In literary research, a division is often made between the “actual author,” who is the intellectual creator of the work, and the “implied author,” which is an image of the author that the reader can infer from the text (see e.g. [Booth 1961] [Booth 1983] [Chatman 1978] [Chatman 1994] [Schmidt 2010]). In this article, my focus is on the question of how authorship is viewed in story generation research, a question that is primarily related to the actual author.

This aspect of authorship is generally associated with the assumption of an individual who can profit from their literary work and is responsible for and deserving of credit for it. This perception of authorship has evolved alongside print technology and the publishing industry [Woodmansee 1994, 35–36] [Finkelstein and McCleery 2005, 69–71] [Febvre and Martin 1984, 165–166]. Before the invention of printing in Europe, authorship was not defined in the same way as it is today, and the authors of texts were not considered their owners [Finkelstein and McCleery 2005, 69]. Writing was seen more as a form of craftsmanship or a channelling of inspiration from a higher power [Woodmansee 1994, 35]. The perception of authorship began to shift as distribution of hundreds of books became possible with print technology, and as copyright legislation began to be drafted in the 18th century [Woodmansee 1994, 36–37]. In addition, inspiration and genius began to be used to characterise authors around the same time [Jefferson 2014, 4, 61, 67–71]. These developments influenced the formation of the current understanding of authorship, as they led to the forming of the profession of writers who could support themselves with the proceeds of their writings [Woodmansee 1994, 36–37]. The author’s role as the party that is responsible for the aesthetic whole of the work is also often associated with the term [Lamarque 2009, 109–111].

In many respects, electronic literature often goes against the grain of this perception of authorship. For instance, in the 1990s, many researchers viewed the changes brought about by computers to writing and reading as a counterforce to the old concept of authorship that emphasised the individual author (e.g. [Gaggi 1997, 106–107] [Landow 1992, 70–77] [Bolter 1991, 4]). Many phenomena that challenge the established perception of authorship, such as collective writing, adaptations, plagiarism, and crowdsourcing ideas, are also made considerably easier with digital technologies [Goldsmith 2011, 2–3]. With computer-generated texts, ascribing authorship is complicated even further, as many factors, such as the design of the program and the data that the program has been given to work on, shape the produced text [Henrickson 2021]. Furthermore, with computer-generated texts, the program’s role as a co-author also comes into consideration [Lebrun 2017], and it has been even argued that computer-generated works can undermine the image of the author as a unique genius [Bök 2002, 10].

These differing perspectives — the common assumption of the author as an individual who is the owner of and responsible for their work, and the varying approaches in electronic literature that challenge this expectation — provide a background for analysing the research material as well. I compare how authorship is framed in story generation research in relation to these points of view.

The following research material is based on a number of story generation studies published over the last decade. In selecting the publications to analyse, I rely on reviews that have tracked the of the state-of-the-art of story generation programs [van der Lee et al. 2019] [Herrera-González, Gelbukh, and Calvo 2020] [Alhussain and Azmi 2021] [Guan et al. 2021] [Chhun et al. 2022] [Ansag and Gonzalez 2023]. The other publications I examine [Ammanabrolu et al. 2021] [Bottoni et al. 2020] [Cheong and Young 2015] [Daza, Calvo, and Figueroa-Nazuno 2016] [Fan, Lewis, and Dauphin 2018] [Guan et al. 2020] [Ippolito et al. 2020] [Jain et al. 2017] [Li, Ding, and Liu 2018] [Liu et al. 2020] [O'Neill and Riedl 2014] [Peng et al. 2018] [Peng et al. 2022] [Porteus and Lindsay 2019] [Tambwekar et al. 2019] [Xu et al. 2018] [Yao et al. 2019] were chosen by cross-referencing these reviews and selecting the articles that are discussed in more than one of them. This is to ensure that although I cannot comprehensively cover all of the recent work on story generation in this article, my research material aligns with what multiple other researchers have found noteworthy. The reviews themselves are also a subject of analysis since they discuss the authorship of computer-generated texts as well. I examine how the term “authorship” is used across these various publications in their entirety, including the descriptions of the programs developed and the evaluations of the stories produced.

