Artifacts Without Authors: Generative Artificial Intelligence and the Question of Authorship

1. Generative AI and its Products

The use of computer programs to generate various types of artifacts is not a recent development. For example, consider AARON, a computer program created by the artist Harold Cohen in 1973 (Garcia 2016). AARON was designed to generate paintings based on specific rules and forms outlined by Cohen. Cohen would then utilize machinery, such as an attached robot hand with a marker or a printer, to physically create paintings based on AARON’s outputs. Although Cohen continually refined AARON’s algorithm to enhance its sophistication, the algorithm’s ‘decisions’ regarding its outputs were ultimately determined by the rules and forms, no matter how complex, imposed by Cohen.⁵

Another example is David Cope’s Experiments in Musical Intelligence (EMI), designed to generate musical compositions. Initially, Cope structured EMI similarly to AARON, defining specific rules and forms for musical composition (Cope 1987). However, he later modified the model to resemble current GAI systems, where the program was trained on a relatively large dataset and combined various elements in a manner that reflects common elements found and recognized in its training data.⁶

The most recent GAI applications, including ChatGPT, Bard, and LlaMa, differ significantly from early examples of generative artificial intelligence in several ways. While many of these differences are not directly relevant to the main question addressed in this paper and will be set aside, a crucial distinction lies in the size of the data used for training these statistical language models. Similar to their early counterparts, recent GAI applications are designed and trained to analyze data, recognize patterns, and generate outputs based on some statistical model in response to user prompts. However, a key difference is evident in the approach to coding rules and decisions into the algorithms. Unlike early examples, recent GAI applications such as ChatGPT do not incorporate a vast set of pre-programmed rules and decisions by their developers to generate outputs. Instead, these language models are trained with an enormous volume of data to produce outputs that mimic those generated by a human agent. Consequently, the programmers or trainers of these models do not make explicit decisions about the content of the outputs in response to specific inputs. This difference is crucial because, as explored later in the paper, part of what it means to be an author of an artifact involves making various decisions about its features. The more decisions developers or programmers make about the content of a particular output from a GAI system, the more claim they have to be at least partial authors of such outputs. In the absence of explicit decisions, it appears that recent GAI systems ‘behave’ more autonomously compared to their earlier counterparts. This autonomy contributes to the distinct nature of contemporary debates about the authorship of GAI-generated entities compared to similar debates surrounding early AI applications such as AARON or EMI.

The integration of GAI into the artistic domain has sparked a heated debate regarding the authorship of works generated by GAI systems.⁷ While it is too early to conclusively determine how the art community will respond to these questions, a corresponding scholarly debate is emerging within copyright and patent law.⁸ It appears inevitable that various legal bodies will eventually address these complex questions. While many of these discussions center on works of art and, in some cases, advanced technological inventions, my focus extends beyond these domains. In fact, I will center my discussion on various sorts of GAI-generated entities that may not urgently demand answers to the question of their authorship. There are several reasons for this decision. Firstly, current GAI applications are used to generate a wide variety of entities, not just works of art. As I will demonstrate below, the question of authorship for these entities is as philosophically interesting and significant as the question of authorship in the use of GAI applications for artistic practice. Secondly, the treatment of authorship questions may differ between works of art, as a subcategory of artifacts, and non-art artifact kinds such as computer programs, food recipes, or letters. Restricting our discussion solely to works of art when examining these questions might be misleading. Finally, the limited literature on the authorship of GAI-generated artifacts has predominantly focused on works of art. Thus, expanding our discussion to encompass other artifact kinds may contribute to our understanding of GAI-generated artworks in return.

Before moving onto the discussion on GAI-generated artifacts, let us briefly examine how authorship is typically characterized in the literature. This discussion will aid us in understanding the minimal requirements for the proper attribution of authorship to artifacts and will help us determine whether artifacts generated by GAI can be considered to have authors.

