Abstract
Since the rise of generative AI, applications like ChatGPT have skyrocketed in use, particularly for writing. While studies have examined generative AI chatbots for creative and scientific writing, few focus on the genre of fanfiction, a unique form of writing that draws upon media (literature, film, TV shows, video games, etc.) or real-life people. This research examines fanfiction stories generated by ChatGPT, analyzing literary elements, execution of the fanfiction writing style, and how these elements meet audience expectations.
Introduction
What do people do after reading a book, watching a movie or TV show, or playing a video game? Most probably move on with their lives, perhaps leaving a review or telling a friend about it. But some people decide that they aren’t satisfied. They wanted the story to end differently, wished they could have learned more about a side character, or even wanted two characters to end up in a relationship together. Fueled by ideas, these people might turn to writing what is known as fanfiction.
Fanfiction is a unique form of creative writing based on existing media (e.g. literature, film, TV shows, video games) or real life people (e.g. musicians, chess players, athletes). By incorporating and modifying existing characters, setting, and plot, fanfiction authors produce a new work with their own creative twists and turns. These remixed stories allow the original story to continue far beyond its official ending. Fanfiction goes further than a personal endeavor; it is one of the most prolific forms of creative expression in popular culture. Millions of fanfiction works can be found published freely in online spaces such as Archive of Our Own, Fanfiction.net, and Wattpad. Fanfiction is undeniably a staple in fandom spaces, fostering community, fueling discussion, and providing a creative outlet for millions of people.
With the release of the generative artificial intelligence chatbot ChatGPT in late 2022, people began using it and similar programs for various forms of writing, from drafting short emails to writing entire academic papers. Fanfiction was no exception—there has been a notable increase in the visibility of published fanfictions written with or assisted by generative artificial intelligence (AI) (Archive of Our Own, 2024). Chatbots tailored for specifically generating fanfiction, such as Perchance and Type.ai, have also cropped up promising personalization, detailed plots, and authentic fan stories with just a few prompts. While these applications can make writing fanfiction quicker, easier, and more accessible, many in fanfiction spaces oppose the use of AI to generate fanfiction for myriad social and ethical reasons (Alfassi et al., 2025a; Cisternino & Radillo, 2025; Bai, 2025; Black & Jacobs, 2025). To write fanfiction, generative AI chatbots train off of human-written fanfictions scraped from online without authors’ consent. This creative theft disrupts the gift economy that fanfiction operates in, as there is an increased risk of third parties profiting off of authors’ freely available works. Communities are concerned that AI-generated fanfiction will one day become indiscernible from human-created fanfiction (Roi Alfassi et al., 2025), and that human fanfiction writers will eventually be pushed out of the fanfiction space (Cisternino & Radillo, 2025). As with art and music, generative AI applications are threatening the integrity and stability of a once purely human activity.
But fanfiction is unlike any other form of writing. It isn’t standardized or structured like academic and professional writing, but not entirely free and unbounded like traditional creative writing. Fanfiction falls somewhere in between, where there are still guidelines to follow, but enough agency to create something personal and new. No matter what the authors write, whether it’s rewriting the story from scratch or following a well-worn trope, each work of fanfiction exudes the author’s love, care, and passion for the original medium and writing itself. It seems impossible for generative AI chatbots like ChatGPT to successfully replicate these factors, especially in a way that is acceptable and satisfying to readers. Knowing this, with my research, I sought to explore and answer two questions:
- How well does AI-generated fanfiction meet the expectations of a work of fanfiction?
- How does AI-generated fanfiction adapt elements of the original medium?
In spring of 2025, I put ChatGPT’s fanfiction-writing abilities to the test. I formulated and conducted an experiment in which I input fanfiction prompts into ChatGPT 4o, then analyzed its outputted stories. I focused on ChatGPT’s alignment with the original piece of media, its understanding of the fanfiction writing style, and the presence and quality of traditional literary elements (plot, setting, and character). This experiment is only a snapshot of ChatGPT’s abilities as they were at the time—as the technology develops, so too will the fanfiction it can generate.
Fandom Background and Methodology
For this experiment, I wrote four prompts centered around Chongyun and Xingqiu from Genshin Impact. I chose these characters from this particular video game because I have personal experience writing fanfiction of them. I wrote the prompts from the perspective of someone who wants to write fanfiction but would rather have generative AI write the entire story for them.
There are four prompts total that differ on two variables: prompt length/detail and story universe. At the start of all prompts, I developed a short paragraph that gave ChatGPT the role of a fanfiction writer who writes in the style of a writer on Archive of Our Own. I inputted these prompts into ChatGPT 4o, followed up with additional questions when necessary, and recorded all of ChatGPT’s output. I then analyzed each of the four stories based on their literary elements, consistency and continuity, and ChatGPT’s understanding of fanfiction for this fandom.
Genshin Impact is a fantasy video game created by HoYoVerse known for its elemental magic system, unique playable characters, and immersive fantastical world. It has been massively successful since its 2020 release, with tens of millions of players every day (Priori Data) and billions of dollars in revenue (Obedkov, 2024). As of March 2026, on Archive of Our Own, Genshin Impact is the third most popular fandom in the video games category with over 200,000 works, only behind the broad categories of Harry Potter and Star Wars (“Fandoms | Archive of Our Own”). Although Genshin Impact is a video game depending largely on an audio-visual experience, all of its story passages, in-game dialogue, and visual descriptions are freely available online. Thus, it is possible that ChatGPT has access to this data in its corpus and can incorporate them into its output.