Story generation programs are part of a larger phenomenon of text generation. Text generation with digital computers has been a method of producing literary works since the mid-20th century [Rettberg 2018, 31–33]. It is a versatile practice where multiple sources of influence are at play. One such influence is dadaist poetry, and methods similar to the cut-up technique are used in computer-generated works [Funkhouser 2007, 32–34]. Furthermore, the work of Oulipo (Ouvroir de littérature potentielle) has also been influential in text-generation, as they explored the intersections of literature and mathematics from the 1960s on and are known for experimenting with combinatory procedures and methods of constrained writing, that is, creating rules and restrictions within which different works could be created [Schäfer 2007, 121] [Funkhouser 2007, 33–34]. Later, text generation has been used in a similar way to produce works based on the rules created by the author [Funkhouser 2007, 34].

One key motivation for text-generation, especially in computer science, has been the imitation of human creativity. Story generating programs that have been developed in the field of computer science have been closely tied to research in artificial intelligence [Gervás 2012]. As Sharples & Pérez y Pérez put it, developing programs that can perform various creative tasks is seen as a way of modelling the human mind, and this work aims to answer questions such as “What is intelligence?” or “What makes humans creative?” [Sharples and Pérez y Pérez 2022, xii]. Here, storytelling is viewed as a task that is particularly difficult for computers, and is thus often discussed as one of the grand challenges of artificial intelligence [Fan, Lewis, and Dauphin 2018, 889] [Xu et al. 2018, 4306] [Guan et al. 2020, 93] [Ammanabrolu et al. 2021, 5859].

The goals and practices of story generation as a field of research also affects how the text-generation process and authorship are framed. As my analysis will show, text generation of literary texts in the context of computer science research is often viewed more in terms of problem solving than in terms of creating art. This marks a difference to many computer-generated works that are created outside of research-oriented environments. Throughout the article, I contrast story generation with text generation outside of computer science research, both to explore story generation as a subtype of electronic literature and to offer an external perspective on the conceptions of authorship that might be useful in the field.

Story generation researchers have historically utilised literary theory in the development of the programs, and some (e.g. [Gervás 2012] [Montfort and Pérez y Pérez 2023, 99]) have even suggested that story generation could work as a means of “testing” different concepts and theories of literature. Story generation researchers have been especially interested in formulations of plot in narrative theory, most notably Vladimir Propp’s famous study Morphology of a Folktale (1968),which has been used either as a source of inspiration or even as a basis for story generation programs in many instances (see e.g. [Turner 1994] [Peinado and Gervás 2006] [Gervás 2013] [Veale 2014] [Ryan 2017]).

However, based on my review, it seems that dialogue between the fields is not necessarily a part of all story generation research. The objective of contributing to or utilising literary theory is not generally present in the articles surveyed, and literary theory concerning concepts such as story and plot is often not referenced at all or only in passing (see e.g. [Li, Ding, and Liu 2018] [Xu et al. 2018] [Fan, Lewis, and Dauphin 2018] [Peng et al. 2018] [Tambwekar et al. 2019] [Ippolito et al. 2020]). Moreover, when literary theory is referenced, it is often in the context of techniques for plot modelling (e.g. [O'Neill and Riedl 2014] [Ansag and Gonzalez 2023]). I wish to suggest that engaging with other areas of literary theory, such as stylistics and genre studies, could also be beneficial for story generation research.

The textual analysis of the research material was conducted by first going systematically through the articles and then identifying recurring themes related to the research question: What or whom do the researchers view as the stories’ author and what does that imply on how authorship is viewed in the research? I observed that discussion related to authorship took place especially in connection with motivating the research, evaluating the outputs, and refining the style of the outputs. Themes that recurred throughout much of this material were reducing the impact of the human author on the outputs, juxtaposing human and computer-generated texts, and avoiding “machine-like” features in outputs. In the sections that follow, I examine how these topics were addressed in the articles and what kind of overall picture of the approach to authorship can be drawn from them.