2. Minimal Authorship Conditions

What is it to be an author of an artifact? The question demands an answer, even if one entertains the possibility of artifacts without authors. Acknowledging the existence of artifacts without authors implies that, for those artifacts, even the minimal conditions for authorship are not satisfied by any agent. Therefore, we must first answer to the question of what is minimally required to be an author of an artifact.

This question becomes especially important for those who subscribe to the view that being a member of a particular artifactual kind essentially requires authors with the right sort of intentions. According to this type of essentialism concerning artifact kinds, being an artifact is, at least in part, being a product of some (human) agent’s creative intentions. There exist various versions of this intentionalist essentialism, with differences arising mainly from distinct views about the content of the intentions necessary for the creation of an artifact. Lynne Rudder Baker, for instance, argues that artifacts essentially have proper functions. According to her, the proper function of an artifact is the one intended by its author(s) (Baker 2007, 52). In Baker’s view, the content of an author’s intention determines the function that their creation is supposed to serve.

Alternatively, some intentionalists propose that the metaphysically relevant content of authorial intentions is not determined by any particular functional, structural, or aesthetic property that the created object is intended to instantiate.⁹ Instead, they argue that the content is at least partially determined by the intention that the created object falls under a specific artifactual kind. This condition is often formulated as a necessary but not yet a sufficient condition, as there appear to be further success conditions for the creation of an artifact.¹⁰ According to these views, a successful creation of an artifact requires that its author has a substantial concept of the specific kind they are taking themselves to be producing (Thomasson (2007) and Evnine (2016)), or a type (Hilpinen (1993)), or a work description (Mag Uidhir (2011)) under which the object falls. It is important to note that these intentionalist accounts are not strictly accounts of authorship but rather identify some intentional property as an essential property for artifactual kinds. However, based on these essentialist views about artifacts, it is possible to derive some basic elements of an intentionalist account of authorship.

Take, for instance, Baker once again. One of the conditions that are both necessary and jointly sufficient for an object’s belonging to an artifactual kind is the following:

Leaving aside the details of Baker’s account of constitution and her views on primary kinds, what is clear from this particular condition is that being an author of an artifact, according to Baker, requires arranging certain entities or selecting certain features of an object in such a way that the created object satisfies or performs its proper function. Baker acknowledges the possibility of cases where this intentional action of arranging or selecting is quite minimal, such as when a natural object is appropriated without modification to serve a certain function (Baker 2007: 53). Thus, on Baker’s view, authorship essentially requires making various decisions about and taking intentional actions towards some object and its intended (functional) properties.

A similar account of authorship can be derived from Thomasson’s and Evnine’s views on artifacts. According to Thomasson, for an object to be an artifact of kind K, it must be the product of a largely successful intention that it is a K. The author has the relevant intention if they have “a substantive concept of the nature of Ks that largely matches that of some group of prior makers of Ks (if there are any) (Thomasson 2003: 600).” The author realizes this intention by imposing properties that are relevant to the kind K onto the object. However, unlike Baker, Thomasson does not restrict these intended properties to be only functional ones. The K relevant properties, or the properties associated with the kind K, may involve aesthetic, structural, historical, or some normative properties as well as functional properties (Thomasson 2003: 509).

Similarly, Evnine argues that intentional making is necessary for the existence of members of artifactual kinds. According to him, the content of the creative intention involves an intention to make an object of an artifactual kind K, “where the kind K provides the conditions on boundaries and flexibility for the objects falling under it” (Evnine 2016: 72). So it seems that on both Thomasson and Evnine’s intentionalist accounts, like Baker’s, to be an author of an object x belonging to an artifactual kind K requires being responsible for x’s falling under the kind K, where this responsibility entails the responsibility of making it the case that x has properties associated with the kind K.