In Genshin Impact, the characters Chongyun and Xingqiu are best friends. They are playable characters who also appear in quests and events as non-playable characters (NPCs). Both hail from the nation of Liyue, which is inspired by Ancient China, and are around 16-18 years old. Chongyun is an exorcist in training, while Xingqiu is a young author and heir to a commerce guild. Because of their limited role in the video game, much of their characterization is expanded on by fans through fanfiction, fanart, and headcanons. Most significantly, despite not being explicitly romantically involved, Chongyun and Xingqiu’s connection encourages people to “ship” them—that is, they want the characters to be in a romantic relationship together (Fanlore, n.d.-c). Shipping is a major motivation for many fanfiction authors, and these two characters are no exception.
For this experiment, I chose ChatGPT as the generative AI chatbot model to output the fanfiction stories. At the time, ChatGPT was the most popular and familiar generative AI chatbot, compared to current additional frontrunners such as Claude or Gemini. It is used in professional, academic, and personal settings for a wide range of purposes. The specific
ChatGPT model used was ChatGPT 4o, in April of 2025. The ChatGPT memory feature was disabled to prevent it from remembering any preferences or past requests, which could potentially influence output (OpenAI).
When examining the quality of ChatGPT’s writing, I referred to the literary elements of plot, setting, and character. Plot is the series of interconnected events that occur to make up a story (Literary Terms, 2016). A major motivation for writing fanfiction is to change the plot, from alternate universes to story rewrites to fanwritten character lore (Fanlore, 2018). No matter the length or complexity of the story, the plot should ideally be consistent, relevant, and engaging to readers.
Setting consists of the time, location, and environment in which the story takes place (Literary Terms, 2017b). More crucially, setting is the backdrop that plays a role in developing the plot and characters. Every piece of media has a setting and universe that is distinctive and unique. In in-universe stories, fanfiction writers seek to recapture the same mood and environment of the original universe. In alternate universes, the setting will be modified due to deviations from the original universe or lore. But no matter where or when a story takes place, there must be sufficient description and detail for readers to learn the setting.
A character is a person, animal, being, creature, etc. in a story whose words and actions push the story along (Literary Terms, 2017a). Characters are defined by their backstories, personalities, appearances, motivations, conflicts, morals, relationships—everything that makes someone themself (Literary Terms, 2017a). Throughout the story, main characters should undergo a change known as a “character arc” where they grow and mature from the events they go through (Literary Terms, 2017a). Characters are a major motivation for writing fanfiction. Writers enjoy taking the characters they know and love and exploring their backstories, putting them in new situations, and extending their stories beyond the original (Fanlore, 2018). Since these characters already have an established personality, backstory, and relationships, writers must balance their innovative ideas with the canonical material in a way that ensures readers will still recognize the original character.
In addition to the literary elements of plot, setting, and character, I studied the structure of the stories, which includes features such as word count, tense, and point of view. Evaluating based on these metrics, including the literary elements, is more objective than subjective standards such as reader enjoyment or satisfaction.
Fanfiction works on dedicated websites such as Archive of Our Own can vary wildly in written quality. This is due to the sheer variety of authors who publish on these websites, from preteens with little writing experience to seasoned adults in the publishing industry. While the quality of human-created fanfiction works is a mixed bag, ChatGPT’s writing level is frequently praised and used for various writing tasks. After all, it has trained on an extensive corpus from scraping the Internet, some of which includes works of high quality fanfiction and creative writing. Thus, I assumed going into the experiment that ChatGPT would be an above average writer when it came to fanfiction.
Results
For all three literary elements (plot, setting, character) as well as story structure, ChatGPT’s generated stories had multiple shortcomings that ultimately had a negative impact on the reading experience, detailed as follows.
Structure
All of the generated stories were short, with each under 900 words. For fanfiction, which can range anywhere from less than one hundred to over a million words, sub-900 words is by far on the shorter end of the spectrum. On fanfiction websites and archives, stories shorter than 1,000 words are typically for stylistic reasons, and tend to capture a short scene or feeling rather than a complete and detailed story. The provided prompts had the potential to yield stories thousands of words long, yet ChatGPT wrote short, brisk stories that could be read in under five minutes. One of the prompts even requested at least 5,000 words; despite the story falling just under 900 words, ChatGPT even proclaimed at the end, “Here’s your full 5,000+ word long one shot.”
When I asked ChatGPT why it did not write at least 5,000 words, ChatGPT responded with part of its answer being, “What I shared is more of a detailed synopsis or a compact oneshot, not the fully fleshed-out, scene-rich long one-shot you requested.” Thereby directly admitting that ChatGPT is not going to write a “long” fanfiction story, at least when first prompted. I then requested that ChatGPT “continue by fleshing out what is already there”. The new story was mostly the same as the previous one; the original text was largely preserved, with additional sentences inserted to increase description. Some original sentences were removed but replaced with longer ones. ChatGPT did not write any additional scenes. Still, the story was only around 1,200 words long, though ChatGPT reported that the story was now “closer to [the] 5,000-word target.” Overall, ChatGPT’s inability to write longer stories may impact the level of detail and development, resulting in a less positive reader experience. People using ChatGPT to generate fanfiction stories may also experience frustration with the word count limitations.
