. This is because the Artists’ free-riding complaint described in Section 2 extends the protections that copyright offers in a way that is akin to the protection offered by patents, and as I shall argue there are important reasons to think that these copyrights ought not to be treated like patents.

5.1 Copyrights & Patents

To illustrate this point, let us briefly compare the examples of copyright law and patent law in the United States. On the Anglo-American conception the purpose of copyright law and patent law are one and the same: “to promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries”, in the words of the U.S. constitution.³⁷ Notably however, copyrights and patents are governed by very different rules.

In general, it is much easier to obtain a copyright than it is to obtain a patent. In the United States a person gains a copyright simply by creating a copyrightable work, whereas a patent must be applied for.³⁸ In the United States, in order to obtain a copyright the work in question need only be original to the author, and fixed in a medium of expression.³⁹ The requirements for patenting an invention are far more stringent: to qualify for a patent, the invention must be novel, it must be useful, and it must be non-obvious.⁴⁰ In addition, the invention must be a “process, machine, manufacture, or composition of matter”.⁴¹ The stringency of these requirements mean that many inventions cannot be patented: between 1996 and 2005, in the United States only 55.8% of patent applications filed ultimately resulted in the grant of a patent.⁴² Even after a patent is granted, the validity of patents are often successfully challenged.⁴³

Furthermore, patents are more restrictive of what others can do than are copyrights. In the United States, the owner of a patent has the “right to exclude others from making, using, offering for sale, or selling the invention[…] and, if the invention is a process, of the right to exclude others from using, offering for sale[…] products made by that process”.⁴⁴ Copyright law provides no right to exclude so broadly construed as this.⁴⁵ While a copyright forbids others from doing some things with a person’s art, it certainly does not forbid all uses of the art and the sale of anything produced by the use of the art, in the way that patent protection would.

Fitting these differences into the Anglo-American justification for copyright law, it would be fair to say that the reasons why it is appropriate that it be far more difficult to obtain a patent than a copyright are precisely because the protections a patent affords are broader. Though the point of both of these types of rights is to promote the progress of the useful arts and the production of valuable works, more broadly distributed rights (like copyrights) merit narrower restrictions.

5.2 Free-riding and Copyright Law

With that in mind, I can now explain why the free-riding complaint asks for copyrights to be too powerful. This complaint is that artists have put in work that the users and producers of AI generated images are benefiting from without compensating those artists. Though artists did not intend to, they produced an essential input into the development of this AI image generating technology. If the existence of AI generated images cuts into the demand for art produced using conventional methods, as many artists fear, then a large part of the economic value of the artistic work they have done will come from the fact that these works were a useful input in the creation of AI image generation software. But because of free-riding, none of the value of their art as an input will be captured by the artists themselves. This seems unjust, because on this conception, artists are entitled to capture the economic value others get from the use of their art, whatever that use might be.

However, copyright should not be construed to protect against this kind of ‘free-riding.’ A patent would perhaps protect against this kind of free-riding use of their art, but a copyright protects against one specific kind of use, copying, and with good reason.

The above conception of what the artist is entitled to argues that their copyrights should give them essentially the same rights that patents provide—the “right to exclude others from making, using, offering for sale, or selling”⁴⁶ the artistic work. The problem with this is that art is ubiquitously and perhaps necessarily “used” by other artists as part of the creative process that produces new art. Virtually all art is produced via an artistic process that involves the artist observing and learning from art produced by others, and involves copying stylistic elements and motifs from these works. But if this is the case, if artists were given this expansive right to exclude, artworks that are inspired by or responding to other works of art would be required to licence the works that inspired them or be copyright infringing, in the same way that “follow-on innovations” require the user to licence the patents that the follow-on innovation depends on to avoid patent infringement.⁴⁷ This environment would not be conducive to the development, production, and dissemination of art. In this environment, if struggling human artists did not adjust their output, then in all likelihood, they would be required to pay far more in licensing fees to the dominant sellers of art and cultural media than they would receive in payment for the use of the art they create.

