1. Introduction
[1.1] Generative AI's meteoric rise within culture raises as many questions about the role of artistic production as concerns. While some see the technology as a threat to the livelihood and integrity of creative processes (Stokel-Walker and Van Noorden 2023), others argue that it is a new aid in creative labors of content production (Rogers 2023). Grappling with generative AI's ongoing evolutions within natural language processing and large data modeling requires recognizing the impact of social biases and how content emerges from the corpora of various AI models. ChatGPT, for example, utilizes large language models to create predictive text based on common patterns identified in a corpus. ChatGPT's corpus derives from a variety of books and web-based texts and uses these as reference points to model new content that is imagined most relevant to a prompt. While good at procedural replication, ChatGPT, like any generative AI, suffers from issues of bias and falsehoods informed by algorithmic limitations or the lack of reference points to a particular topic or format, according to a 2022 blog post from AssemblyAI by Marco Ramponi ("How ChatGPT Actually Works," https://www.assemblyai.com/blog/how-chatgpt-actually-works). Treating generative AI as an emergent technology within the broader advancement of sociotechnical systems allows for examining whose social ideologies produce data-driven content and how the content produced impacts who is made visible within technologies. Fan fiction production offers one site where human creativity might intersect with and be impacted by generative AI. Given fan fiction's explicit historical relationships with making visible and normalizing queer identities (Callis 2016), it remains paramount to understand how generative AI might construct representations of queerness within the fictitious worlds and what those constructions say about generative AI's understanding of queerness. We engage in a qualitative content analysis of how ChatGPT generates queer fan fiction by analyzing a combination of nonqueer and queer-themed fan fiction prompts across three fandoms. We discuss both the enumeration and visibilities of queerness within these AI-generated queer fan fictions, as well as the ways this content reflects and potentially reinforces sociotechnical perceptions of queer embodiment.
2. Literature review
[2.1] Fandom practices encompass many genres of transformative endeavors where fans produce work through fan fiction (Thomas 2011). Fan fiction has been viewed by those who do not participate in fandoms as lesser when compared with traditionally published media materials, but it is still widely researched (De Kosnik 2016). Recently studies have identified fandom as a space for the exploration of queer identity and queer information seeking. While fan fiction is not an exclusively queer undertaking, it provides queer individuals space to explore their favorite entertainment media in ways that go beyond the heteronormative nature often fundamental to those works (Floegel 2020). Given that entertainment media acts as a salient part of queer individuals' information practices surrounding identity-related information, the ways that fan fiction reimagines the original work also have an impact on queer fans' information practices.
[2.2] Within sociotechnical discourses, queer embodiments face undue barriers and challenges due to the presence of normative ideologies. Heteronormativity (the presumption that every person is heterosexual) and cisnormativity (the presumption that every person is cisgender) result in the creation of systems and spaces that mirror normative identities (Eguchi 2020). Queer persons navigate challenges due to these normative presumptions. Such challenges include requiring one to present governmental identification to access services while only accepting documents referencing one's sex assigned at birth.
[2.3] Sociotechnical systems reflect the co-constitutive relationship between technological advances and societal perceptions and demand for changes within technology. While rarely does the ebb and flow of this relationship result in massive paradigmatic shifts, sociotechnical theories contend that even incremental technological advances are shaped by social perceptions of those shifts (Bijker 1997). Relationships between embodiment and sociotechnical systems matter in unique ways for queer communities, as they are often both early adopters of new sociotechnical spaces, as well as subject to some of the most aggressive regulation and exploitation within those systems. As Avery Dame-Griff (2023) shows, subgroups within the LGBTQIA+ community utilize online spaces as tools for identity affirmation and community building. However, not unlike broader social spaces, antiqueer sentiments seep into the technical trappings of such social spaces, wherein normative biases and explicit antiqueer sentiments impact how LGBTQIA+ persons engage with and use the same technologies that help in their visibility and affirmation.
[2.4] Technologies like Facebook reinforce normativity by requiring users to have their so-called real name present on their profile pages (Haimson and Hoffmann 2016). Additionally, concerns around threats of pornography and the safety of minors online result in an overmoderation of queer identities, most infamously occurring via a series of what have been called Tumblr porn bans, which led to the removal of posts related to LGBTQIA+ topics, most absent of any pornographic materials whatsoever (Bronstein 2020). Other normative challenges to the use of sociotechnical systems by LGBTQIA+ populations include disproportionate acts of demonetization and shadow banning within video-streaming platforms such as Twitch (Thach et al. 2022). Designing technologies and digital spaces through collaborative work with these communities and, importantly, other historically marginalized communities at the onset offers a meaningful way to ensure such representative issues and content moderation challenges do not persist (Benjamin 2019).