An account of authorship for artworks along these intentionalist lines is defended by Christy Mag Uidhir. According to his view, to be an author of a work w that satisfies some sortal description F entails the following requirement:

This condition, according to Mag Uidhir, serves as a minimal condition of authorship of works (of art) that any account of authorship must capture (Mag Uidhir 2011: 376).¹¹

It seems that what is common between these different intentionalist proposals is that the authorship of an artifact minimally requires that the author’s intentions and the activities guided by them provide a certain degree of control over the kind their creation falls under and its various properties. This control may take various forms. For instance, in the case of concrete artifacts such as chairs, the author may intentionally select certain materials or shapes them for their creation and implement those choices by physically modifying some matter. However, it would be a mistake to think that such a level of control and physical modification is necessary for proper authorship attributions.¹² This is because there are examples of artifacts where no such physical modification is imposed on some previously existing material. So, it appears that the control over the author’s product and the selection of its properties may be minimal. An agent may become an author of an artifact by merely intentionally selecting some preexisting object and presenting it as belonging to some artifactual kind.¹³ In the case of ready-made artworks, this selection and presentation seem to change the kind of the object in question. For instance, a urinal may become a work of art by merely being selected and presented as such. Therefore, the minimal authorship of artifacts may only require an intention to select and present an object as belonging to some (other) artifactual kind.

It is important to note that this is only a minimal condition of proper authorship ascriptions. In most cases, authorship may require much more control over the properties the created object is supposed to instantiate. To be the author of a chair may require more than merely selecting and presenting a piece of wood as a chair without modifying it in various ways so that it instantiates certain properties typically associated with the kind chair. Differences with respect to the degree of control authorship requires may be the result of variations between different kinds of artifacts and their unique nature. For concrete functional artifacts such as chairs or airplanes, mere selection and presentation may not suffice, as what is being selected and presented must be able to support the functional properties typically associated with these kinds of artifacts. However, the case may be different if we are thinking of works of art, especially the kind of artworks such as ready-mades and found art.

In what follows, I present various cases where intentionalist accounts of authorship fail to attribute proper authorship of particular artifactual objects to any human agents. Some of these cases may as well be accommodated within an intentionalist framework if the conditions of minimal authorship are further relaxed. However, as argued below, there are cases that even the most relaxed intentionalist account or minimal conditions of authorship cannot successfully justify authorship claims to any human agents. Therefore, I conclude that we are forced to recognize the possibility of authorless artifacts.

3. GAI and authorless artifacts

There is already a growing controversy in the public domain about the authorship of various entities generated with the help of GAI applications. The use of GAI applications in art-making predates the public’s recent attention and participation in these debates.¹⁴ Often, these public debates focus on issues surrounding the creativity of the artist who employed a GAI application in their artistic practice or the creativity of the work generated with the help of GAI.

Recently, however, one particular example of such public controversy has significantly focused on the question of authorship. The controversy is over an image that won the Colorado State Fair’s annual art competition in the category of emerging digital artists. Jason M. Allen, who submitted an image to the contest used Midjourney, which is one of the GAI platforms where a user can produce images using text prompts. With the prompts Mr. Allen entered the platform, the GAI-generated a number of images, three of which were then selected by him and submitted to the competition. One of those images, with the title Théâtre D’opéra Spatial won the first place and the controversy ensued (Roose 2022). Théâtre D’opéra Spatial may be one of the first GAI-generated works that was recognized by a certain art community without any deception about its origin and the production process. I will not speculate about the future of this controversy, nor will I discuss the moral, political, or various aesthetic questions or concerns GAI applications such as Midjourney has already ignited. My main concern is about the question of authorship of entities such as Théâtre D’opéra Spatial. Therefore, the main question here is who is the author(s) of Théâtre D’opéra Spatial?