Authorship is arguably one of the most important topics in story generation research, especially when research motivations are examined. Developing a “computer author” that mimics human creativity is one of main objectives expressed in these articles. This goal is expressed, for example, by linking the research to the development of computational creativity, which is referred to simply as an imitation of human creativity (see e.g. [Jain et al. 2017, 2] [Alhussain and Azmi 2021, 1] [Ansag and Gonzalez 2023, 878]). Some researchers define the term “creative” further, asserting that a program is creative if the output it produces is different enough from the parameters and other information given to it [Alhussain and Azmi 2021, 25] or if the output is unpredictable to its programmers [Daza, Calvo, and Figueroa-Nazuno 2016, 14]. As [Yao et al. 2019, 7381] put it, the aim is often to make a program that passes the Turing test.[2]

Over the time period during which the articles examined were published, the methods of story generation have altered greatly. At the beginning of the 1960s, in the early days of story generation research, the programmer played an integral part in shaping the information required for storytelling into a computer-readable form [Ryan 2017]. Methods such as simulation, planning, and story grammars have been and continue to be utilised in story generation.[3] In recent decades, however, neural network methods, and even more recently LLMs such as generative pretrained transformers, have surged in popularity and are regularly used alongside other methods. With the use of neural networks, in which LLMs are also based on, it is possible to train the program to imitate a set of training data it has been given, for example by recognising which words typically appear together [Henrickson 2018a, 9–10] [Paaß and Giesselbach 2023, vii]. This has enabled greater levels of automation in the story generation process.

This shift in methodology has also affected how the authorship of the produced stories is discussed. With advances in natural language generation, such as the increased use of neural networks, previously dominant methods such as planning and story grammars are often criticised. [Yao et al. 2019, 7378] and [Tambwekar et al. 2019, 5982] criticise older approaches for relying heavily on extensive domain knowledge and annotation crafted by the programmers. “Hand-crafting” the information that is given to the program requires a lot of effort from the developers and significantly impacts the construction of the text, which is viewed as a shortcoming. In contrast, the use of neural models means there is less need for “hand-crafted” knowledge, an outcome that is regarded as desirable [Liu et al. 2020, 1725].

The researchers’ contributions are often contrasted with previous research by showcasing how much they have achieved “authorial reduction” (e.g. [Porteus and Lindsay 2019, 1074]) or lessened the “authorial burden” that goes into making a story generation program [Bottoni et al. 2020, 21] [Ansag and Gonzalez 2023, 878]. Correspondingly, previous work can be praised if generating the story requires “almost no human intervention” [Daza, Calvo, and Figueroa-Nazuno 2016, 10]. In other words, the aim is most often to minimise the role of the human(s) — especially the people who develop the program — in the creation of the text.

The goal of minimising the developer’s contribution is reflected in how the authorship of the stories is characterised. Notably, the researchers do not generally refer to themselves as authors or writers of the produced stories.[4] Instead, the researchers typically name the program they have developed and refer to that program as the writer of the stories. Alternatively, the researchers often refer to the writer as a “model” or “system” and in some cases even as an “agent,” which implies some degree of authority over the output [Tambwekar et al. 2019] [Ammanabrolu et al. 2021].

As these findings show, authorship — or more narrowly, defining the extent to which a story is computer-generated — is closely tied to the goals of story generation research. However, the “authorship” term is not explicitly defined in any of the publications, as would be the case with literary research that focuses on authorship. Despite this, it is apparent the concept “authorship” is ascribed a meaning that is more specific than the everyday use of the word. Based on the research material, the developers of story generating algorithms seem to discuss authorship almost entirely by assessing how minimal their own role is in the text’s construction. This approach might partly stem from the generally desired features in software design: The less hand-crafted material there is in a program, the more flexible it becomes [Xu et al. 2018, 4307–4308] [Liu et al. 2020, 1725]. In such an approach, by focusing on the “authorial reductions,” the concept of authorship is reduced to a technical question: who or what has written each part in a story.

The effort to diminish the role of the human contributors as authors differs from many other works of electronic literature that are produced using text-generation programs. Take, for example, Philippe Bootz’s La série des U. As Bootz who is a renowned poet in the field of electronic literature, describes it, “The work generates a different text-that-is-seen [...] each time it is played through subtle variations in the timing at which the textual elements appear and the relation between the verbal text and the sonic component [...]” [Bootz 2005]. Like story generators, Bootz’s work is not a single, stable text but a text generator that produces different iterations according to principles that he designed. The process of creating the program that produces these texts is similar to story generation. Yet, in contrast to the perceived computer authorship in story generation, Bootz considers himself the (co)author of La série des U. This is evident from how he refers to the work as “a typical example of adaptive generation that illustrates my present approach of e-poetry” [Bootz 2005, emphasis added]. Research on this work also refers to it as being authored by Bootz. For instance, Alexandra Saemmer considers La série des U as an illustrative case of “authors who [...] purposefully reveal details of their program” [Saemmer 2009, 485]. The comparison of story generation programs to similar works outside of the field highlights how differently the roles of the program and its developer can be viewed in terms  of authorship.