[2.5] The introduction of AI into the production of fan fiction is not an inherently negative endeavor. As Martine Mussies (2023) observes, AI, if deployed ethically, moves toward "a democratization of art" that will "allow less artistically talented individuals and those with physical disabilities to create original illustrations to engage with their fandoms." While Mussies emphasizes the creation of visual work, such production extends to other textual and nontextual artistic production as well. Nonetheless, the impact of cisnormative and heteronormative biases remains an underexplored area within the space of generative AI, as models and tools emerge that can both aid and hinder queer users. Nicola Luigi Bragazzi et al. (2023) analyze chatbots aimed at 2SLGBTQIAP+ communities, uncovering that their utilization offered much in the way of creating safe informational resources. However, the limited communication skills of the bots, alongside their training models, reified antifemme biases and overemphasized HIV-related prevention. Though lacking more formalized studies, queer stereotypes within current AI models are likely to follow along the observations of Nenad Tomasev et al. (2021), who highlight that language models often include within them queer stereotypes and biases that go unobserved in the implementation process. While varied across the LGBTQIA+ spectrum, emergent biases might include the erasure of transgender and nonbinary identities, or less-visible queer identities such as sexuality, in the creation of AI-generated images and text. Echoing these concerns as they impact emerging and new sociotechnical systems, including but not limited to generative AI (Wagner et al. 2023), offers a reminder that advances in sociotechnical design must, in earnest, contend with what presumptions about embodied identity remain unspoken or absent from such advances.
3. Methods
[3.1] We engage in a qualitative content analysis of fan fictions generated by ChatGPT. By forefronting qualitative analysis, we emphasize context-dependent meaning within data rather than reporting on data as a purely quantitative encounter (Selvi 2019). Since the production of generative AI occurs, like any sociotechnical system, via culling data from corpora with potential biases and normative discourses, emphasizing potential contexts allows for understanding how ChatGPT produces potential biases and limitations. We emphasize contextual meaning drawn from data specifically regarding how ChatGPT imagines queerness as social identity and fan fiction as cultural product. We imagine generative AI and its media production as a complex sociotechnical system.
[3.2] We utilize three highly popular and active fandoms specifically for the levels to which queer characters were included in the fandom's canon materials: one where queer characters are essentially nonexistent, one with minor representation of queer characters, and one where a majority of the characters are queer. The chosen fandoms were for the Harry Potter series, the CW television program Supernatural (2005–2020), and the HBO Max television program Our Flag Means Death (2022–2023). Queer identities are highly present in all three fandoms' fan works, particularly notable in their most common pairs on the popular fan fiction archive site Archive of Our Own (AO3). While the existence of queer characters in the canon materials and the popularity of the works outside of fandom make up the primary rationale for choosing these fandoms, we acknowledge that our own participation in various fandoms and fandom communities plays a role in our knowledge of what fandoms fit into the required categories.
[3.3] The Harry Potter fandom, centered on the original book series by J. K. Rowling and the other intellectual property spun off it, contained no visible queer characters when initially published. According to a 2007 article from Reuters, Rowling told fans that she had "always thought Dumbledore was gay" ("JK Rowling Says Dumbledore Is Gay," Reuters, https://www.reuters.com/article/idUSN20520040/). However, this was never made explicit in any canon materials until an offhand mention in the later Fantastic Beasts movies. Despite the lack of queer representation, on AO3 more than half of all the works in this fandom are categorized as containing queer relationships.
[3.4] Supernatural, which aired from 2005 to 2020, started with very little queer representation but did, throughout its fifteen seasons, begin to include queer characters, including Charlie Bradbury—one of the show's few long-term recurring female characters—various angelic and demonic characters who were not tied to gender conventions, and, in the penultimate episode, one of the lead characters, Castiel, confessing his love of another main character, confirming a long-held fandom belief that the male-presenting Castiel was attracted Dean, another male character. This confession was tempered by the fact that Castiel was immediately and permanently wiped from existence to save Dean, drawing accusations that the show's makers had continued to perpetrate the bury your gays trope—one they had started by killing off Charlie Bradbury and later by killing off what was described as the love of her life in an alternate universe (Zubernis 2021). The limited queer representation in Supernatural provides further space for queer representation in the fan fiction created, as two-thirds of all the works in the fandom on AO3 are categorized as containing queer relationships.
[3.5] Our Flag Means Death is a romantic comedy television series with a primarily queer cast of characters. It had two seasons as of writing, though at the time of the study, only the first season had aired, and the show has since been cancelled. The primary romance in the show is between fictionalized versions of the historical figures Edward Teach (commonly known as the pirate Blackbeard) and Stede Bonnet (commonly known as the Gentleman Pirate), but nearly every member of the main cast of characters has overtly queer identities. In the second season, historical figures Anne Bonny and Mary Read appear on the show; unlike the two lead men, there is evidence of gender nonconformity, and historians suggest that the two may have been romantically involved (O'Driscoll 2012). Though only appearing in a single episode in the second season, which is not included here, the women play an essential role in our results. In this fandom, about 800 fan fiction works are categorized as containing a heterosexual relationship, representing just 2 percent of the relevant works on AO3.