As I have argued above, AI image generation software does not contain copies of copyrighted works or use copies of copyrighted works in the production of new images. This feature of the AI generated image’s relationship to the copyrighted materials they allegedly infringe on ought to be a difference-making feature, at least on the Anglo-American conception of copyright law. These ways of using the art of others may be a kind of free-riding in a certain manner of speaking—it uses art made by others as a necessary input and does not compensate them for this use. However conventional visual artists use antecedent art in this same way when they create new works.⁴⁸ This is why it is appropriate that artists should be protected from the copying of their artwork, and not from others making any use of their artwork.

5.3 A More Narrow Complaint?

But perhaps the above claim can be more narrowly tailored—rather than claiming that artists are entitled to protection for any useful application of their work, we could instead claim that they should have the right to exclude their work from being used to train generative AI specifically.⁴⁹ From the perspective of the economic-displacement complaint, there would seem to be clear public policy considerations that count in favour of doing this: it would allow artists the opportunity to get a cut of the profits made with generative AI, or alternatively opt out of this form of participation in the creation of generative AI tools.⁵⁰

However, it should be noted that there are public policy considerations that weigh against this narrow restriction as well: the interests that non-artists have in the development and use of AI-image generation technology. Insofar as artists are concerned that they will lose business to these machines, their loss is the gain of those who would like to consume AI-generated images. Furthermore, this transfer is not zero-sum: insofar as (1) many uses of art are well substituted for by AI-generated images, and (2) the production of AI-generated images is more quick and efficient than the production of art by conventional visual artists (the fear that 1 & 2 are true is precisely what grounds the artists’ fears of economic displacement), then collectively we should expect that consumers will gain more than the artists lose.

In any event, as I have said in Section 2, copyright law is not the right tool to address economic displacement complaints. Copyright holders, such as artists, are no more obviously deserving of this protection than other workers displaced by technological changes, and copyright law clearly is not for protecting these other workers.

Is there a reason, stemming from the free-riding complaint, to create this narrowly tailored restriction? On the contrary: from the perspective of the free-riding complaint, adding this narrowly-tailored restriction would seem to be, in an important sense, ad hoc. I cannot see a reason, other than the public policy reasons considered above, why a copyright holder ought to have the right to exclude this use of their work by AI but ought not to have the right to exclude analogous uses of their work by other human artists. If we extended copyright protections in the contemplated manner, it could only be motivated by a desire to protect artists from the disruption of their profession by this new technology—not by an understanding that there is something unique to this AI use of art that makes it morally akin to theft or free-riding.

6 On the French-Kantian Conception, Should Copyright Law Redress the Artists’ Complaints?

On what I have called the French-Kantian conception, the reason why copyright law is morally necessary is because of the special relationship that an author has to their works.⁵¹ Kant claims that an authored work represents the author’s speech, and that speech “belong[s] exclusively to the person of the author, and the author has in them an inalienable right always himself to speak through anyone else, the right, that is, that no one may deliver the same speech to the public other than in his (the author’s) name”.⁵² Adopting the Kantian perspective, we should then think that (1) where that speech is presented, it ought to be attributed to the author, and (2) the author ought to have control over the ways in which speech attributed to them is presented.

What kind of control ought authors to have? A detailed theory of this is not necessary for my present purposes. Rather it should suffice to note that intuitively, there are circumstances in which I would clearly wrong an author by taking away their control of the things attributed to them. I could clearly wrong a person by attributing to that person something they had not said. Likewise, it will often seem that I wrong someone when I divulge to others things told to me in confidence. As such, even when I am honestly attributing things to an author that they have said, if I do so before an audience which the author would prefer not to hear the author’s words, it may be that I wrong the author. But of course, we should not think of this right of the author to control what is attributed to them as absolute. An author has no complaint when I quote something that they have said in public without getting their permission, even if I have good reason to believe that the author would not like me to quote them.

