Common Tells and Tics of AI-Generated Text
Preface + How AIs Generate Text
This is not meant to be an exhaustive list. It’s a collection of some of the most common AI writing tells and patterns. Recognizing AI prose is an exercise in seeing how many of these tells are present in a story in aggregate, as well as learning the characteristic rhythms of AI-generated sentences.
All of these tells stem from how Large Language Models (LLMs), like Claude and ChatGPT, generate prose. The chief thing to keep in mind is this:
LLMs have no way of knowing whether any particular sentence makes sense. They can’t actually understand anything. They are not sentient.
What they do have are giant sets of data they were trained on, i.e. existing writing, pretty much all of which was used without permission or compensation. When someone tells an LLM to generate a story, the LLM will spit out arrangements of words that have a high probability of being “correct” based on:
User-provided prompts and context
Constraints programmed into the LLMs themselves (which are incredibly complex)
When we say “correct,” please know it has nothing to do with factual correctness, or even internal story cohesion. It’s all about how likely it is that the next word to appear in a sequence matches the patterns that have been set based on the three constraints above.
LLMs excel at patterns that stick to consistent patterns and rules, so: spelling and some basic rules of grammar.
It often starts choking on its own spit at the scene and story level.
Keep all these things in mind as you go through the list of tells below.
You’ll also notice that a lot of these tells are things real people do—people fuck up sex scene blocking all the time! People also suck at writing concisely (*shoves tens of thousands of words cut from first drafts under the rug, where they bulge grotesquely*), mess up timelines, forget where their characters are in the room or even in the country, etc. AI was trained on human writing, and all its tics and foibles are based on human patterns.
That said, it’s still possible, with practice, to spot AI-generated prose, especially if it hasn’t been human-edited. We’re not arguing that every fic that uses em dashes, or rules of three, or contains continuity errors is AI-generated. It’s more about learning how to recognize patterns, seeing how often the patterns appear, and taking in the whole picture. Even then, it’s not a guarantee—unless the author has ‘fessed up to using AI, or unless they’ve accidentally left a prompt in the fic, this is more about building a case piece by piece.
Here are some building blocks for making that case.
Common Tells for AI-Generated Text
Extremely prolific output, sustained over a period of months
In our opinion, this is the one strongly dispositive sign in this list, and it’s also the only one that requires zero subjective judgment. The Heated Rivalry TV show fandom is unique because it has a very recent start date and an improbable number of highly prolific fanwriters who started publishing fics in immense quantity right after the show came out. EvilHarlowe, possibly the most prolific writer in the fandom before she deleted her account in early April of 2026, would publish up to tens of thousands of words a day, requiring her to type at a rate of something like 3000 to 5000 words per hour (assuming she wrote eight hours every day, seven days a week). Tendonitis alone would’ve posed a problem at those rates.
Some of these prolific authors claim to have pre-written these works before the TV show came out because they’d read and loved the books, and it’s possible a few may be telling the truth, but it’s just not plausible for the vast majority of the highly-prolific writers to be doing this. If someone’s been writing 80k novels every month for four months, it’s almost certainly AI.
And listen: hyperfixation can get our ass; whomst amongst us hasn’t on occasion output an ungodly number of words in a short amount of time because a giant wad of dopamine had been dumped directly into our brain. But the key thing about these bursts of productivity is that they’re short-lived and draining: writing is work and requires energy. Maintaining that kind of output over a period of months is basically impossible unless the writer resorts to using an LLM.
Polished at a sentence level but incoherent at the paragraph and scene level
Drafting quickly and hitting publish on an unbeta’d draft can result in all kinds of structural issues and mistakes—timelines might get mixed up, passages can be accidentally repeated, or, on the flipside, crucial setup might be accidentally omitted. However, those kinds of mistakes are usually accompanied by other signs of hasty drafting: sentences that are less polished, typos, grammatical errors. That doesn’t tend to happen for AI. Perfectly clean sentence-level prose accompanied by a messy or incoherent story is one of the specialties of LLM-generated fiction.
Concise, precise writing is not a strength of AI prose. There’s a tendency to bloat sentences with complex constructions, too many metaphors, and irrelevant details (many of them bizarre). More detail on the metaphor and irrelevant detail issue below.
Grammatically correct but nonsensical sentences/metaphors.
These read smoothly but have no meaning, e.g., “He filed his hands down the bedspread.” These sentences often feel a little like something made up by an alien who’s trying to demonstrate its humanity—look at how poetic and profound!