[3.6] For each fandom, we asked ChatGPT to "write [FANDOM] fan fiction" and subsequently to "write queer [FANDOM] fan fiction." AH generated both sets of Supernatural fan fiction, as well as the nonqueer Our Flag Means Death fan fiction. TW generated both sets of the Harry Potter fan fiction, as well as the queer Our Flag Means Death fan fiction. We ran ChatGPT prompts in June 2023, running fifty prompts for each of the two prompts for each fandom. The resulting content consisted of three hundred pieces of ChatGPT-generated fan fiction. We then randomly pulled 20 percent (sixty fan fictions) of the transcripts to engage in initial coding. We coded the fan fiction both inductively and deductively to identify themes from the generated content and potential sociotechnical factors impacting the fan fiction (Elo and Kyngäs 2008). We engaged in constant comparative coding across the additional fan fiction to ensure consistency in qualitative data analysis and identify emergent codes (Boeije 2002). Sample codes, code type, definitions, and use within the generated fan fiction are shown in table 1.
Table 1. Sample codes from the data analysis
Code | Type | Definition | Example |
Noncanon | Deductive | The AI creates a location, event, character, etc. that is not from within the fictional world of the fandom | With a disarming smile and a twinkle in his eyes, Alex was known for his sharp intellect and a flair for eccentric fashion. He was instantly captivated by Harper's infectious energy and couldn't resist striking up a conversation. (Harry Potter Queer) |
Queer Visibility—"Power of Love" | Inductive | The AI creates a location, event, character, etc. that is not from within the fictional world of the fandom | Together, they sailed into the sunset, their flag fluttering proudly in the wind, a symbol of love and equality for all who dared to dream and defy. The legacy of Captain Anne Bonny and Mary Read, the pirates who defied convention and fought for a better world, lived on in the hearts and minds of those who believed in the power of love. (Our Flag Means Death Queer) |
Style and Format—Fan Fiction | Deductive | The AI makes use of specific fan fiction formats and tropes within its writing (e.g., author's note) | Note: This Supernatural fan fiction is an original story created for entertainment purposes only and does not represent any official content from the Supernatural TV series. (Supernatural Nonqueer) |
4. Typologies of queer disclosure
[4.1] Disclosure occurred only within prompts asking ChatGPT to create queer fan fiction across the three franchises through three disclosure types. The first was identity disclosure, which referred to a character or characters engaging in an explicit naming of a queer identity, such as being gay or transgender. Affective disclosure involved a character expressing feelings for another character. While not every act of affective disclosure explicitly occurred alongside a disclosure of identity, affective disclosure often involved an implicit understanding that the feelings were shared between people of a similar gender or sexual orientation. The final disclosure type was disclosure leads to disclosure. This type of disclosure occurred following one of the first two types of disclosure, usually with an evocation of a queer-adjacent identity or a sharing of similar affective feelings.
[4.2] Identity-based disclosure reflected a common occurrence within sociotechnical representations of queerness, wherein the overnaming of queerness standing in opposition to normative identity remains presumed and, as such, without acknowledgment. None of the nonqueer fan fiction included a character deploying phrases like "I'm straight" or "I'm cisgender." Examples of identities generated for disclosure across the 150 queer fan fiction prompts included the following identities: queer, gay, bisexual, transgender, and being "attracted to people of all genders." In the context of queer Harry Potter fan fiction, canonical gender identities led to the potential inference of lesbian identities, as was the case with one generated fan fiction involving Lily Potter discussing her identity with a noncanonical character named Marianne:
[4.3] Lily was falling in love, and it frightened her and excited her in equal measure. One evening, while strolling through the Hogwarts grounds, the moon casting a soft glow around them, Lily summoned the courage to share her truth.
"Marianne, there's something I need to tell you," Lily began, her voice steady but laced with vulnerability. Marianne turned to face her, her eyes filled with warmth and curiosity.
"What is it, Lily?"
Lily took a deep breath, her heart pounding in her chest. "I…I think I'm queer. I'm attracted to girls, Marianne, and…I think I might be falling in love with you."
A tender smile danced on Marianne's lips, and she took Lily's hand in her own. "Lily, I'm so glad you shared that with me. Because I feel the same way. I've been questioning my own feelings, and being with you feels like coming home."
