Spotting AI-Written Fanfiction
Preface: Scope, Sources, and Why This Guide Exists
About the examples used in this guide: The quoted lines and excerpts referenced throughout were taken from published fanworks that were assumed to be 100% AI‑generated at the time they were discussed or archived. However, many of these notes were compiled from an older running list of AI tells or analyses of individual fics, and in several cases I no longer have full original context like chapter placement or surrounding prose for individual lines. The examples are meant to illustrate patterns, not to function as isolated proof.
How this guide was compiled: I’ve been noticing and tracking signs of AI‑written fanfiction for roughly ten months. Over that period, Archive of Our Own tags across countless fandoms have see a noticeable influx of works from different authors that shared the same tone, prose rhythms, sentence structures, metaphors, and stylistic quirks... often to an uncanny degree. These similarities went beyond common fandom tropes, author demographics/linguistic background, or genre conventions. What first stood out was not that the writing was “bad” (by all means, it is grammatically correct and all), but that it was recognizably similar in ways that human writing rarely is across unrelated contexts. As these works accumulated, the repetition became very difficult to dismiss as coincidence. The same rhetorical structures, the same emotional beats, the same emphasis patterns, and even the same phrasing began appearing across unrelated fics, fandoms, and accounts, with an absence of divergence. From there, I started actively cataloguing recurring traits and comparing them against one another with methods of inductive reasoning, which is what ultimately became this guide.
I also want to note that my motivation here is not to claim personal superiority or “expertise” in detecting AI. This guide comes from a long-standing personal interest in languages, linguistics, and writing itself. I’m a linguistics student with a focus on syntax, semantics, and pragmatics, and I have been writing and publishing fiction and fanfiction since childhood across various platforms. The intent is to share observations and (admittedly) novice-intermediate level analytical frameworks.
On AI-detection websites: Automated AI-checking tools are not reliable indicators for CREATIVE WRITING. Most are designed for professional, technical, or academic prose and primarily flag vocabulary density, sentence complexity, or so-called “smart” phrasing. Fiction, especially stylized, emotional, or experimental prose, falls outside what these tools can accurately evaluate. This guide relies instead on your own pattern recognition across narrative, syntax, vocabulary, and rhetoric, not numerical scores or percentages.
On reader perception and accessibility: The ability to notice the tells described here varies widely. Some readers pick up on them immediately; others may never consciously register them at all. Neither response is a failure. Additionally, readers whose native language is not English may reasonably interpret certain structures, repetitions, or metaphors differently, or may be less sensitive to subtle awkwardness in English prose rhythms. This guide assumes a relatively high familiarity with English stylistics (in creative writing) and should be read with that limitation in mind.
More broadly, this work is part of a resistance to growing anti-intellectualism, creative deskilling, and late-stage capitalist pressures that frame generative AI as positive, neutral, inevitable, or harmless, when it is far from any of those. The increased reliance on AI to produce language for us has real cognitive costs [A study from MIT showing lower metacognitive engagement, lower brain connectivity, shallow encoding (creation of memories), and pathways to "echo chamber" thinking, associated with using LLMs (specifically ChatGPT in this study) https://arxiv.org/pdf/2506.08872v1]. Additionally, the environmental impact of AI data centers (chiefly, the pollution of communities where humans live and the draining of our freshwater resources) and the technological strain required to maintain and run large language models contribute to broader societal and ecological concerns [https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/] [https://escholarship.org/uc/item/32d6m0d1]. The AI industry also often depends on exploitative and abusive labor practices. When outsourced thinking becomes normalized, our relationship to reading, writing, meaning, and creating itself degrades. Naming these patterns is one small way of pushing back against this.
What this guide is (and is not): This guide is an attempt to articulate why certain works feel mechanically similar or hollow/uncanny to many reader, and to put words to patterns that are otherwise dismissed as a vague “AI vibe”. It's NOT a tool for accusation or a diagnostic checklist. Importantly, it's also not meant to be applied at the sentence level!! Isolated lines, screenshots, or excerpts stripped of context are not sufficient evidence of anything. The goal is transparency and media literacy, definitely not harassment or purity testing. It is best used holistically.
⠀
Important note: One or two of these signs on their own do not mean a fic is AI!!!!!!!!!!!!!!!! Many are common stylistic choices that your average Joe can use normally. It becomes suspicious when multiple patterns appear together, repeatedly, and densely within the same work.
A disclaimer about the word “excessive” in this: Many of the traits below are normal in moderation (em dashes, italics, similes, etc.). This guide focuses on when they cross into noticeable patterned overuse to the point that they feel automatically spit out rather than intentionally chosen.
⠀
1. Metadata + Posting Behavior Red Flags
These don’t prove anything on their own and are weak and circumstantial, but they can strengthen suspicions when paired with prose-level tells and corroborate.