It seems that the GAI application is not its author. This is especially clear if, together with the intentionalists, we assume that proper authorship ascriptions require conscious agents capable of forming various mental states, such as beliefs, intentions, and desires. This condition does not exclude the possibility of non-human animals being authors of their products, as certain animals may be capable of forming such mental states in ways that differ from humans, perhaps to a limited degree.¹⁵ While I will not argue for this, as it is beyond the scope of this paper, it seems quite clear that current GAI applications, such as ChatGPT and Midjourney, are not conscious agents. They lack beliefs, desires, or intentions.¹⁶ Therefore, they cannot be agents responsible for the entities they generate.

Perhaps the programmers of the GAI or those who trained it with data can at least partly be the authors. This is in fact what philosophers such as Hilpinen and Reicher might argue for. According to Hilpinen, whoever created a computer program such as AARON is the author of its products (Hilpinen 1993).¹⁷ Reicher, on the other hand, thinks that programmers are what she calls “shared authors” of works generated by computer programs. This is because, on her view, authorship of a work (of art) essentially involves making decisions, the kind of decisions that determine the work’s internal properties (Reicher 2015: 115). Since, she argues, the programmer makes various decisions about what kind of outputs will be generated in response to the kind of inputs given to the program, the programmer in effect makes decisions about the internal properties of the work produced by the computer program. Therefore, the programmers are shared authors of such works.

The problem with Reicher’s analysis is that the kind of computer programs she was basing her analysis on is quite different from the most recent examples of GAI applications. The programmers of current GAI systems do not make decisions about the kind or qualities of outputs in response to various inputs.¹⁸ Instead, the GAI system, with its training based on an enormous size of data, tracks, recognizes, classifies, and reproduces patterns in its training data set. Therefore, simply put, the programmers do not make decisions about what kind of properties Midjourney, for instance, will produce in response to Mr. Allen’s series of prompts. Since they do not make decisions that determine the internal properties of entities like Théâtre D’opéra Spatial, it is not plausible, even on Reicher’s account, to attribute even partial authorship to the programmers of Midjourney. This leaves us with the user, here in this case, Mr. Allen, and the authors of images used to train the GAI. The latter can quickly be dismissed as a potential author of Théâtre D’opéra Spatial, given the vast number of images fed into the GAI. There is simply no ground to attribute meaningful authorship claims to the creators of images used in the training process.

One might argue that the case at hand is better understood as a collective or joint making involving the user, the programmers, and the authors of images used to train the GAI.¹⁹ It appears that each party contributes to some degree to the production of the images generated by Midjourney. Therefore, intentionalists might argue that the situation represents a case of collective or joint making.²⁰

I will address the user’s claim for authorship below, so my focus here is on the other candidates for collective authorship. One immediate problem with this proposal is that typical cases of collective making, such as the production of artifacts like a car or a smartphone, involve each party having some degree of control not only over the existence but also over some of the properties of the final product. However, if the arguments above are correct, the roles of the programmers and the authors of training images are minimal, if at all, in determining how the images produced by Midjourney in response to the user’s prompts look. In the absence of significant control over the properties of GAI-generated entities, it seems difficult to motivate the claim that the case at hand is a case of collective making.

Furthermore, collective making typically requires at least coordinated or cooperative action to achieve a goal, which might further require individuals sharing an intention to achieve said goal. However, in this case, the candidates for collective making do not participate in coordinated or cooperative action nor share an intention to produce an image that Midjourney generated in response to Mr. Allen’s prompts. This is evident with the authors of training data. The programmers of the GAI application are clearly involved in coordinated and cooperative action with some shared intention to produce the GAI application Midjourney itself. However, it seems quite implausible to think that they are equally, or to a significant degree, involved in the production of the images generated by Midjourney. Therefore, the generation of Théâtre D’opéra Spatial cannot be understood as collective or joint making. Consequently, the only conceivable candidate for authorship remaining is the user of the GAI.