Some writers who use computer-programs to generate literary texts do not view the computer program(s) as author(s) in any sense. A computer-poet Allison Parrish, for example, approaches making poetry with text-generation as putting together a collage. For instance, her work Two of Pentacles is based on seventy-eight different snippets written by her and rearranged using different techniques. These include assigning each snippet a vector of words and letter sequences that represent it and a program that gives measurements on the snippets’ similarities based on the vectors [Parrish 2025]. She views the different programs that she uses in the process as a medium of the assemblage, not as authors or collaborators. As she writes, “I never [...] attribute authorship to a program or a model. If I publish the results of a text generator, I am responsible for its content” [Parrish 2021]. It is apparent that in contrast to the way computer scientists think about authorship, Parrish views authorship as a broader concept than merely estimating the ratio of human and computer effort.

The authorship of story generation programs could also be approached from the perspective of remixing and reusing material. Digital media theorist Leah Henrickson has proposed this kind of approach for another kind of work of algorithmic authorship: Twitter bot Pentametron made by Ranjit Bhatnagar, a bot that gathered tweets in iambic pentameter in its feed and also arranged them into longer poems (henrickson, 2018b, 12–13). Pentametron inspired people to use the material in creative ways, and [Henrickson 2018b, 13] (henrickson, 2018b, 13) argued that instead of examining Pentametron’s retweets and poems as final literary texts, the chains of appropriation of texts are the interesting subject of analysis. Even though story generation programs are quite different from social media bots like Pentametron in terms of their purpose and audience, they share a similar function of providing material for readers that can be used in subsequent texts. Story generation programs are mainly examined by other researchers that can utilise them in creating their own programs. Approaching story generation programs from this perspective could also bring a new element in discussion of their authorship. It highlights the collaboration of not only between the program and its developer, but between researchers.

One reason for the narrower approach to authorship in the research material is that in story generation producing the stories is only of an instrumental value, as the purpose of the research is technological advancement. While story generators are developed for purposes such as storytelling [Peng et al. 2018] [Porteus and Lindsay 2019] [Ansag and Gonzalez 2023], creating works of fiction [Daza, Calvo, and Figueroa-Nazuno 2016], interactive entertainment [Alhussain and Azmi 2021], and games [Cheong and Young 2015], these aims are described as the ultimate goals of story generation, not as direct products of the story generators in their current state. This is perhaps the most crucial difference between story generation and works such as La série des U or Two of Pentacles: The process and products of story generation are not viewed as literary works but rather as steps towards the goal of computational creativity. The developers seem to view their work in reducing the “authorial burden” as a technical task, not as an artistic endeavour. Thus, their view of authorship is also technical, and other dimensions of the term, such as the social role and responsibilities of the author, are not discussed.

This technical definition of authorship that leaves out so many aspects typically associated with literary authorship also leads one to ask whether text generators and the texts they produce should be approached as literature. This does not seem to be the framing in computer science, as the programs and the texts they produce are never referred to as works of art or literature. However, this does not necessarily mean that the programs and their outputs should not be regarded as literature. In fact, this has been typical in older story generation programs that are regularly mentioned as part of electronic literature as well. Developers of older story generators such as Tale Spin or Mexica originally framed their work primarily as research of computational creativity and imitation of human cognition [Meehan 1976, 212–226] [Sharples and Pérez y Pérez 2022, 119–120]. Yet, examining these programs as literature has produced valuable discussion, for example on whether works of electronic literature should strive for simulating human writing [Aarseth 1997, 131–141] [Wardrip-Fruin 2009, 155].