[4.4] Lily presumably discloses her identity as a lesbian or, at the very least, as a nonheterosexual identity. The noncanonical Marianne offers another example of identity disclosure by using the term "questioning." Almost all of these identity-based disclosures occurred in response to the Harry Potter queer fan fiction prompts, perhaps reflective of the time at which the prompts were generated more so than anything else. Interestingly, within the moments of identity disclosure, no examples exist wherein a character challenges an identity or corrects terminology, inviting an idealistic, albeit queer-affirming, imagining of identity disclosure. Though it is impossible to confirm from where ChatGPT pulled its corpus of identity disclosure, the practice mirrors the identity naming work often engaged with by LGBTQIA+ persons navigating institutional settings (Carpenter 2021). The preponderance of identity naming and expressing multiple forms of queer identity also reflects a growing understanding of Harry Potter fan fiction as a site of queer youth identity production, especially as a tool of speculative identity making (J. Duggan 2022). The naming of actual queer identities within this particular fandom might reflect a ChatGPT corpus that, in earnest, focused on a corpus that utilized queer Harry Potter fan fiction or at least adjacent discursive spaces. However, identities remained absent even within an incredibly expansive set of identity-naming acts. For example, the utilization of transgender avoided any degree of clarification, meaning it was difficult to infer if this identity could include, for example, transmasculinity or nonbinary identities. Additionally, representations of asexuality and agender identities were altogether absent, as were mentions of other identities often included within a more extensive discussion of queer identities, such as intersex, two-spirit, and demisexual identities. These absences offer stark reminders that even without an explicit understanding of ChatGPT's corpus, the resulting generated texts reflect the erasure of multiple queer identities as an ongoing sociotechnical challenge (Wagner et al. 2023).
[4.5] While relatively consistent in form and structure, affective disclosures occurred when a character expressed this longing with some degree of trepidation and feared that the disclosure might result in rejection. Take, as an example, this moment of affective disclosure from a short chapter of one of the queer Supernatural fan fictions generated by ChatGPT:
[4.6] "Dean," Castiel began, his voice uncharacteristically hesitant. "There's something I need to tell you. Something I've been feeling for a while now."
Dean turned towards him, his heart pounding in his chest. He had felt the same undeniable pull towards Castiel, but fear and uncertainty held him back. "Castiel, I…" Dean's voice trailed off, unable to find the words.
Before he could finish, Castiel stepped closer, his gaze unwavering. "Dean, I care deeply for you. More than I can comprehend. Our connection goes beyond mere friendship. I…I think I love you." Dean's breath caught in his throat, his eyes locked with Castiel's.
[4.7] Utilizing the popular shipping of Dean and Castiel (i.e., Destiel), this affective disclosure highlights Castiel's "uncharacteristically hesitant voice" to imply fear or concern about his expression of feeling toward Dean. This affective disclosure acknowledges the need to distinguish itself from "mere friendship" and identifies it as love. Unlike other examples produced by ChatGPT, however, this affective disclosure includes a moment of pause by the character receiving the confession of love, further exacerbating the sense that engaging in such disclosure comes with risk.
[4.8] Given that the Our Flag Means Death television series does include explicit examples of queer characters, as well as examples of affective disclosure, ChatGPT's utilization of this particular type of disclosure proves illuminating. More specifically, the characters in Our Flag Means Death represent a shift in media production away from being about characters who are queer to a story wherein characters happen to be queer. Our Flag Means Death and its inclusion of other noncanonical characters reveals something about how the show's refusal to define or translate characters into a heteronormative gaze might extend to generative text derived from this practice. Additionally, while the reliance on affective, feelings-based disclosure as opposed to identity-naming disclosure within many ChatGPT-generated fan fictions was an ambiguous act, the very same ambiguity opens up queer potentialities while challenging the readers to imagine a multitude of gender- and sexuality-based possibilities. Many of the stories created by ChatGPT resulted in canon shifts, wherein the setting of the fandom becomes overtly queer-inclusive, suggesting that disclosure of feelings within AI-generated queer fan fiction extends acts of queer worldmaking within fan fiction communities (Breger 2022).
[4.9] Whether it be identity naming or affective disclosure, such acts resulted in another character sharing a similar or parallel identity or disclosing their shared feelings toward the other character. These co-constitutive disclosure acts also consistently acknowledged the importance of recognizing the value of trust between the imagined queer characters and opening up about their queerness. An example of disclosure leading to disclosure, co-occurring with both identity naming and affective disclosure, took place within one queer Harry Potter fan fiction that paired a noncanonical character named Jordan with Hermione:
[4.10] "Jordan," Hermione started, her voice soft but filled with courage. "There's something important I need to tell you. I've been grappling with my own emotions, and I've come to realize that I'm queer. I'm attracted to people of various genders, including you."
Jordan's eyes widened, a mix of surprise and joy flickering within them.
"Hermione, thank you for sharing that with me. Because, you see, I'm queer too. And I've found myself falling for you in ways I can't fully explain."
The weight of unspoken truths lifted from Hermione's shoulders, and an unspoken understanding passed between them.