In the case of Mr. Allen, the user of the program, it appears that he employed a series of complex prompts to generate his envisioned image. Subsequently, he further manipulated these images digitally using other computer programs before printing and submitting them for the competition. Assuming that his actions, such as providing prompts, selecting specific images, and presenting them to the contest, were driven by the intention to create a work of art—specifically, digital art—his case seems to satisfy the minimal conditions of authorship required by most intentionalist accounts discussed earlier. This is because not only were Mr. Allen’s creative actions guided by the intention to create an artifact of a specific kind, but he also exercised a degree of control over various properties of the resulting entity, namely Théâtre D’opéra Spatial. This control was exerted through the complex series of prompts given to the GAI until the images produced by Midjourney aligned with Mr. Allen’s artistic objectives.

Even if the degree of control he had over the images generated by the GAI were significantly lower than assumed here, there is another reason to recognize him as the author of the work. As mentioned previously, the limiting case of authorship appears to be in cases of authorship by appropriation. Many intentionalist philosophers, including Reicher, acknowledge this type of authorship as a limiting case for the creation of an artifact:

Even in the absence of substantial control or decision-making regarding the properties of the GAI-generated outputs, it is evident that the question of authorship for such artifacts can be resolved by meeting the minimal condition of authorship. This is accomplished through the act of selecting and presenting a specific object as belonging to the appropriate artifactual category, in this case, a work of art.

The case of Théâtre D’opéra Spatial is relatively straightforward for most intentionalist accounts of authorship, particularly those acknowledging the possibility of authorship by appropriation. However, there are instances where intentionalist accounts face serious challenges. Below, I present various cases that complicate matters for intentionalist views. While presented as possible cases, they can be easily generated using one of the currently available GAI systems.²¹

Firstly, consider a case in which I use a specific GAI application to generate a simple computer program. Let us assume that I have no predetermined function or purpose for this program. It is worth noting here that I may not even possess the knowledge of how to write a computer program to complete this task. With these considerations in mind, suppose I input the prompt “generate a computer program” into the GAI application. In response, the GAI application produces an entity that appears to be a computer program, let us call it CP, written in a specific code. Let us further assume that although CP is simple, it has a novel and functional code that does not already exist. Subsequently, I store CP on my computer and later utilize it in various ways.

While this case shares similarities with Théâtre D’opéra Spatial, there are significant differences. First, I have almost no control over the properties of the object, that is, CP, generated by the GAI. Because the prompt I provide to the machine is exceedingly simple, it does not impose significant demands or restrictions on the qualities of the computer program I request the GAI to generate. Second, I receive a single output and decide to use it, without necessarily selecting it for presentation to some institution. Third, it appears that the output, CP, generated by the GAI in response to my prompt is already a computer program even before my decision to choose and use it. This is crucial, as in Mr. Allen’s case, one might argue that his act of selection and presentation changes the ontological status of the output, from a mere digital image to a digital painting. That is, before his act of selection and presentation, the outputs existed merely as digital images and were not yet classified as works of art. However, upon Mr. Allen’s decision to choose and present them to the competition, they acquired the status of works of art. This is in line with typical cases of appropriation, such as ready-mades or found art, where the artist’s decision to regard an object of a specific kind brings about a new object belonging to a different kind. Contrary to these examples, in the present case, it seems clear that the output generated by the GAI is already a computer program. Thus, my decision to select and use it does not alter the ontological category to which CP belongs.

Despite these differences, an intentionalist might argue that the minimal authorship requirement is satisfied through my intentional actions, making me the author of the computer program. Even if I lack control over the qualities of CP, its existence can be attributed to my intentional action—a creative act directed at an object of a specific kind. Furthermore, the decision to use the computer program is similar enough to typical cases of appropriation, even if the act of selection does not alter the ontological category of the object in question. Indeed, there are also cases of appropriation in art that do not change the kind of object created but still constitute the creation of a numerically distinct work of art.²²