Whether the framework of “literature” is useful in examining story generation programs depends also on the purpose and content of the produced texts. Older story generation programs tended to produce outputs in an identifiable literary genre, such as fables or fairy tales [Meehan 1976] [Riedl and Young 2010], detective stories [Klein et al. 1973], or short stories [Bringsjord and Ferrucci 2000]. In comparison, most stories presented in the research material are much shorter and not easily recognisable as belonging to a particular genre. Still, the researchers generally do have aesthetic objectives for the produced texts. These aims include creating suspense [O'Neill and Riedl 2014] [Cheong and Young 2015], literary language [Daza, Calvo, and Figueroa-Nazuno 2016], deepening the characters [Bottoni et al. 2020], following a narrative structure [Porteus and Lindsay 2019] [Liu et al. 2020], or arousing interest in the reader [Yao et al. 2019] [Peng et al. 2018] [Jain et al. 2017]. In this sense, it is worthwhile to analyse their outputs in literary terms.[5]

Developers of story generation programs not viewing themselves as authors or co-authors of the stories is reflected in the juxtapositions of “the human” and “the machine.” This is most evident in the evaluation processes of the programs. In studies that conduct a so-called “human evaluation” of the texts, it is common to include “human-written” texts alongside computer-generated ones to be read by the participants of the study (see e.g. [Cheong and Young 2015] [Daza, Calvo, and Figueroa-Nazuno 2016] [Guan et al. 2020] [Guan et al. 2021, 6396] [Chhun et al. 2022, 5797]).[6] Since the goal of story generation is to mimic human intelligence and creativity, “human-written” stories are used as a yardstick against which the success of the program is measured. For example, the evaluation scores of the stories can be used as an indicator of the program’s effectiveness [Liu et al. 2020, 1731], “narrative intelligence” [Chhun et al. 2022, 5799], or the stories’ suspensefulness [Cheong and Young 2015]. By contrasting the “human-generated” and computer-generated texts in this way, the human and the machine are positioned as separate entities. 

A study conducted by [Daza, Calvo, and Figueroa-Nazuno 2016] is an especially interesting example of this phenomenon in terms of authorship, because it juxtaposes experimental works by authors and texts that are generated with the program developed by the researchers. Image 1 shows a list of texts that participants were asked to evaluate without knowledge of their authorship. The picture depicts a dichotomy, where authors such as James Joyce are named but the texts generated by the program are put under the generic label of “machine.” This separation of the two is further repeated in the results, where the performance of “artificial texts” and “human texts” in the evaluation is presented. This division reflects the perception of authorship as a technical either-or question: A text is either is or is not a result of a text-generation process.

The picture also shows that multiple experimental works are included in the comparison. The authors of the article state that this is because they want to stress that the texts they produced are “at least as experimental” as the included human written works [Daza, Calvo, and Figueroa-Nazuno 2016, 18]. However, the article does not discuss in more detail how the experimental texts in question relate to authorship. William Burroughs, for instance, famously aimed to obfuscate authorial authority in his works: He has described writing Naked Lunch without conscious control [Heal 2016], and suggested that he used the cut-up method in writing it [Gontarski 2015, 173]. In this way, Naked Lunch and Burrough's attempts to minimise his own conscious role in writing it provides a kind of precursor for similar pursuits in story generation. This nuance is lost in the human-machine dichotomy.

Furthermore, the line between “the human” and “the machine” is often blurrier than the dichotomy suggests. For one, the “computer-generated” works are often largely shaped by the developers of the program. For example, in the case of Suspenser, the developers have formulated a list of fifty-seven events that the program can pick from to include in a storyline, which has an effect on the produced stories. To evaluate the results, Cheong & Young have also asked a “human author” to construct stories based on the same pool of events, and these stories are then read and evaluated by study participants [Cheong and Young 2015, 46]. The categories “human-generated” and “machine-generated” do not accurately describe the process of creating these stories. Both types of stories are shaped by the programmers, who have chosen what events are possible in a story, regardless of whether the storyline itself is constructed by the human author or the Suspenser.