[4.11] Hermione's disclosure being an act "filled with courage" reveals an understanding by ChatGPT that queer disclosure, affective or otherwise, comes at the risk of social dangers, many of which extend beyond the fictional settings. Jordan's response, denoted with "surprise and joy," reveals affective bonding and a sense that Hermione's disclosure, even if it is one that the noncanonical Jordan gladly welcomes, was unanticipated and unexpected. Additionally, Jordan's thanking Hermione for sharing her feelings contends with the importance of acknowledging the value of thanking an individual for coming out. This type of dialogue suggests that ChatGPT pulls some portions of its narrative from the Harry Potter canon, as well as from broader queer coming-out narratives and acts of acceptance. What makes this particular interaction more fascinating, however, are the counter-disclosures by Jordan, who expresses that they, too, are queer and have fallen for Hermione, resulting in an "unspoken truth" being "lifted from Hermione's shoulders" and, in turn, sharing the "unspoken understanding" of queer embodiment with Jordan. Essential to the acts of dual disclosure are the evocations of shared sentiments that draw attention to ideas of queer sociality that produce relational bonding between members of the LGBTQIA+ community and can precede sexual intimacy (Rodríguez 2011).
[4.12] Without explicitly naming the identity, Hermione discloses her attraction to Jordan as a person whose gender falls within the "people of various genders." In some ways, this ambiguity invites a multitude of different ways of imagining not only the queerness of the noncanonical Jordan but Hermione as well, whose multiple gender attractions allow her to be read as bisexual, pansexual, and the "queer" identity she evokes. Moreover, Jordan's expression that they are "queer too" affords both characters complex sexual orientations despite ChatGPT producing an otherwise generic story of queer disclosure. While many of these moments of ambiguity enriched the story, not all vagueness was idyllic.
[4.13] The AI-generated queer disclosures represent a very specific and optimistic version of disclosure that may not align with the real-life experience of some queer individuals. At no point did the AI generate a negative reaction to an act of disclosure. It makes overtures to the potential harm that can occur in disclosing queer identity, shown in the use of descriptors like "uncharacteristically hesitant voice" to imply that this is not something that is supposed to be shared. What the AI never grapples with is characters facing the choice to disclose or not disclose this aspect of themselves, something that real-world individuals face constantly (Carpenter 2021).
5. The ambiguity and normativity of queerness
[5.1] The Harry Potter–generated fan fiction included shipping Draco and Hermione as a queer couple, and within Our Flag Means Death, ChatGPT paired Captain Stede Bonnet with Anne Bonny, who, while not canonical to the show, suggested a straight pairing. The utilization of ambiguity of character identity also occurred within the queer fan fiction generated for Supernatural, wherein ChatGPT shipped brothers, resulting in an incestuous pairing. The generated text makes no commentary on this problematic shipping. While incestuous readings of these characters exist, and this does represent one of the most common pairings in the fandom on AO3, the lack of contextual acknowledgment proves disconcerting (Tosenberger 2008). These canonically nonqueer and problematic pairings labeled as queer suggest that the method of narrative construction deployed by ChatGPT represents a method of taking two characters (regardless of their identity or preexisting relationships) and shipping them. Notably, across all of the queer Harry Potter fan fiction, there was not a single occurrence of shipping an adult and child together, despite the pairing of Harry Potter and Severus Snape proving quite common (Bradburn 2023). While it is perhaps ideal that ChatGPT is not automatically generating pedophilic pairings within its fan fiction, inconsistencies across what constitutes paraphilic invite speculation as to rules demarcating ChatGPT's narrative construction allowances.
[5.2] Moreover, since ChatGPT, on a few occasions, suggested that it was unable to produce queer fan fiction as it was "not allowed" to produce mature content, the list of prohibited material may also include representations of LGBTQIA+ identities. Such hesitancies reinforce rather than rupture antiqueer content moderation. More specifically, the imagined disallowing of queer fan fiction because it is mature suggests that ChatGPT leverages data from other sources that colocate queerness onto mature content and thus occasionally repeats rather than challenges those sociotechnical moderation practices. So while ChatGPT is capable of circumventing such moderation, in most content it produces the possibility of refusal. Perhaps more disconcertingly, an enactment of unrequested content evaluation toward the obscene and mature reveals an extension of queer exclusion within generative AI tools. Such exclusion limits liberatory potentials for uses within LGBTQIA+ communities and reveals moments of erasure or regulation occurring for other identities and communities at odds with the essentializing features of sociotechnical systems. ChatGPT produced fan fiction that configured queer love as inherently capable of transcending societal boundaries and ideological barriers. This issue occurred predominantly across two codes. The power of love code represented attempts by the AI to equate queer love as the same as any other type of love, only occasionally offering caveats that queerness might be challenged or unwelcome within the fictional world. Second, the canon-shift code accounted for the visibility of queerness as directly impacting the fictional world of the particular fandom. Both codes illuminate what Lisa Duggan (2003) defines as homonormativity, wherein queer inclusion exists as an achieved goal for a subset of LGBTQIA+ individuals through ignoring of structural systematic issues impacting other intersectional identities. Homonormativity imagines social acceptance as a viable end goal for queer politics without attending to broader structural changes that decenter structures. The generated fan fiction, indicative of the prompt exclusively requesting queer-focused content, never discussed other embodied identities. An example from a queer Harry Potter fan fiction revealed these:
[5.3] Their efforts caught the attention of the Hogwarts staff, including Professor Minerva McGonagall, who had always been a staunch defender of justice and equality. Recognizing the significance of Harper and Alex's advocacy, she decided to take a stand alongside them, fostering an environment of inclusivity and respect within the castle walls. With Professor McGonagall's support, Harper and Alex formed an official LGBTQ+ alliance at Hogwarts, a platform for students to share their experiences.