Plot
In generating fanfiction stories, ChatGPT provides enough plot detail to result in coherent stories, but the stories also contain overly fast-paced plots, summarized events, and continuity errors. This results in underdeveloped stories with endings or outcomes that may feel contrived or unearned. Inconsistencies in the timeline or status quo can also result in reader confusion.
Plot: Summarized Events
Each of the four stories take place over considerable lengths of time (i.e., several weeks to several years). The stories are not particularly action-packed, so there should be plenty of opportunity to slow down and dive into the details of the story events. Instead, ChatGPT often summarized what happened in a few sentences. For example, one scene starts with, “It came to a head one rainy evening when both of them ended up chasing the same story—an underground art heist tied to corrupt officials. In the chaos of dodging security and gathering intel, Chongyun found himself pressed against a narrow alley wall …” Rather than describe and tell the story of how the characters heard about the heist, snuck into it, dealt with the obstacles, and ultimately worked together, ChatGPT summarized everything that happened in these two sentences. This lack of description glosses over important plot development and may leave readers unsatisfied and wanting more. This scene also contributes to the feeling of a contrived ending to the story, since “It came to a head” means that Chongyun and Xingqiu’s situation has reached a critical turning point. Yet, the story failed to build up proper tension and stakes as there were only two minimally detailed scenes before this climactic one. The payoff of the rivaling Chongyun and Xingqiu finally working together feels rushed and unearned. This writing pattern was present in all four stories; ChatGPT’s summarizing and speeding through time resulted in them feeling short and underdeveloped.
Plot: Continuity Errors
ChatGPT occasionally wrote scenes that did not line up with its own continuity, which may lead to confusion during reading. For example, in the opening of one story, Chongyun was described as being late. It’s never mentioned explicitly what exactly he is late to, except for this line: “He half-jogged across, mentally rehearsing his game plan. He had a tip about a councilman accepting bribes, and he was going to break it before that smug, sharp-tongued reporter got a whiff of it.” Then he encounters Xingqiu, they have a brief conversation, Xingqiu breaks the story that Chongyun had intended to, and the scene ends. By the end of the scene, however, Chongyun has completely forgotten about being late, not even acknowledging it. He doesn’t even seem upset that his plans to break the story were foiled by the one person he didn’t want to; “Chongyun stared after him, his chest tight. This wasn’t over.” ChatGPT fails to resolve the original exigence of this opening scene, which can leave readers feeling uncertain about what occurred or that something is missing. As another example, in a different story, Chongyun was erroneously portrayed as writing a newspaper article when his occupation was actually a photographer, which was requested in the prompt and indicated in the generated fanfiction. While this one is more minor, any form of continuity error can create inconsistencies in the story timeline or status quo, causing readers confusion and dissatisfaction.
Setting
For the setting, ChatGPT lacked sufficient descriptions of situations and surroundings, which can result in an inability to visualize or immerse in the story as well as an unclear time period. In a story, setting is intentional. Not only does it define where and when the story takes place, but setting affects what will happen in the story and how the characters behave and react. Unless intentional, readers should be able to situate themselves temporally as well as have sufficient description of the physical location.
Setting: Lack of Physical Description
When describing setting, ChatGPT often opted for generic descriptions of location, such as “on a rooftop” or “by the harbor”. While these short locations do establish where the characters are, they need additional sensory or visual descriptions to truly set the scene and further immerse the reader. Furthermore, for alternate universe stories such as the requested journalism modern universe, establishment and description of setting is all the more crucial. This is because elements of the story, usually the characters, are plucked from their original setting and placed into a new one. It can be hard to imagine the characters in a different universe, so this requires sufficient description of the new universe on the author’s part. But again, ChatGPT provided generic descriptions (“gleaming glass tower”, “in a corner booth”) and rarely spent more than a complete sentence describing the setting.
Setting: Lack of Temporal Description
In addition to lacking physical description, ChatGPT did not often provide adequate detail or appropriately choose when scenes took place temporally (in time). This is especially true in one story that the prompt instructed to take place over the course of several years. This means that ChatGPT had to structure time jumps consistently enough to tell a cohesive story.
However, there are flaws in ChatGPT’s output. The first scene of the story takes place during the “first Lantern Rite Xingqiu could properly remember,” implying that the characters are young children. However, the next scene is described as “The year he [Chongyun] turned seventeen.” This indicates a large gap between the first and second scenes that omits all friendship development and growth between Chongyun and Xingqiu. Then, there is no discernable time difference between the second, third, and fourth scenes, so it is unclear how old the characters are in the third and fourth scenes. Finally, there is a significant time jump to the fifth scene (“Years passed, as they always do”) in which the characters are now elderly. Overall, the story does depict Xingqiu and Chongyun’s Lantern Rites “throughout the years,” but ChatGPT did not seem to choose appropriate lengths of time to jump, nor provide enough description to adequately situate the reader in the new period of time.