Even with programs where the role of the developers is less central than with Suspenser, they shape the produced stories significantly. The programs need some sort of rules to function, for example concerning the length and genre of the stories, the kinds of texts that the program should imitate, and whether the program should focus mainly on coherence, suspense, or literariness. These rules are designed and put in place by the developers of the programs. As an example, [Daza, Calvo, and Figueroa-Nazuno 2016, 19] state that their program produces output based on a single word that the user of the program gives to it as a prompt, and that they do not have any control over the substance of the stories produced. In a limited technical sense — who or what has constructed the text at the sentence level — this might be the case. However, if authorship is understood more broadly and includes coming up with the concept of the work, the developers can be regarded as co-authors of the texts. The researchers undeniably shape the production of the stories by setting the program goals such as coherence and literary style, and by choosing the training data.

Furthermore, “purely” human-written texts are becoming an increasingly difficult category to identify, given the recently risen popularity of LLM based programs. As LLMs have become more widely known and available, creating partly computer-generated literary works has been possible for a larger number of people. This is evident, for example, on book platforms such as Amazon that have experienced acceleration of publishing rates [Cabezas-Clavijo et al. 2024].[7] The fact that human and machine writing are increasingly merging outside of the research context is making it increasingly difficult to distinguish between “human-written” and “computer-generated” texts. Especially when story generation programs are examined as a form of literature, a more accurate approach could be to acknowledge that the outputs are not void of human influence, and that whether they are written by a human or a machine is a sliding scale.

As demonstrated, the main goal of story generation programs is to imitate human intelligence, and “human-written” stories are used as a measure of the success of computer-generated stories. This aim is reflected in attempts to reduce the “mechanical” feel of the generated texts. These attempts are most clearly seen in evaluations of the stories, where the desirable and undesirable features of the produced stories are defined.

Story features that are deemed “undesirable” are discussed especially in the evaluations of the quality of the stories. These features include, most notably, repetition and incoherence in stories (see e.g. [Yao et al. 2019, 7383–7384] [Tambwekar et al. 2019, 5986] [Guan et al. 2020, 93] [Herrera-González, Gelbukh, and Calvo 2020, 82] [Guan et al. 2021, 6396]). These qualities are often discussed in contrast to the stories written by people and thus are deemed too machine-like. Bottoni et al. describe how stylistic features such as repetition can make the text sound “mechanical”:

Sentences in stories are affected by the sentences around them, and most readers can quickly notice if they don’t flow together “correctly”: e.g. “Noah went to the store. He went to the store to buy eggs.” or “Sara saw a spaceship. She boarded the spaceship.” There are quotation marks around “correctly” because prescriptively these sentences have correct grammar. However, they sound robotic and inhuman — “incorrect” in a rhetorical sense. [Bottoni et al. 2020, 24]

The style of the text — in this case, the monotone syntax of the sentences — is deemed incorrect because of its “mechanicalness.” Repetition in story generation can mean repeating the same syntactic structure, as in the case of the program by [Bottoni et al. 2020], but also what developers characterise as repetitiveness of plot, where the same event is depicted repeatedly.

Incoherence in stories is another feature that is often considered undesirable. What researchers regard as incoherence can manifest in the stories in various ways. For example, [Xu et al. 2018, 4313] note that some of the stories produced have a “chaotic timeline,” meaning that the stories go backwards in time. Conflicts of logic or impossible events can also result in incoherence. Guan et al. give the following extract of a story with conflicting logic:

He noticed a car in the road. He decided to stop. He got out of his car. He drove for half an hour. [Guan et al. 2020, 103]

In this example, beginning to drive the car after exiting it is what creates incoherence. Researchers can also define incoherence as general “chaoticness” in stories. For example, Liu et al. characterise one of the stories generated by their program as “inferior” and “chaotic,” because it lacks a clear main character [Liu et al. 2020, 1731]. As in the case of repetition, the researchers often attribute the incoherence to the computational origins of the text. For example, when describing types of incoherence in stories, Guan et al. point out that “unrelated events and conflicting logic, or globally chaotic scenes” are common in natural language generation models [Guan et al. 2021, 6396]. Because coherence is regarded by the researchers as an integral feature of successful stories, they try to avoid all instances of incoherence.