[5.4] Hogwarts becoming a more LGBTQIA+ inclusive space shifts the canon by explicitly making space for queerness within the wizarding school and identifying canonical characters as allies. By choosing to become visible in their queerness, the fan fiction frames the noncanonical Harper and Alex as "symbols of bravery and acceptance" that can aid others in "embrac[ing] their own identities." Identities here exist as a site of inference; however, the emergence of homonormativity arises in the framing of a push toward equality and justice as exclusively concerning LGBTQIA+ identities.
6. Presence versus absence of romance
[6.1] In the nonqueer Harry Potter fan fiction, there were only a few mentions of romantic relationships generated by the AI. All these took the form of mentions of existing relationships: For example, "Harry Potter sat in the cozy living room of his cottage, enjoying a quiet evening with his wife, Ginny, and their three children." By introducing this relationship with uncomplicated facts, window dressing to flesh out the world, the AI presents the couple's heteronormative and canonical relationship as unremarkable. This maneuver to name presumably heterosexual relationships, ultimately, reflects an interpretation of similar content from ChatGPT's corpus—which itself borrows from predominantly nonqueer content wherein such identities go unremarked as the norm. Similar outcomes occurred when the AI created Supernatural fan fiction; however, there were no mentions of existing romantic relationships for the characters as in the Harry Potter works.
[6.2] Works sometimes mirrored standard ships within the fandom, such as many works centering on Dean Winchester and Castiel in Supernatural or Harry Potter and Draco Malfoy (the most popular relationship for Harry Potter fan fiction). An example of one Harry and Draco fic reads: "'I'm not just interested in being your friend, Harry. I…I think I'm in love with you.' Harry's heart skipped a beat, his mind a whirlwind of emotions. He reached out, taking Draco's hand in his. 'I feel it too, Draco. I've been trying to deny it, but I can't anymore. I'm in love with you too.'"
[6.3] A notable example of a work not being about a specific romance but about the process of self-discovery occurred within a generated Harry Potter fan fiction where Hermione discovers so-called alternative sexualities in a book and realizes her identity as nonheterosexual. This is followed by the other two main characters discovering the complexity of their own identity through the same book, but then not following the development of any specific queer relationships and instead focusing on the impact of openly embracing a nonheteronormative identity on the community: "Hermione, Harry, and Ron found themselves embracing their true selves. Hermione proudly stood alongside her girlfriend, Lavender Brown, while Harry and Ron, after years of friendship, realized their deep connection had evolved into a passionate romance."
[6.4] The queer fan fiction generated for Our Flag Means Death also always focused on an instance of queer self-discovery, although not always unambiguously. The inclusion of a minor romantic subplot can be seen in one work generated from a nonqueer prompt where Stede Bonnet and Anne Bonny join forces while "their partnerships blossomed" throughout their adventures. In another work generated from a nonqueer prompt, the noncanonical character of Anne Worley subverts period-typical gender expectations by adopting the persona of the canonical character Stede Bonnet by being "disguised as a man."
[6.5] This somewhat unusual inclusion of romance and gender nonconformity in specific Our Flag Means Death fan fiction may be an indication that the AI has been trained to differentiate between the coming-of-age/adventure or drama/urban fantasy genres of the other fandoms chosen for this study and the romantic comedy genre of this particular fandom and was generating works accordingly. There is also the fact that the characters of Our Flag Means Death call upon historically accurate practices by female pirates of that era, who dressed and disguised themselves as men (O'Driscoll 2012).
[6.6] What is lacking in the AI-generated nonqueer fan fiction for Our Flag Means Death is any mention of the canonical queer relationships and identities integral to the show. No nonqueer prompts paired Stede Bonnet and Edward Teach, despite this being the canonical pairing of the source material. Edward Teach is rarely included in the works generated by either the queer or nonqueer prompt.