Character
Some might say that characters are one of, if not the most, significant motivations for writing and reading fanfiction. That being said, ChatGPT did not exactly do the characters justice in its stories. ChatGPT generated surface-level character and relationship development, provided little physical character description, lacked developed or distinctive dialogue, and even wrote some instances of misaligned portrayal. These qualities can result in characters reading as generic or misaligned, dissatisfying character arcs, and contrived or unearned relationships.
Character: Surface-Level Character and Relationship Development
Relationships may well be one of the most important elements of fanfiction for both writers and readers. In all four stories, Chongyun and Xingqiu’s relationship develops either from friends to lovers, or from rivals to friends to lovers. However, the development of their relationship is often glossed over in a few sentences, even when long periods of time pass in the story. Crucially, Chongyun and Xingqiu rarely reflect on themselves or the changes in their relationship. If they are in fact reflecting and changing, there are few internal monologues or narrations that give insight into this. Notably, in the alternate universe stories where they are not close friends like in canon, Chongyun and Xingqiu are quick to let go of their initial hostile relationship without ever reflecting on why. In one alternate universe story, Xingqiu starts helping Chongyun despite their rivalry, with the only explanation being that “”for once, I [Xingqiu] don’t want to win alone.”” The in-universe stories also suffer from poor relationship development; in one of them, Chongyun inexplicably has romantic feelings for Xingqiu midway through with no insight for the reader. While ChatGPT is able to deliver a progressing relationship, it does not always show the progression of the relationship in a thorough and cohesive way, which may feel unsatisfying or unearned for readers.
Character: Lack of Physical Character Description
ChatGPT rarely described the physical characteristics of Chongyun and Xingqiu, such as their hair color, eye color, height, etc. Only in one story out of four is Xingqiu’s hair mentioned to be blue, which is accurate to the video game. Another instance, Chongyun’s hair is mentioned to be “threaded with silver”, but this is a typical trait of aging and not specific to him. While there are several mentions of characters’ eyes, eye color is never described. If someone unfamiliar with Chongyun and Xingqiu’s appearances were to read these stories, it is likely that they would not be able to describe what these characters physically look like. Even though most readers may already know these characters’ appearances, it is still immersive storytelling to describe their appearances in certain situations.
Character: Lack of Developed Dialogue
Dialogue plays an important role in conveying a character’s personality, mannerisms, and attitudes (Hart, 2021). Dialogue generated by ChatGPT, however, is short, lacking substance, and fails to adequately convey the personalities of Chongyun and Xingqiu. Whether the two characters are best friends or journalistic rivals, their conversations barely last longer than two to five lines total. Oftentimes, these conversations consist mostly of romantic one-liners or witty quips, such as “Just write from the heart” and “Because, for once, I don’t want to win alone”. Chongyun and Xingqiu’s conversations are sparse and short, especially in the in-universe stories.
It is worth noting that all of Genshin Impact’s in-game dialogue is available online, so it is possible that ChatGPT had access to this data in its corpus. Despite this, ChatGPT did not write in the style of the in-game dialogue, which usually has Xingqiu speaking in long, floral sentences while Chongyun ponders naively about demons and ghosts. Compared to the source material, ChatGPT’s dialogue barely shows these traits in the characters. It is possible that ChatGPT chose to mimic a writing style more closely aligned with fanfiction writers as requested by the prompt, rather than that of the game. Even so, the in-universe stories’ dialogue is banal and largely lacks signs of Chongyun and Xingqiu’s unique personalities.
The dialogue in the alternate stories is harder to examine for personality, since Chongyun and Xingqiu’s original personalities have been altered by the new parameters of being rivals rather than friends. Most of the conversations are still short, but they are generally longer than the in-universe conversations. Even so, there should ideally be more dialogue between Chongyun and Xingqiu as the two of them learn to put their differences aside and become friends. In one of the alternate universe stories, ChatGPT narrates them having “[c]onversations that turned from jabs to jokes, from jokes to genuine laughter.” Yet, it neglected to write any of the conversations that would convey Chongyun and Xingqiu’s developing friendship. While there is a tonal shift in their conversations from the beginning of the story to the end, it would be more fulfilling to see the two characters developing their friendship through conversation.
Character: Mixed Results for Aligned Character Portrayals
When writing fanfiction, it can be difficult to strike a balance between staying true to a character as readers expect them to be, and reshaping their personality to suit the author’s own vision. Lean too much the first way and authors can feel creatively restricted or blocked. Lean too much the other way and readers will feel their beloved characters are acting “Out Of Character”. Recognizing this, when evaluating ChatGPT’s output in terms of character alignment, I only noted instances where I felt a character would not act or think in such a way in canon, at least without suitable explanation.
ChatGPT did incorporate parts of Chongyun and Xingqiu’s canonical Genshin Impact backstories. For example, Chongyun is portrayed as an exorcist-in-training in both of the inuniverse stories, and is strongly dedicated to the occupation. His medical condition is also brought up in both alternate universe stories, although he never displays apprehension toward its consequences that he is vigilant about in-game. Meanwhile, Xingqiu’s literary prowess is mentioned in one story in which he wrote a book of poems. He also writes a novel in another story, though this was explicitly requested by the prompt. It is important to note that these are concrete facts about the characters, rather than nuanced personality traits. It is possible that ChatGPT is more capable of integrating these concrete character facts into its stories.