Now, let us modify the scenario slightly. Instead of generating just one entity, let us assume that the GAI produces five distinct outputs, namely CP₁, CP₂, CP₃, CP₄, and CP₅. Each has unique codes, functions, etc., in response to either repeated prompts or a slightly different prompt requesting the generation of five computer programs instead of just one. Subsequently, I store all five outputs on my computer. However, I later choose to use only CP₁ while leaving the remaining four unused. Similar to the previous case, an intentionalist might attribute authorship of the entity I chose to use, namely CP₁, to me, arguing that the act of selection and subsequent use satisfies the minimal conditions of authorship. However, unlike the first case, there are now four additional entities generated by the GAI that I never select or use. These entities, in terms of their internal structure, are very similar to the one I decided to use and thus do not seem to differ in any metaphysically significant way, except that they were not chosen for use. This is why it would be implausible not to classify these unused entities as belonging to the same artifactual kind as CP₁, namely the kind computer program. However, if this is correct, intentionalists cannot justifiably claim that I am their author. Even if they appeal to a minimal condition of authorship in the limiting cases of artifact creation, such as authorship by appropriation, the four unused entities do not satisfy this condition. In other words, if all five entities generated by the GAI are considered computer programs before I decide what to do with them, then they would all continue to be so, even if I select one and never choose the remaining four for later use. Since the minimal condition of authorship in the limiting case of artifact creation, i.e. authorship by appropriation, is not met for the four unused computer programs, they are artifacts without authors.

Therefore, I conclude that there are, or could be, artifacts generated by a GAI application that have no authors. It is important to emphasize that not all GAI-generated artifacts lack authorship. As demonstrated in the example of Mr. Allen and Théâtre D’opéra Spatial, it is possible to attribute authorship based on the satisfaction of minimal conditions. However, the existence of various examples of authorless artifacts suffices to challenge the intentionalist account of artifact authorship, which is the primary objective of the project undertaken here.

4. Objections and replies

One way to resist the possibility of authorless artifacts in both cases is to argue that the outcomes of GAI processes, namely CP in the first case and CP₁ to CP₅ in the second case, are not artifacts until they are selected for use or appropriated in some other way. Authorship by selection and presentation/use already explains why Théâtre D’opéra Spatial or the case where I select and use CP does indeed have authors. Before any such intentional action, the products of the GAI application are not yet artifacts. In other words, the case where the GAI produces a number of outputs in response to my prompt “generate computer programs” is not a case where all of its outputs are indeed computer programs. Similarly, when Mr. Allen used Midjourney to generate a number of images, the outputs of the GAI system were just mere images and not yet digital paintings. The act of selection changes the ontological category of these products. Since, in either of these cases, there are no artifacts before an intentional act of selection, presentation/use, or appropriation, these cases do not constitute examples of authorless artifacts.

This line of response, I believe, aligns with the approach taken by intentionalist philosophers, such as Reicher, when addressing similar hypothetical cases:

One obvious problem with this particular response is that the case at hand is not entirely analogous to swamp cases—cases where something with various properties shared with an artifact of a particular kind is generated through some random natural process. According to the intentionalist view, in such swamp cases, the object’s origin is not of the right sort, meaning the object does not come into existence as a result of some intentional human action. However, in the cases discussed above, the origin of the objects in question—namely, the outputs of the GAI application—is a result of intentional human action, specifically the user’s prompts provided to the GAI. The images or codes generated by the GAI exist partly because users prompted the machine to produce an object of a particular kind and are not a result of some random natural or artificial process. In other words, they are not accidental products of a natural event or a human action; therefore, they are not random products.

More importantly, if they were not digital paintings in Mr. Allen’s case or computer programs in the other two cases, what would they be? Perhaps, in the former case, one might insist that they are digital images but not yet works of art. As one might argue, to be a digital painting and not merely a digital image requires something more than an arrangement of various shapes and colors; it requires some artistic meaning. Unlike Théâtre D’opéra Spatial, the rest of the images generated but not selected for presentation by Mr. Allen lack such artistic meaning and are thus not yet works of art. I believe that works of art may indeed have different conditions for their existence compared to other kinds of artifacts, especially those with an identifiable function. More specifically, opponents of authorless (art) works such as Reicher may be correct to insist that for something to be a work of art, there needs to be some artistic decisions made by an author. Therefore, they may be correct to reject the possibility of authorless works of art.