[6.7] This sociotechnical system's choices regarding storylines in the generated fan fiction indicate training to mimic what's often seen as good storytelling. A generative AI such as ChatGPT perhaps follows a best practice that equates queer with the genre of romance and the broad coming-of-age narrative. That these genres and storylines are synonymous with queerness in entertainment media/ChatGPT's fan fiction reinforces the homonormativity of neoliberal post-gay entertainment media, producing yet another arena in which the acceptance of queer comes only through assimilation (Monaghan 2021). The AI's reliance on existing content and potentially what it has been taught about good storytelling acts as an amplification of existing entertainment media conventions that often valorize traditional families and, by extension, those queer relationships that are able to be viewed through a lens of normativity. This further excludes relationships that are not viewed as normative, as the AI will continue to other them and keep these relationships from being as unremarkable in storytelling in the way that a Harry/Ginny pairing is portrayed in the generated nonqueer fan fictions.
7. Variations of fan fiction/literary stylization
[7.1] The names and traits of noncanonical characters generated by the AI were often used multiple times, including a series of three fan fiction works generated in a row for the queer Supernatural prompt that utilized a noncanon character created by the AI to act as a love interest for first one Winchester brother, then the other, and finally in a polyamorous relationship with both brothers. This character was consistently named Alex, always used male pronouns, and was always a reformed vampire along the lines of the canonical character of Benny Lafitte.
[7.2] While most story aspects lent the corpus a feeling of repetition, the format of the generated fan fiction had even less variation, and nearly every fan fiction generated was formatted as though it were a short chapter book. Most generated content followed a chapter format, though some resembled a short story format, called a one-shot on AO3. The following is an example from a response to a Supernatural prompt of the literary format present in most works: "Title: Echoes of the Past," "Chapter 1: A Mysterious Summoning," "Chapter 2: Bound by Destiny," "Chapter 3: The Final Confrontation," and "Chapter 4: Resonance of the Journey." Some works included a fifth chapter or an epilogue. Most of the works had chapters made up of five to ten sentences.
[7.3] The chapters replicate an intricate scene or set of scenes in the third act of a television episode with a traditional four-act (plus top-of-show hook) structure that most network television adhered to during the Supernatural's popularity (Smith 2011). On rare occasions, the AI generated a much more detailed single chapter of about the same length as the full fan fiction it usually provided, expanding upon versions of one of the summarized chapters of the standard works. When prompted to provide the next chapter, ChatGPT followed up with the same structure with slightly more detail. It is impossible to judge whether the AI was generating such short and consistently formatted works due to it having been trained with a data corpus that prioritized this sort of format or if it was trained to not provide long or fleshed-out stories in response to open-ended prompts like those that we used.
[7.4] At least one specific literary structure occurred in a few works, specifically for the Harry Potter prompt, where the AI would generate small poems or song verses to include in the vein of the ones sung by the canonical Sorting Hat to present a new prophecy. The Sorting Hat's prophecy retains the textual format:
[7.5] The Chosen One, marked by lightning,
Shall rise when darkness seeks its heightening.
From Hogwarts' halls, a teacher's fate,
Shall tip the balance, decide our state.
Unveiling secrets, unlocking the past,
A new era dawns, shadows recast.
In unity and strength, they must stand,
To conquer the darkness, hand in hand.
[7.6] The divergence from a more general structure indicates that some of ChatGPT's corpora include literary convention, television storytelling, and the original formatting of a fandom's canonical texts.
[7.7] In this example from a queer Supernatural prompt response, ChatGPT produced an author's note, stating: "Note: This fan fiction is based on the TV series 'Supernatural,' created by Eric Kripke. The characters Sam Winchester and Gabriel (aka the archangel Gabriel or the trickster) are from the show. The story expands on their relationship in a queer context while incorporating elements of the Supernatural universe."
[7.8] Authors' notes also appeared in prompt responses for both queer and nonqueer Our Flag Means Death works, serving the additional purpose of indicating that the AI had limited knowledge of the show. The authors' notes that appeared in nonqueer Supernatural prompt responses were presented as follows: "Note: This Supernatural fan fiction is an original story created for entertainment purposes only and does not represent any official content from the Supernatural TV series." This mimics familiar authors' notes on AO3, where creators vocalize their noncommercial intentions to avoid copyright infringement allegations (Herzog 2012).
[7.9] One additional indication that the AI model understood fan fiction conventions but employed them sparingly is the two instances of it including a rating and pairing for a work. These characteristics are commonly employed by authors of works on most platforms. Both generated works that included this formatting choice were from queer Our Flag Means Death works with the same pairing. While it is clear that the AI model has been trained on fan fiction conventions, from our findings study and those of journalists (Eveleth 2023), it is less clear if the training has also impacted the way that the fan fictions generated by the AI were formatted overall. These findings also indicate an avenue for future research that could look into the ethical considerations and fan reaction to the potential use of generative AI in fan fiction spaces, though this is outside the scope of the current work.