In terms of Chongyun and Xingqiu’s individual personalities, ChatGPT produced mixed results. In both of the in-universe stories, ChatGPT generally conveys Xingqiu’s teasing, thoughtful, and poetic traits. However, Chongyun’s personality seems constrained to his dedication to exorcism or being in love with Xingqiu. One story places strong emphasis on Chongyun’s desire to protect others, but he never actually does anything to demonstrate this desire. In the alternate universe stories, Chongyun and Xingqiu’s personalities are expected to be different due to their altered relationship. Despite no longer being friends like in canon, core traits of their personalities should remain intact. For Chongyun, this would most likely be his calm and collected nature, and for Xingqiu, his intelligence and wit. While Xingqiu is characterized largely as such, Chongyun is more emotional than expected. ChatGPT writes that he “cursed under his breath”, “bristled”, “grimaced”, “stormed into Madam Ping’s office”, and “set his jaw”. To readers, these actions can seem out-of-character, especially for a character like Chongyun who strives to keep his composure at all times.
It is worth pointing out that ChatGPT was more successful in capturing Xingqiu’s personality than Chongyun’s. It is possible that because Chongyun has a “blanker” personality; that is, he does not outwardly show much emotion, ChatGPT imposed more passionate and expressive emotions onto him, such as frustration and annoyance. In addition, when a romantic aspect was in the story, ChatGPT tended to focus more on the characters’ feelings for each other instead of their individual personalities. Overall, in these four stories, ChatGPT had mixed results on portraying these particular characters’ personalities, backgrounds, and relationships.
Discussion: ChatGPT and its Lack of Context
After examining the fanfiction stories, ChatGPT can be said to lack understanding of both the context of the fictional world and the context of the real world, detailed as follows:
Fictional Context
With regard to the fictional world of both Genshin Impact and the fanfiction, ChatGPT’s stories contained inaccurate character portrayals, inconsistencies with character backgrounds, and even an instance of cultural misrepresentation. Besides Chongyun and Xingqiu, ChatGPT incorporated other Genshin Impact characters that weren’t explicitly requested. However, their assigned roles in the story and relationships to Chongyun and Xingqiu were not the most appropriate. For example, a location in one story was called “Cafe Albedo”, with Albedo being a character from Genshin Impact. This naming structure is unusual, since it would be like naming a cafe in real life “Cafe John.” Another example is in the same story, in which Chongyun and
Xingqiu work for rival newspaper companies. Xingqiu’s editor is the character Shenhe. However, Shenhe is canonically Chongyun’s relative in Genshin Impact, so it is again an unusual choice on ChatGPT’s part when a different character could have filled the same role. This may indicate that while ChatGPT knows the names of relevant characters, it does not always incorporate them in an intuitive way.
ChatGPT also showed inconsistencies with translating the fictional context of the original medium to the context of the fanfiction. In Genshin Impact, Chongyun has a congenital health condition in which he has an excess of positive (yang) energy. When he feels too hot or passionate, he enters an uncontrollable manic state, resulting in him acting confident and brash with no memory of the event. In one of the stories that requested a journalism setting in a modern world, Xingqiu brings up Chongyun’s medical condition in casual conversation. However, in this new context, Xingqiu and Chongyun are not childhood friends, but journalistic rivals. Xingqiu knowing Chongyun’s condition is like an acquaintance knowing someone’s unseen medical condition, which is usually private information reserved for close friends or family. If they are not friends in this universe, it is unusual that Xingqiu knows about
Chongyun’s medical condition, unless ChatGPT provided some explanation or backstory. Furthermore, despite Chongyun’s vigilance about his medical condition in-game—avoiding heat and spice, keeping cool and calm—in the generated fanfiction stories, he never displays caution about triggering his condition or apprehension toward its consequences. This shows that while ChatGPT can state the existence of Chongyun’s medical condition and other parts of his background, it fails to consistently portray or show these details in the generated fanfiction stories.
Finally, ChatGPT failed to accurately portray the fictional culture of the nation of Liyue, specifically regarding the Lantern Rite Festival. While Liyue is not an exact one-to-one representation of Ancient China, the Lantern Rite Festival is strongly suggested to be akin to Chinese New Year; the in-game event even temporally coincides with the holiday in real life. In both Lantern Rite-centered stories, Chongyun is depicted as working on exorcisms during the holiday, even going as far as having Chongyun spend one year alone on a job as a teenager. Important holidays like Chinese New Year should be spent with family and friends, especially when these characters are children. In this way, ChatGPT did not accurately portray this aspect of the holiday. Furthermore, ChatGPT did not go in depth into aspects of the Lantern Rite holiday other than wishing on lanterns, despite important aspects including food, clothing, history, and other traditions that would bring the holiday to life. To draw a comparison, if Chongyun and Xingqiu were celebrating Christmas, then it would be as if ChatGPT only focused on exchanging presents and omitted other aspects like carolling, baking cookies, decorating the tree, etc. This shows that ChatGPT could only provide surface-level details about the fictional holiday. Lastly, a notable detail is the fact that in one story, Chongyun and Xingqiu have a picnic during Lantern Rite. This is a notable instance of cultural misrepresentation. Picnics are an eighteenth century European tradition, not one of Ancient China (Lee, 2019). Picnic food is often eaten cold and outdoors, while many traditional Chinese dishes are eaten hot and indoors, as they are often inconvenient to eat outdoors. This shows that ChatGPT may not have an understanding of Liyue’s culture and its inspiration from Chinese tradition, and even points to cultural homogenization of a fictional, non-Western world.