However, the case is different for other kinds of artifacts, especially those with an identifiable function. Consider, once again, the outputs of the GAI in response to my prompt “generate computer programs”. The ones that I do not select and use have similar functions or forms as the one that I do select. It seems that the ones stored in my computer and never used are, in many ways, or ways that metaphysically matter, just like the selected one. My mere selection does not seem to change the kind of object, i.e., from a mere random product to a computer program, generated by the GAI. Unlike works of art or digital paintings, artifacts such as computer programs do not seem to require a further condition to be categorized as such. Let me emphasize that, unlike some alleged cases of authorless artifacts mentioned by Reicher, the outputs of the GAI are a direct result of a user’s prompt and thus a human intentional action. The problem for the intentionalist here is that said intentional action—namely entering the prompt “generate computer programs”—does not, by itself, constitute what is minimally required for proper authorship attributions. In particular, in this case, the user has no control or makes no significant decisions about what kind of computer program the GAI generates. Therefore, these outputs, CP₂ to CP₅, like the one I select for use, CP₁, are indeed computer programs, but unlike CP₁, they have no authors.

One might argue that instead of recognizing my intentional action of selecting one and discarding the remaining four as the source of my authorship, we should consider the prompt I provide to the GAI as the true source of authorship. This approach not only explains why all five are correctly classified as computer programs but also claims that I am the author of all five. This line of response is also consistent with some of the essentialist intentionalist accounts of artifacts discussed earlier. Recall that, according to some intentionalists, to be an artifact of kind K is to be a (successful) product of an intention to create an artifact of the kind K. If we take the user’s prompt “generate a computer program” as the content of their creative intention, then it seems that this intention, by specifying the particular kind of artifact the user intends the GAI to produce, suffices for its outputs to be the products of their intentional action. Since all five entities come into existence as a result of such an intentional action, the user is considered their author. One might further argue that this is analogous to a case of commissioning a work of art, where an artist specifies the work they intend to create but commissions the actual labor to those who follow their instructions.²³

One obvious problem with this line of response is that it appears to give up on the idea that authorship attributions require a certain degree of control over the qualities of the created entities. Commissioning cases in art also typically involve authors who exert significant control over many properties of their intended work. The distinction from paradigmatic cases of creation lies in the fact that the author is not the one physically producing the work itself through direct manipulation of concrete materials. Even in cases of appropriation, the author typically exercises control by selecting an object as their own, determining whether it aligns with their goals, demands, desires, requirements, etc. In the case at hand, however, there is no control of any kind, except for whatever is present in the content of the prompt given to the GAI.

Compare this case with a case where you ask your friend to generate a computer program for you. Let us assume that your request is as general as the prompt “generate a computer program.” Further, let us assume that your friend takes up this task and produces five distinct computer programs. This case differs from GAI cases, as your friend qualifies as a potential author (unlike the GAI application) and thus has a legitimate claim to be the author of their work. To accommodate this difference, let us rather focus on the question of co-authorship or shared authorship of the five computer programs your friend wrote. Suppose that after seeing the first two programs, you decide to choose the first one and use it for some purpose. The intentionalist view at hand implies that you are a co-author or a shared author of all five computer programs, even though you have not even taken a look at the remaining three programs, let alone used them for some purpose. It is true that you are partially responsible for the existence of all five computer programs; however, this responsibility does not seem to suffice for your claim to be the co-author of those programs. Your claim is admittedly stronger for the co-authorship of the program you chose, as it is similar to cases of appropriation. However, your claim seems unfounded for the remaining programs, as you have no control over their features except that they fall under the artifactual kind computer program. This is because, as I argue, partial existential responsibility does not suffice to warrant partial or full claims of authorship. Authorship of an object further requires some degree of control over its qualities, which is lacking in these cases.