- Suspicious backdating or upload patterns
- Extremely rapid updates accumulating unusually high word counts
(Please note that excuses of conditions like ADHD leading to one hyperfocusing on writing does not apply when all following tells are present, either.) - Chapters that feel mass-produced rather than drafted over time, a narrative that is pushed and pushed and pushed just for the sake of continuous content and validation
- CONSTRAINTS:
Some humans genuinely prewrite large chunks. Some fandoms reward fast serial posting and encourage batch uploads.
2. Structural + Macro-level Issues
- Generated "scene chunks"
This is one of THE strongest tells. If: chapters are split into multiple paragraph clusters that each feel like a self-contained mini-scene; each chunk has its own emotional arc or conclusion (despite occurring within the same in-universe moment); there are scene breaks, dividers but a lack of smooth transitions between these chunks; or it reads like stitched-together outputs rather than continuous narrative flow—that's a big sign. - Illusion of progress without movement
Scenes that feel like they advance emotionally but don’t actually change anything within the narrative, and repeated internal realizations that don’t actually affect character behavior
3. Sentence-Level Patterning (Very Common)
- Negation structures ("Not X, but/just Y")
Often used constantly to sound insightful while adding little meaning.
Examples:
“And they both liked music—not just listening, but making it.”
“Not loud—just enough to fill the corners.”
“He wasn’t calm—he just worked really hard to convince himself he was.”
This structure becomes suspicious when it appears in nearly every paragraph, the negation doesn’t correct a real assumption the reader had, and it creates a feeling of depth without actually adding it. Instead of making a direct claim, it spends time dismantling strawman assumptions that weren't there in the first place, and negates generic imagined premises. - Rhetorical echoing (“And now? Now…”)
Pattern: A statement followed by a short rhetorical question or fragment that repeats it
Example: “None of it had made a difference. And now? Now he was teetering on the edge of failure."
"But right now? Right now, he was dangerously close to losing it."
CONSTRAINTS: Humans absolutely use this. It's particularly common in YA and dramatic genres. The key tell is predictability and repetition across unrelated works. - Negative statement followed by clarification
Pattern: “He doesn’t do X. No, he does Y instead.” "He doesn't do X, just Y."
Example: “He doesn’t kiss him right away. No, he starts slow, deliberate…”
Used sparingly, this can be effective, but when repeated constantly, it becomes formulaic and extremely easily spotted. - Rule of threes
Frequent strings of groups of three adjectives or verbs. Especially noticeable when paired with other AI patterns. Humans do this too, but LLMs tend to do it much more frequently because triads are culturally overrepresented as “good writing.” On its own, weak, but when stacked with other tells, can corroborate. So this will be counted as a tell that is best treated as supporting evidence ONLY.
4. Punctuation + Formatting Overuse
Both of these are also best treated as supporting evidence only!
- Excessive Em Dashes
Em dashes are normal and can be used to replace parentheses- or comma-encircled clauses, but AI tends to use them multiple times per paragraph, stack emphasis inside them, and replace commas, periods, or sentence breaks entirely.
CONSTRAINTS: Some authors genuinely love em dashes (ME). Stylistic mimicry can obscure this. The key tell is density and stacking. - Emphasis Within Em Dashes
Stacked emphasis often looks like repetition inside dashes and intensifiers layered SUPER redundantly.
Example:
"The question comes out more horrified than he intended, and [character]—[character], the unshakable, unreadable enigma of a man—lets out an incredulous laugh."
“Because for half a second—just a half second—his eyes flashed crimson.”
“So seeing him fidgeting—fidgeting—is practically a flashing neon sign…”
5. Vocabulary + Word Choice Patterns
- The “un‑” adjective cluster!!!!!!!
AKA: "unspoken", "unnameable", "unreadable", "undeniable", "unnamed", and other forms of these that may come up.
Often used to describe characters who are NOT the POV character and their expressions, reactions, or mannerisms, in an attempt to demonstrate what could possibly be third-person limited perspective, but comes across as formulaic and lazy.
Example:
"Every conversation, no matter how mundane, left something unspoken hanging in the air. They'd exchange small things-how annoying the heat was, how obnoxious their fellow interns were. Each time, it felt like they were slowly peeling back layers." - "Exhale/exhales/exhaled"
No idea. Just seen this one a lot. - “The kind of _ that...” “the type of _ that...”
Also no idea why. Need to look into this further. - pattern/group of two adjectives mid-sentence or at end of sentence... still working on this... need to update... like "'Don't start with that,' he murmurs, slow, controlled."
- Repetitive descriptor recycling
A trait is established once, then restated word-for-word throughout the fic. It feels like the model is reintroducing information it doesn’t remember already establishing, which makes sense when taking into account that LLMs have limited memory [Two papers explaining this and discussing this topic: https://arxiv.org/html/2506.08184v2 https://arxiv.org/html/2504.15965v2] [This following article was written in the context of using LLMs to write code, but still explains the limited memory thing: https://blog.bytebytego.com/p/the-memory-problem-why-llms-sometimes]
Example:
A character repeatedly described as “untouchable,” “effortless,” or having “dark, unreadable eyes” in nearly every scene, even though it was established very early on.
Again, salient traits are re-introduced because the LLM treats them as probabilistically relevant each time, and there’s no internal sense from the author of “I’ve already done this for stylistic reasons”.