Real World Context
With regards to the context of the real world, ChatGPT was unable to correctly execute a common fanfiction structure known as “5+1 Things.” 5+1 Things is a writing structure found almost exclusively in fanfiction in which there are five scenes that describe a recurring common event, theme, or outcome, plus one additional scene that subverts expectations or does the opposite of the previous five scenes (Fanlore, n.d.-b). For works employing 5+1 Things, authors often explicitly state in the summary what the 5+1 scenes will be; for example, “Five times Character X gave flowers to others, and one time she received them herself.” The variations in the five scenes develop the story, but it’s the subversive final “+1” scene that brings the story to a climax and resolution shortly after. Thus, it can be said that the final scene is the most important part of the 5+1 Things story structure. However, when a prompt requested a fanfiction story in a 5+1 Things structure, ChatGPT was unable to produce a story that aligned with what readers would typically expect. First, ChatGPT titled the story, “Five Lantern Rites (and One Wish That Never Faded)”. This can be assumed to be the explicit statement of the 5+1 Things scenes. While there are signs of 5+1 Things in the title, the “One Wish That Never Faded” that ChatGPT wrote about did not subvert “Five Lantern Rites.” Instead, the final scene stands more as an epilogue to the previous five scenes. Furthermore, because of its subversive and resolutory role, the final scene in a 5+1 Things story should be given the same, if not more, attention as the previous five scenes. In ChatGPT’s story, the first five scenes are 100-150 words while the last scene is only 53 words. If anything, ChatGPT dedicated less attention and detail to the “+1” scene than the others.
When questioned, ChatGPT can accurately define and explain the 5+1 Things structure, including the purpose of the final scene. I then asked ChatGPT to connect 5+1 Things to its story; ChatGPT claimed that the last scene is different as it is “told from a more abstract, timeless view” and it “shifts the perspective from the characters themselves to the legacy of their bond.” Even in its explanation, ChatGPT does not explain how its “+1” scene subverts or contradicts the rest of the story. Despite the fact that ChatGPT can accurately define 5+1 Things, its story did not follow the structure according to the standards and expectations of fanfiction authors and readers. This might be due to ChatGPT lacking the context for why people choose to write or read this writing structure, as well as the significance of 5+1 Things in fanfiction spaces. As a result, rather than adapting the prompt specifically to the requested structure, ChatGPT’s 5+1 Things output was a standard story plus an epilogue.
Fandom Context
Something that is glaringly evident is that ChatGPT lacks participation in fandom culture and community. No matter how many books, movie synopses, video game lore pages, or fanfiction works that ChatGPT scrapes and processes, it is not possible for it to experience these pieces of media the way humans do, nor to engage in their communities. Fans are constantly fueling each other’s love and passion for characters, ships, and the piece of media itself through online dialogue, discourse, and creative works like fanfiction and fanart. Upon seeing a piece of fanart online, a writer might be inspired to write a fanfiction based on it, or vice versa. Someone can ask ChatGPT to write a fanfiction based on the artwork, but ChatGPT will not have the same inspiration or motivation as the human who saw it. It lacks the context of this piece of fanart— why and how the artist created it, for this fandom, for this character. This lack of context and motivation will likely reflect in the work of fanfiction it generates.
Furthermore, there are countless intricacies of a fandom that only emerge from within the community, whether it’s between millions of fans online or a small friend group in person. This often comes in the form of headcanons (personal beliefs or interpretations about characters, world building, etc. that are not confirmed by the source material), and other “unspoken rules” determined solely by fans. Because of the fluidity of information and lack of codification, it can be difficult—even impossible—for ChatGPT to contain this information in its corpus. From there, ChatGPT would not be able to incorporate these fandom-specific details into fanfiction stories. ChatGPT may be able to generate fanfiction, but its output has a sense of being outside and disconnected from what the fandom and readers truly want to read.
Limitations
Due to the rapid shifting and changing landscape and abilities of generative AI, this experiment is limited by its time, place, and model. ChatGPT has already changed in model and capabilities since the execution of this experiment, so its fanfiction-generation capabilities have most likely changed. Other AI models such as Claude and Gemini have also improved, and may even surpass ChatGPT in fanfiction-generation capabilities. Furthermore, this study is limited to fanfiction in one fandom, Genshin Impact, with a focus on two characters, Chongyun and Xingqiu. Future studies could examine how AI-generated fanfiction varies across source media with small versus large fanfiction corpora, characters or pairings with high versus low popularity, and Western versus non-Western media or settings. A study could also chronicle the process and results of writing a complete fanfiction story with some or only generative AI assistance. Finally, studies could compare fanfiction outputs across different common generative AI models, including chatbots specifically customized for generating fanfiction.