This approach, which emphasises the necessity of reducing features that supposedly give away the text’s computational origins, is different from those of many other genres of electronic literature. Many writers specifically use computational methods because of the features that they can bring into the text. For example, random generation of text can be used to create surprising associations [Hayles 2008, 26]. In fact, “randomness,” which can be computationally simulated in many ways, has been an important part of the formation of electronic literature and other forms of digital artistry [Montfort et al. 2013]. In these cases, the “chaoticness” is integral to the works, which is in some ways an opposite approach to story generation.

Furthermore, in computer-generated literature outside of story generation research, the fact that the text is a result of a joint effort of human and computer can be embraced, as the interplay between the two is often what makes the work interesting. The interaction between the programmer and the program can be intentionally made visible to the reader as an integral part of the work, either by revealing the code that produces the text [Marino 2020, 2015–220] [Hongisto 2023], or by letting the reader uncover the underlying principles of the program by observing multiple iterations of texts produced by the same program [Schoenbeck 2013]. The value that the human-computer interaction brings to the text can also be partly unintentional. Pressman et al. note that in digital culture, surprising results or glitches are often appreciated either as an aesthetic quality of the work or as an opportunity to catch a glimpse of the structure of the program [Pressman, Marino, and Douglass 2015, 41]. Whereas with these kinds of works the ambiguity of authorship is something to be examined and pondered by the reader, the goal of story generation is to eliminate such considerations. As with the Turing test, the aim can be that of deception [Natale 2021, 28] — to trick the evaluators into thinking a text is written by a human (see e.g. [Daza, Calvo, and Figueroa-Nazuno 2016]). Contrasted with the approaches that highlight the human-machine interaction, the aim to imitate human writing by avoiding “mechanicalness” in texts can be distinguished as a feature that characterises story generation as a subtype of electronic literature,  at least in the form that the research material represents.

Considering the aim of mimicking human writing in story generation, avoiding features that might lead the reader to believe that a text is computer-generated is understandable. However, categorically avoiding features such as repetition and incoherence may not always serve this goal, as it does not take into account their value as stylistic devices. The value of the unexpectedness of computer-generated outputs can be observed through stories that have been labelled failures. The examples of “unsuccessful” stories are often intriguing, sometimes arguably more so than their “successful” counterparts. Take for example these two short stories generated by the model by [Guan et al. 2020] below:

[MALE] was driving around in the snow. Suddenly his car broke down on the side of the road. [MALE] had to call a tow truck. The tow truck came and took [MALE] home. [MALE] was happy he was able to get home. [Guan et al. 2020, 103]

[MALE] was on thin ice with his job. He had a friend over to help him. [MALE] was able to hold his breath the entire time. he was so cold that he froze in his tracks. [MALE] finally felt good about himself. [Guan et al. 2020, 104]

Out of these examples, the former is the “successful” story and the latter the “unsuccessful” one, as Guan et al. aim to produce coherent stories and have labelled the latter story as “chaotic” [Guan et al. 2020, 103–104]. However, there is something enthralling in the story’s use of metaphors that have to do with freezing and that merge with the protagonist feeling cold, and there is an element of surprise in the protagonist’s contentment at the end of the story. It is not a coherent story by the standards of [Guan et al. 2020], but it arguably generates more interest and questions, as well as providing more room for interpretation than the “successful” story.

Another well-known case of similar “unsuccessful” stories turning out more amusing than the “successful” ones are Tale-spin’s failed stories [Meehan 1976]. The so-called “mis-spun” fables generated by the program first raise genre expectations only to soon drop them, which creates a humorous effect. These “unsuccessful” stories have been circulated much more widely than the successful ones, partly because of their entertainment value [Bolter 1991, 180] [Aarseth 1997, 131] [Wardrip-Fruin 2009, 130]. The popularity of these stories is an example that there can be value in the stories that the programs' creators have deemed unfit.

Similarly, avoiding all instances of repetition may not effectively serve the aim of producing texts that imitate human writing. Yao et al. give the following example of an unsuccessful story with repetition on the level of plot:

Anna was cutting her nails. She cut her finger and cut her finger. Then she cut her finger. It was bleeding! Anna had to bandage her finger. [Yao et al. 2019, 7380]

It is easy to understand why this story is considered a failure, since Yao et al.) aim to describe “a sensible sequence of events” [Yao et al. 2019, 7378]. However, as they also want to “mimic human practice in real world story writing” it is worth pointing out that by forbidding repetition, they dismiss its potential as a stylistic device [Yao et al. 2019, 7379]. In the above story, for example, the repeated finger cutting brings a chilling tone to the text, implying self-harm.