6. Simile Overload
One of the most noticeable AI tells when overused.
- Density
Excessive similes, especially using "like", in a short fic (e.g., 5k words). Once you notice it, it becomes impossible to unsee. - Illogical or vague comparisons
Examples:
“Heat continued to suffocate the dorms like wet cloth."
"...so when [Character] started pulling away, it hit like a bruise." (nonsense comparison. there's no element of surprise related to bruises, bruises don't do anything, much less "hit", they're a damaged portion of skin....)
“Disappearing from the air like it was trying to erase itself."
"But it felt like something. Like a slow orbit drawing tighter." (????)
"The apartment fell into a strange, hollow rhythm, like it was waiting, holding its breath." (if you're saying it's waiting or holding it's breath why would you say that it also fell into a rhythm? talking about a state of movement and simultaneous stillness??)
Given that this guide is made for English speakers and noticing or analyzing vocabulary/word choice depends on proficiency, it's okay to not spot these right away. They may be easy to spot for native English speakers because of their sheer nonsensicality. They are made-up comparisons that don’t map cleanly onto the sensation described, similes that merely sound poetic on a surface level but don’t clarify their meaning. This is the LLM spitting out something that "sounds" like human writing but really isn't.
These are also more suspicious if the author has native-level English proficiency and still includes these, as anyone else who does would see that they don't make sense.
CONSTRAINTS: Humans can write bad similes too. Experimental prose can blur logic intentionally. A key tell is frequency and lack of reasonable payoff.
similarly...
7. Faux‑Poetic but Empty Language
Again, this is highly subjective, but often paired with other tells, and the pattern does exist. Traits include adjectives stacked for tone rather than clarity, and metaphors that don’t resolve into concrete imagery. Flowery and poetic but tasteless. Think of semantic fog. Metaphors without referents, and adjectives chosen purely for tone. The LLM knows what “poetic” sounds like but doesn’t experience the image it’s describing.
Examples:
“The aftertaste of something bitter made bearable.” (the aftertaste of something bitter can easily be imagined. but something bitter made bearable... when has that ever been said??? doesnt make sense at all)
“Just enough to fill the corners.” (of what???? how??? i dont have the original context of this line though sorry)
CONSTRAINTS: Some human writers genuinely like abstraction. The catch is when the abstraction replaces meaning instead of complementing it. This is still the section most vulnerable to unfair application, so I'm still working on it.
8. Tense + Voice Inconsistencies
- Summary written in present tense, fic written in past (or vice versa)
- Generally perfect/standard grammar interrupted by awkward or incorrect tense shifts
Both of these can suggest human edits layered onto generated text in an attempt to either disguise the LLM text or to add details that they could not get the LLM to generate.
Final Reminder
AI detection in fanfiction is fundamentally about pattern recognition. No single trait or stylistic choice indicates AI use, and creative experimentation should never be conflated with automated writing. Context, repetition, and density of patterns are far more telling than any isolated sentence or single indication. Some human writers, especially prolific fanfic writers, write pieces that may naturally resemble LLM-generated content, because they’ve internalized the same dominant internet prose rhythms that the model was trained on. This guide is intended to help readers articulate why something feels off or mechanically similar, without making direct accusations.
As a practical aid, readers may find it helpful to compare works suspected to be AI-generated with works that are more confidently identified as AI-written, such as those openly discussed or flagged in communities. Doing so can illustrate recurring patterns in tone, phrasing, or structural tendencies, and strengthen one’s understanding of the characteristics described here.
It is also worth considering possible motivations for AI use. Some writers may turn to AI because they feel that their English (if their native language is not English) is not strong enough, or they struggle to translate complex ideas into written form. While these users should not be shamed, it's still important to encourage skill development and genuine improvement in writing rather than reliance on automation and rotting their brain. In contrast, AI use for instant gratification, such as accumulating kudos or comments, seeking fame or notoriety, or artificially growing a fanbase, is ethically problematic and should be recognized as such. In any case, habitual use of AI in creative writing undermines the development of writing skill, and readers and writers alike should treat its application critically.
Writing is not meant to be effortless or about churning out large volumes for clout. Within the process, there is meant to be labor, revision, frustration, and passion. Making mistakes is a necessary part of growth, especially for those learning a language or honing craft. Readers deserve authentic storytelling, and for the sake of real creative work, personal development, and the planet, AI-generated fics should not be presented as genuine writing. If you recognize AI writing, do not reward it with attention or praise, or distance yourself from it if you wish to prioritize human-created content, as instant gratification is the primary incentive keeping AI-generated content alive, and discuss with others if you feel keen to. Much of the discomfort around AI-generated fiction comes from a sense of loss of humanity. Do not be afraid to create.
Reference works widely discussed as AI-generated to compare and contrast with:
https://archiveofourown.org/works/66992944/chapters/172945585
https://archiveofourown.org/works/73082056
https://archiveofourown.org/works/73995916
https://archiveofourown.org/works/70005076/chapters/181714291