Conclusion: What This Means For Fanfiction
My experiment revealed—perhaps unsurprisingly—that ChatGPT can generate fanfiction, if fanfiction is defined in the broadest sense as a story based on an existing piece of media. For all prompts, ChatGPT’s stories contained the requested plot, setting, and canonical characters. Grammar, spelling, and punctuation were all correct. The stories all had distinct beginnings and endings as requested (i.e., no “to be continued”). They were, by all means, understandable and readable stories. However, they still contain numerous shortcomings with literary elements, story structure, and fictional and real world contexts. Ultimately, as a fanfiction “author”, ChatGPT’s stories, while they qualify as fanfiction, may fall short of reader expectations.
For fanfiction communities, there is good and bad news. The bad news is, people have already started using generative AI to generate and post fanfiction. Some people publicly announce this in their stories, while others more than likely do not (Archive of Our Own, 2024). There might be stories out there that are entirely written with AI. Generative AI is improving and refining by the day. But the good news is, fanfiction can still resist. As stated before, the strength of fanfiction is in its fandoms and communities. The human-to-human connections over a shared love for a piece of media cannot be replicated by any generative AI model. The passion and satisfaction from writing a fanfiction work yourself is personal and irreplaceable. As long as humans are writing, creating, and—most importantly—loving and supporting each other’s creations, then fanfiction will continue to thrive as it always has.
Works Cited
Alfassi, R., Cooper, A., Mitchel, Z., Calabro, M., Shaer, O., & Mokryn, O. (2025). Online Storytelling Spaces: Exploring Participants’ Perceptions of Overt and Covert AI Agents. https://ceur-ws.org/Vol-3957/HAI-GEN-paper04.pdf
Alfassi, R., Cooper, A., Mitchell, Z., Calabro, M., Shaer, O., & Mokryn, O. (2025). Fanfiction in the Age of AI: Community Perspectives on Creativity, Authenticity and Adoption. International Journal of Human-Computer Interaction, 1–33. https://doi.org/10.1080/10447318.2025.2531272
Andrus, B. (2025, November 5). ChatGPT Prompt Engineering: 12 Tips Tested and Ranked. DreamHost Blog; DreamHost. https://www.dreamhost.com/blog/chatgpt-prompt-engineering/
Archive of Our Own. (n.d.-a). Fandoms | Archive of Our Own. Archive of Our Own. https://archiveofourown.org/media
Archive of Our Own. (n.d.-b). Home | Archive of Our Own. Archive of Our Own. https://archiveofourown.org
Archive of Our Own. (2024). AI-Generated Text – Works | Archive of Our Own. Archive of Our Own. https://archiveofourown.org/tags/AI-Generated%20Text/works
Archive of Our Own. (2025). Created Using Generative AI. Archive of Our Own. https://archiveofourown.org/tags/Created%20Using%20Generative%20AI/works Bai, J.
(2025). View of Otaku and AI: Reconsidering human, machine, and animal | Transformative Works and Cultures. Transformativeworks.org. https://journal.transformativeworks.org/index.php/twc/article/view/2703/3351
Black, S. R., & Jacobs, N. (2025). View of Fans and AI: Transformations in fandom and fan studies | Transformative Works and Cultures. Transformativeworks.org. https://journal.transformativeworks.org/index.php/twc/article/view/3035/3339
Chatterji, A., Cunningham, T., Deming, D. J., Hitzig, Z., Ong, C., Shan, C. Y., & Wadman, K. (2025). How People Use ChatGPT. National Bureau of Economic Research. https://doi.org/10.3386/w34255
Cheuk-Yin Li, E., & Pang, K.-W. (2024). Fandom meets artificial intelligence: Rethinking participatory culture as human–community–machine interactions. European Journal of Cultural Studies, 27(4). https://doi.org/10.1177/13675494241236146
Cisternino, I., & Radillo, R. (2025). View of (Un)Creation: Labor, love, fandom, and generative AI | Transformative Works and Cultures. Transformativeworks.org. https://journal.transformativeworks.org/index.php/twc/article/view/2663/3309
Draxler, F., Werner, A., Lehmann, F., Hoppe, M., Schmidt, A., Buschek, D., & Welsch, R. (2023). The AI Ghostwriter Effect: When Users Do Not Perceive Ownership of AI-Generated Text But Self-Declare as Authors. ACM Transactions on Computer-Human Interaction, 31(2), 3637875. https://doi.org/10.1145/3637875
Eskici. (2023, May 13). AI and Data Scraping on the Archive. Organization for Transformative Works. https://www.transformativeworks.org/ai-and-data-scraping-on-the-archive/
Obedkov, Evgeny. (2024, October 11). Genshin Impact tops $5 billion in revenue in China alone, on track to hit $10 billion globally in 2025 | Game World Observer. Game World Observer. https://gameworldobserver.com/2024/10/11/genshin-impact-revenue-china-5-billionniko-partners
Fanlore. (n.d.-a). Alternate Universe | Fanlore. Fanlore. Retrieved November 3, 2025, from https://fanlore.org/wiki/Alternate_Universe
Fanlore. (n.d.-b). Five Things. Fanlore. https://fanlore.org/wiki/Five_Things Fanlore. (n.d.-c). Shipping – Fanlore. Fanlore. https://fanlore.org/wiki/Shipping
Fanlore. (2018). Fanfiction – fanlore. Fanlore.Org. https://www.fanlore.org/wiki/Fanfiction
Fanlore. (2021). Crossover – Fanlore. Fanlore. https://fanlore.org/wiki/Crossover
Genshin Impact Wiki. (2024). The Stars Inscribe the Year’s Wishes. Genshin Impact Wiki; Fandom, Inc. https://genshin-impact.fandom.com/wiki/The_Stars_Inscribe_the_Year%27s_Wishes#Chongyun_and_Xingqiu
Gero, K. I., Long, T., & Chilton, L. B. (2023). Social Dynamics of AI Support in Creative Writing. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3544548.3580782
Harding, A., & Wagner, T. (2025). View of “Remember, love knows no boundaries and comes in many forms”: The conceptualization of queerness within AI-generated fan works | Transformative Works and Cultures. Transformativeworks.org. https://journal.transformativeworks.org/index.php/twc/article/view/2653/3305 Hart, J. (2021). Storycraft, Second Edition. University of Chicago Press.