One could argue that pointing out the humour and tellability of these stories is beside the point of story generation. The difficulty of producing long, coherent texts without repetition using natural language processing models is well established [Fan, Lewis, and Dauphin 2018] [Ammanabrolu et al. 2021] [Peng et al. 2022]. From this perspective, it is understandable that the researchers want to develop these capabilities in story generation models. More generally, the critique of ignoring common literary devices in the development of story generation programs applies best to research where the intent is to produce a literary output. Still, avoiding incoherence and repetition entirely and dismissing them as “mechanical” features does not seem consistent with the goal of mimicking human writing. Although computers are prone to producing texts with such features, they are not inherently machine-like but common stylistic devices in literature. Even when randomness is not used as a foregrounded literary device, is some amount of randomness generally desirable even in computer-generated creative texts [Hua and Raley 2020]. If  these features are left out of writing, part of the expressive power of fiction is lost. 

The approach to authorship remains somewhat consistent throughout the material presented here. While the articles do not explicitly define the concept of authorship, a closer examination suggests that the term is used in a way that differs from its everyday use. The aim is to automatise the creation process of texts and to minimise the role of the human authors: a challenge that seems to be viewed as a technical one by the researchers. This view is reflected in the technical use of the term authorship, mostly referring to who (or what) has composed which parts of the text, while the aspect of who (or what) is responsible for the work is not discussed.

This narrow perspective on authorship affects how the human-machine relations in the creative process of the texts are discussed. The stories are most often referred to as computer-generated and contrasted with texts written by humans. This view differs from a large part of computer-generated literature, where authorship is often regarded as a broader concept, and the role of the designer of the program as an author is often empathised more.

This difference in approach to authorship is also reflected on the content of the produced texts. One consequence of the dichotomy of human and machine authorship is that the stories can often be evaluated based on the absence of features that are regarded as “mechanical,” mainly incoherence and repetition. Developing programs that are less prone to producing texts with such features undoubtedly improves the understandability and fluency of texts, but without them texts mimicking human writing would remain lacking in stylistic versatility.

Examining the approach to these “mechanical” features in texts reveals a larger issue within the evaluation of computer-generated texts. Many researchers stress the difficulty of reliably evaluating the quality of the stories (e.g. [van der Lee et al. 2019, 355] [Herrera-González, Gelbukh, and Calvo 2020, 86] [Chhun et al. 2022, 5794]). Indeed, the task of objectively and quantifiably evaluating stories is fundamentally challenging since it enters the territory of personal taste and aesthetic judgements — an area where absolute claims of quality are impossible. However, incorporating more contextual information in the process of both development and the evaluation of the program (Whom are the stories aimed at? What types of reading situations are they meant for?) could help in formulating the evaluation criteria. Different media, text types, genres, and reading contexts all affect which stylistic and narrative features are desirable, and thorough knowledge of these topics could help in designing programs for different purposes. Developing evaluation criteria for the stories might be an area where story generation could benefit from an interdisciplinary approach.

The authorship of the produced texts can be discussed differently depending on whether story generation is placed in the realm of science or art. On one hand, in the sample of articles analysed in this article, the programs or produced texts are regarded not as literature but as technical improvements of natural language generation technology. This view explains the rather technical outlook on authorship. On the other hand, story generation can be approached as a genre of electronic literature. This approach evokes a more diverse image of authorship than can be inferred from story generation research: It refers not only to how the text technically came to be, but to who (or what) is accountable for it. Outlining the extent to which story generation aims at literary expression is an interesting question from the perspective of both computer science and electronic literature studies. In computer science, considering the question would mean defining the intended use of the programs more precisely, and if the aim is literary expression, taking into account the wide range of features of narrative and fiction. In the context of electronic literature research, on the other hand, story generation research must be viewed as a broad and varied field, some parts of which can be more justifiably included in the scope of electronic literature than others. Navigating the research output of story generation research and its relations to electronic literature — and literature more broadly — deserve close examination also in the future.

This research was funded by Emil Aaltonen Foundation (Emil Aaltosen Säätiö).