Even if I am mistaken and the minimal requirements of authorship are further relaxed, as indicated above, to allow claims of authorship in the cases discussed on the grounds that an intention to generate or create an artifact of a particular kind, such as a computer program, without specifying its features or functions, is sufficient for authorship attributions and thus constitutes exercising some control over the product, the challenges for intentionalists with GAI-generated artifacts still persist.

To illustrate this, let us consider one last case. In this scenario, imagine that I feed the GAI with a prompt that does not even specify the kind of artifact I want it to generate. Suppose I give the following prompt to the machine: “generate something.” As a result, let us assume that the GAI generates what appears to be a simple computer program. Unlike the previous case, I do not intend for the GAI to generate any particular artifact as a product. This means that whatever kind of entity is produced by the GAI is not determined by my intention or my intentional action. Therefore, if what the GAI produces is indeed a computer program, it has no author. This consequence cannot be avoided even if we accept the intentionalist response described above.

As a final attempt to salvage the intentionalist account, one might concede that what the GAI generates in all these problematic cases are indeed computer programs, but deny that all computer programs are artifacts and must have authors. That is, CP₂ to CP₅ and the computer program generated in response to the prompt “generate something” belong to the same kind namely computer program, but some members of this kind are not artifacts. Only the authored members of the kind are artifacts.²⁴ This line of response should, of course, be extended to other sorts of entities that are or could be generated by GAI applications, such as poems, stories, novels, musical works, and digital paintings, among others. Notice that this response already concedes that some computer programs have no authors, and therefore is consistent with some of the consequences of my arguments. The main disagreement is over the general kind artifact or whether the members of this kind must have authors. If we wish to maintain what seems to me a very intuitive claim that kinds like computer program, poem, or musical work are artifactual kinds and thus their members are artifacts, then I believe we should instead accept the conclusion that some of the members of these kinds, although artifacts, have no authors, rather than claim that such entities, though they belong to these kinds, are not artifacts.

5. Conclusion

It is commonly assumed that it is necessary that artifacts have authors. Recognizing the necessity of authors for artifacts, various intentionalist proposals have been put forth to account for artifact authorship. It seems that the minimal conditions of proper authorship of artifacts require authors to hold a specific intentional attitude towards the object of their creation. This does not necessarily entail physical modification of material; the limiting case of such creation might involve selection for presentation, use, or some other form of appropriation. I argue that even this minimal condition of authorship for the limiting case of artifact creation is not satisfied with the growing examples of GAI-generated artifacts. I provide several examples where possible candidates for authors—the programmers, trainers, original authors of the training data, or the user—cannot qualify as the authors of such artifacts. Therefore, I conclude, contrary to the received view, that there are artifacts without authors.

However, this conclusion does not conflict with authorship attributions in cases where users have some degree of control over the products of GAI applications. In fact, in most cases, users of these applications exert significant control over the properties of the final products through specific prompts. Often, users will continue specifying their prompts in response to the products of the GAI application until they are satisfied with the final product. Therefore, quite frequently, the user of a GAI application is entitled to be the sole author of whatever is produced by these machines. The arguments I provide above, inasmuch as they show the possibility of authorless artifacts, also demonstrate how and under what circumstances such author ascriptions can be justifiably made.

The possibility of authorless artifacts relies on the assumption that proper authorship ascriptions minimally require a conscious agent with some degree of control over the products of their creative attempt. This, in turn, implies that authors must have a particular sort of intentions and perform various actions guided by those intentions. One of the goals of this paper is to clarify the minimal conditions required to establish authorship. However, this task is incomplete if we do not examine similar questions concerning related concepts such as ownership and responsibility for GAI-generated artifacts. I hope to have shown that these questions might have answers that carry philosophically interesting consequences and are worthy of our attention.

Competing Interests

The author has no competing interests to declare.