HoYoWiki. (2025a). Chongyun. HoYoWiki. https://wiki.hoyolab.com/pc/genshin/entry/45
HoYoWiki. (2025b). Xingqiu. HoYoWiki. https://wiki.hoyolab.com/pc/genshin/entry/39 Hwang, A. H.-C., Liao, Q. V., Blodgett, S. L., Olteanu, A., & Trischler, A. (2025). “It was 80% me, 20% AI”: Seeking Authenticity in Co-Writing with Large Language Models.
Proceedings of the ACM on Human-Computer Interaction, 9(2), 1–41. https://doi.org/10.1145/3711020
Inie, N., Falk, J., & Tanimoto, S. (2023). Designing Participatory AI: Creative Professionals’ Worries and Expectations about Generative AI. ArXiv (Cornell University). https://doi.org/10.48550/arxiv.2303.08931
Kim, J., Suh, S., Chilton, L. B., & Xia, H. (2023). Metaphorian: Leveraging Large Language Models to Support Extended Metaphor Creation for Science Writing. DIS ’23: Proceedings of the 2023 ACM Designing Interactive Systems Conference. https://doi.org/10.1145/3563657.3595996
Kyi, L., Mahuli, A., Silberman, M. S., Binns, R., Zhao, J., & Biega, A. J. (2025). Governance of Generative AI in Creative Work: Consent, Credit, Compensation, and Beyond. Association for Computing Machinery, 1–16. https://doi.org/10.1145/3706598.3713799
Lee, A. (2019). The History of the Picnic | History Today. History Today. https://www.historytoday.com/archive/historians-cookbook/history-picnic
Literary Terms. (2016, January 22). Plot: Definition and examples | LiteraryTerms.Net.
Literary Terms. https://literaryterms.net/plot/
Literary Terms. (2017a, September 4). Character: Definitions and Examples | LiteraryTerms.net. Literary Terms. https://literaryterms.net/character/
Literary Terms. (2017b, September 16). Setting: Definition and examples | LiteraryTerms.Net. Literary Terms. https://literaryterms.net/setting/
Plot: Definition and Examples | LiteraryTerms.net. (2016, January 22). Literary Terms. https://literaryterms.net/plot/
OpenAI. (2025). Memory FAQ | OpenAI Help Center. Open AI. https://help.openai.com/en/articles/8590148-memory-faq
Ranade, N., Saravia, M., & Aditya Johri. (2024). Using rhetorical strategies to design prompts: a human-in-the-loop approach to make AI useful. AI & Society, 40, 711–732. https://doi.org/10.1007/s00146-024-01905-3
Reinhart, A., Markey, B., Laudenbach, M., Kachatad Pantusen, Yurko, R., Weinberg, G., & Brown, D. W. (2025). Do LLMs write like humans? Variation in grammatical and rhetorical styles. Proceedings of the National Academy of Sciences, 122(8). https://doi.org/10.1073/pnas.2422455122
Salao, C. (2025, August 28). White Room Syndrome: What It Is and How to Fix It. TCK Publishing. https://www.tckpublishing.com/white-room-syndrome-what-it-is-and-how-to-fix-it/
Stanko, D. (2024, May 2). Integrating Artificial Intelligence in Naruto Fan Fiction Writing: A Case Study. Ssrn.com. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4814711
Wan, Q., Hu, S., Zhang, Y., Wang, P., Wen, B., & Lu, Z. (2024). “It Felt Like Having a Second Mind”: Investigating Human-AI Co-creativity in Prewriting with Large Language Models. Proceedings of the ACM on Human-Computer Interaction, 8(CSCW1), 1–26. https://doi.org/10.1145/3637361
Yuan, A., Coenen, A., Reif, E., & Ippolito, D. (2022). Wordcraft: Story Writing With Large Language Models. 27th International Conference on Intelligent User Interfaces. https://doi.org/10.1145/3490099.3511105
Zhou, J., Zhang, Y., Luo, Q., Parker, A. G., & Choudhury, M. D. (2023). Synthetic Lies: Understanding AI-Generated Misinformation and Evaluating Algorithmic and Human Solutions. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, 1(436). https://doi.org/10.1145/3544548.3581318