David Davidson’s third complaint, that he has lost control over the content of the work in the art style that he has created, seems to map rather well onto this French-Kantian conception of copyright protections. On the French-Kantian conception, perhaps the process that produces AI generated images can wrong artists by failing to properly attribute their works, and by usurping the control artists ought to have over how their art is used and displayed, and because it wrongs them in this way it ought to be illegal. And so we should ask: does the production of AI generated images wrong artists in this way? I will argue that it does not.

First, I noted in discussing the stylistic control complaint in Section 2 that artists are in a sense vulnerable to impersonation by AI generated images. But it must be said that mimicry and impersonation are not the same thing—impersonation involves at least a misconception about the identity of the author, whereas mimicry need not. In this case, a misconception that the work made by the AI was in fact made by another artist. Merely producing an image that mimics the artists’ style is not enough to do this. At minimum, it requires that the end-user to whom the image is delivered present that image in a way that will cause others to conclude that it was made by the artist in question. It is possible for an artist to be impersonated with AI generated images, just as it is possible for an AI generated image to violate Warner Brothers’ copyright by creating an image of Bugs Bunny. But the process of generating images itself with this software, even images made ‘in the style’ of a particular artist, does not intrinsically involve impersonation in virtue of the production process.

The main issue is whether or not the author’s control over the expression of their work is being usurped by the process that produces AI generated images. This usurpation could potentially happen at one of two stages of the process: (1) when the AI image generation software is produced, and (2) when the final image is produced by the AI image generation software.

At point one, the author’s work is not publicly displayed, and so it seems hard to construe the way that the author’s art is being used at this point as ‘compelled speech’, in the sense Kant is concerned with in his essay “On the Wrongfulness of Unauthorized Publication of Books”⁵³—this is because this use of the work does not involve the communication of anything to an audience.⁵⁴ Presumably when the text-image pairs used to produce AI image generation software are collected, they are collected from publicly accessible places on the internet where it is uncontroversially legal for any ordinary user of the internet to view them. No more use of these images is made than ‘viewing them’ by the machine learning process that creates AI image generation software – this process does not disseminate a copy of any image, as none of the images involved in that process are embedded in the AI image generation software, or used by that software to create new images. This kind of ‘copying’ as an intermediate step in the process of creating software does not usurp the ability of the author to express their work as they wish because it does not really involve the expression of the artist’s work by another to begin with.⁵⁵

At point two, unless the image produced by the AI image generation software is a look-alike of the antecedent art (which is possible with a prompt engineered to achieve this outcome, but certainly not necessary or ordinary), the displaying of this image does not usurp the author’s control over the expression of their work. This is because, as I have argued in Section 5, this image is not the antecedent author’s work any more than the image would be if it had been produced by a human artist.

Lastly, in my David Davidson example in Section 2, I mentioned that he had de facto enjoyed control over the body of works created in the David Davidson art style. This is in virtue of the fact that David Davidson was uniquely qualified to produce these works. He valued that control, and that control is now gone because of the images AI image generation software is capable of producing. Is this a kind of control that David Davidson is legitimately entitled to according to the logic of the French-Kantian tradition?

To make the case that he is entitled to this control: David Davidson is in an important sense the ‘author’ or ‘inventor’ of this artistic style—not only his particular artworks. Producing artworks in this style would thus seem to involve the ‘expression’ of that style in those artworks. As a result, creating new works in this style could be said to involve the expression of something he has authored, and so it may be that according to the French-Kantian conception he morally ought to have control over this expression and this entitlement should be reflected in copyright law.

However, I do not think that this sort of argument should be endorsed, because its potential scope of application is far broader than the production of AI generated images. Once again, this is an issue that applies to the final product of the AI image generation process rather than the process of production itself. If producing a work that copies an artistic style is morally wrong, it is wrong for humans who do it via conventional artistic means, and for humans who use AI image generation software to do the same. This moral complaint does not pick out AI generated images as the problem.