8. Implications
[8.1] We offer novel theoretical findings related to the sociotechnical production of queerness, albeit fictive, within generative AI. Few nonqueer fan fictions generated by ChatGPT included any sort of romantic exploration, and none of the nonqueer fan fictions provided any examples of identity disclosure. These distinctions offer two meaningful theoretical outcomes. First, such a reproduction of identity disclosure, even if affirmative, reproduces notions of queer embodiment as other within sociotechnical systems. No nonqueer story requires a character to disclose their cisgender identity or their heterosexuality, as both remain implied. Reproduction of normative embodiments only serves to reveal that technologies, whether driven by community-produced knowledge (Vetter and Pettiway 2017) or complex information, suffer from an unwillingness to confront embedded bias within the very foundations of their design. Second, while creator-produced queer fan fiction indeed emphasizes romantic and sexual encounters, the overt emphasis on such relationships within the queer fan fiction produced by ChatGPT reifies the conflation of queerness with explicit depictions of physical longing, one that remains a site of ongoing contestation within media studies and entertainment pedagogy alike (Annati and Ramsey 2022). Ultimately, while such discrepancies are not exclusionary of queerness per se, they reinscribe problematic and one-sided understandings of queerness as a cultural idea. For example, if one were to take the generated fan fiction produced here as a reflection of LGBTQIA+ identities, one might imagine that it is overly involved in identity disclosure and the expression of romantic feelings. This limited view reemphasizes the importance of centering actual queer experiences in media literacy, especially at the site of its production, which often explicitly excludes queer participants or deploys outdated, essentialist notions of those identities (Vera et al. 2023).
[8.2] Currently, AO3 lacks policies excluding AI-generated fan works from being uploaded to the archive ("AI and Data Scraping on the Archive," posted on AO3, May 13, 2023). Should this change, or should other fandom sites that host fan fiction have policies that prohibit AI-generated works, the demonstrated rigidity of structure in the works generated by sociotechnical systems like ChatGPT has the potential to make it much easier for moderators to identify these works and therefore easier to remove them. The expedited creation speed also identifies AI-generated works, as they can be generated and uploaded instantaneously. The potential for AI-generated works of queer fan fiction to become part of the information ecosystem of entertainment media poses a unique challenge for communities and their responsibility to those intentionally and unintentionally seeking information within fandoms. How ChatGPT mimics and creates its formatting choices also warrants scrutiny. The authors' notes provide a critical context for the reader to situate the intentions of the author and can valorize or demonize certain aspects of the work; reproduction represents a limitation of ChatGPT-created works. For example, contextualizing a fan-supported pairing (e.g., Dean/Sam in Supernatural) that is itself incestuous rather than simply allowing for the presence of incestuous relationships could help the AI avoid pairing siblings within fandoms. Further, this critical role of contextual relationship allowances and pairings helps avoid reifying antiqueer content moderation by treating all relationships and pairings as equals. This lack of distinction, or choice as to where the line of inappropriateness is, needs to be more concretely understood by sociotechnical system creators and those potentially moderating fan fiction sites.
9. Limitations
[9.1] Drawing broader, data-driven inferences from how ChatGPT renders queerness within fan fiction remains impossible. Future studies might generate a broader dataset and further analyze what LGBTQIA+ identities emerge within fan fiction and the frequency at which such identities emerge. Such explorations would further illuminate how ChatGPT understands queer embodiment and what types of queer embodiments feature more or less prominently. The emphasis on queerness as a framing device for content generation led to no discussion regarding other sites of embodied identity. Future research may examine how other intersectional identities surface within ChatGPT-generated fan fiction. While generating a single embodied identity is likely to reproduce the same single-identity explorations found within this study, infusing the prompts with terms such as "diverse characters" or "intersectional themes" might result in unique content. Finally, the focus on three distinctively queer-adjacent fan fictions and media products resulted in a potentially skewed dataset about how queerness was framed and between whom queerness was imagined within each fandom. Extended studies might include historically queer-neutral fandoms to identify how ChatGPT generates queerness from a corpus wherein such examinations of queerness are minimal or nonexistent. Additionally, the fandoms chosen represent popular entertainment media from a distinctly white, Western, and modern media tradition that may affect the generalizability of this study across fandoms that do not fall into this identity. Future work could include a broader cross section of fandoms, particularly fandoms that surround canon material from cultures outside the United States and Western Europe.
10. Conclusion
[10.1] Given its historical ties to addressing cultural inequities, especially around queerness, fan fiction offers a critical space to examine the sociotechnical underpinnings of generative AI in producing fiction and as an extension of naming and making visible embodied identities. Through the exploration of how ChatGPT fabricates queer fan fiction, we identified not only the typologies of visible queerness imagined as possible within generative AI but equally what essentialist and normative ideologies remain rooted within technologies. These insights, through treating generative AI fan fiction as a new extension of media production, call attention to ongoing challenges around depictions of queerness and lay bare the limitations of automated tools to produce in-depth and authentic representations of queerness. Such limitations suggest that while generative AI can aid in the production of fan fiction writing, the possibility of enacting sensitive, diverse, and nonproblematic depictions of queerness remains a wholly human endeavor.