We could say instead that it is not only wrong for the artistic style to be copied in the final image but also as an intermediate step in the process that produces the final image—even if the final image does not copy the artistic style. From this we could conclude that the possibility of wronging people in this way should be reflected in the entitlements conferred to authors by copyright law. But again, this does not pick out something that the AI image generation software does that human artists do not. As I have said above, the normal process of human artistic production involves ‘copying’ artistic styles in this same way. The claim that there is a sense in which David Davidson ought to control how his artistic style is used might be independently plausible, but this argument cannot pick out the way in which AI image generation software necessarily usurps his control without making the same claim about human artists inspired by David Davidson’s work.

We will have to look elsewhere for an explanation of why artists, in virtue of the fact that they have created works used by AI image generation software, are wronged. This is because the kind of ‘copying’ of works that is involved in AI image generation can be expected only to reproduce styles and motifs in the finished product. As a result, it is difficult to see how the operations of the AI image generation software violate the author’s moral rights without making the same claims about the ordinary process of artistic production. The French-Kantian conception of copyright law cannot make the ‘AI art is theft’ allegation plausible because we cannot adopt an interpretation of the French-Kantian conception on which the use that generative AI makes of antecedent artworks ought to be copyright infringing, without saying the same of human-made artworks.

7 Conclusion: The Economic Displacement Complaint Revisited

If the preceding analysis is correct, then the claim that artists are being essentially stolen from by having their work used by AI-image generation software without their permission is incorrect. Copyright does not protect its holder from all uses of their copyrighted work, but from copying specifically, and with good reason.

But the purpose of this analysis has not been to dismiss the anger of these artists as simply based in ignorance of how this technology works.⁵⁶ A better interpretation of the concerns motivating these accusations of theft is that they are based in the perception that this new technology is going to make it harder for visual artists to make a living, and the idea that this is illegitimate.

Of the three artists’ complaints I identified in Section 2, while I found none were related in the right way to copyright law, the economic displacement complaint is clearly a very important concern. While human artists ‘use’ antecedent art in their creative process, there is a morally relevant difference between the way that human artists ‘use’ antecedent art and the way that AI makes use of it: the effect on the economic prospects of professional artists in general. When a conventional artist makes art that depends on the existence of genre conventions and motifs developed by other artists, they contribute to the very body of art they are drawing from such that their work could conceivably be used by those same antecedent artists in their future works. This process of the creation of new art using antecedent art seems ‘sustainable’ for artists, in that each new artist can only take on so much work and is only specialized in producing a narrow range of art, such that there is still room for those antecedent artists to work in.

These are points of contrast with the way that AI image generation software uses antecedent art. Conventional artists fit into this environment like native plants—they grow out of the soil made by the plants that have come before, and while they in a sense compete with the other plants that got there first, they also contribute to the maintenance of this system that allows plants to grow and thrive. AI generated images are like an invasive species, threatening to crowd out native plants and destroy or at least substantially inhibit this ecosystem. The problem is not that the invasive plant is growing out of soil that it did not create—the native plants do that too. The problem is that it is poised to destroy the ecosystem. Likewise, the problem with AI art is that it is poised to extract a huge proportion of the available value from the antecedent pool, leaving too little for the conventional artists to subsist on.

This can be true even while ‘theft’ is the wrong way to describe the problem. Before the advent of AI image generation software, it was already notoriously difficult for visual artists to make a living from their work, and it looks poised to become even more difficult. This is, in my view, a social problem worth taking seriously. I worry that tying these grievances to a mistaken interpretation of what the project of copyright law is about could become an obstacle to addressing this problem. Calls to strengthen copyright protections have long been justified with appeals to the supposed benefits to struggling artists.⁵⁷ Though copyright protections have been strengthened many times in the United States, the problem persists. As Zimmerman argues, a further strengthening of copyright law may in fact make the production of new artworks by conventional means more difficult, by making it harder for new art to avoid infringing on the copyrights of existing works.⁵⁸ The problem that AI generated images represent for visual artists brings into sharper relief what was visible before: that copyright protection will never be able to secure adequate compensation for artists. Artists would be better served by advocating for a more direct solution to this problem, such as a radical expansion of public funding for the arts.⁵⁹