The bizarre quality of these constructions come from the LLM making a series of associations and leaps based on probability and how closely words appear in its training set. When a word has multiple meanings, the LLM can stumble; it can fall on its face outright if the word has verb and noun forms that are spelled identically but used very, very differently. In the “He filed his hands down the bedspread” example, the chain of association the LLM went through probably looked something like:
File → Used on hands (nail file) → Filing something down
Hence: “He filed his hands down the bedspread,” a sentence no human would write on purpose in a Heated Rivalry fic.
These nonsensical sentences are especially jarring because they tend to be correct in every other particular. Spelling, punctuation, syntax are all correct (at least for reasonably formal registers). But the arrangement betrays a lack of understanding of how humans think or how the world works. There's fluency at the word level, but no fluency at anything else. The result is distinctly uncanny valley.
Excessive and weirdly specific details
A pattern many LLMs have picked up on is that sensory detail enriches a text. Lacking human discernment, however, they jam these details everywhere, all the time, regardless of their impact on the pacing, or even if they make sense. If a weird metaphor can be fused with the specific detail, even better. LLM-generated prose in particular is full of smells, and often they’re bizarre, like “seared fish oil.” Hockey rinks smell like Zamboni exhaust, or something burnt, or…stick tape??? Things are constantly humming—the air, the silence, emotions, blood. Situations, bodies, and things can rarely exist as they are; they need to be gussied up.
Themes and figurative language that never reoccur.
LLMs, as noted before, struggle with maintaining cohesion across long stretches of text. This results in a lot of thematic imagery or metaphor being used once and then abandoned, only for a different image to sprout up in its place—e.g., someone has a shark-like smile in chapter one, a Cheshire cat smile in chapter two, and a smile like a knife blade in chapter three. Complicating this tell however is:
Over-usage of pet phrases and metaphors
Some LLMs have pet words and figurative constructions that they use over and over again: hum/humming is a popular one, as is something being etched/drawn/felt etc. in the bones. Silences are never just silent; something has to happen to them. (Oftentimes…humming.) This repetition isn’t a matter of, say, three or four times over the course of a fic—they’ll sometimes appear multiple times in a paragraph.
LLMs also, for unknown reasons, favor certain names: ChatGPT generating Elara as a girl's name over and over again is well-documented, for example. Claude repeatedly uses Marcus Webb, Sarah Chen, Volkov, Voss and a few others—just search Reddit for Sarah Chen or Marcus Voss. The number 47 is a number Claude tends to throw out by default.
The absence of these isn’t dispositive, of course; they’re extremely easy to edit out, and names like Marcus and Sarah are reasonably common. But if a fic almost exclusively uses 47 as a number, it might be worth looking at it more closely. If it straight-up names a character Marcus Webb, it’s almost certainly AI.
Repeating scenes or actions
Once again: humans will sometimes forget to delete a new take of a scene, or repeat themselves inadvertently. LLMs do it in a distinct fashion, however: highly-polished prose describing substantially the same thing in fairly rapid succession.
Weird or implausible sex scenes
Is a character lying on their stomach but in the very next sentence—sometimes the next phrase—their navel is visible? Do the limbs never add up? Is the character licking a hip crease before the clothes come off? Many LLMs, either due to the exclusion of sexually-explicit material from the training set or programmed restrictions on generating sex scenes (or both), struggle with them. Tracking how bodies interact and where they are in space is genuinely complex; LLMs in particular struggle with sequencing (e.g. clothes need to come off first before you can lick bare skin) and body blocking.
Violations of the laws of physics
Does the cum run down their dick and end up in their belly-button? Did someone leave so much spit while sharing a spoon for ice-cream that the spit floats on top of the ice-cream on the next scoop? (This one is bizarre on many, many levels.) Is sweat flowing upward somehow? Again, LLMs don’t actually know anything—and that includes basic laws of nature that we take for granted.
Losing track of locations
Did the characters start the scene in Montreal, but suddenly flip to Ottawa, and then back to Montreal again in a few paragraphs? Wait, were they in a hotel, or were they in an apartment bedroom? Again, LLMs can’t track whether the locations in any given story make sense; all it has are the user prompts and what it has been programmed to spit out as most likely to be correct.
Time warps and other timey-wimey nonsense
Did a character resolve to wait two weeks on a Friday, and then five days later, it’s Tuesday and two weeks have passed? Did a character begin performing a six-hour surgery at midnight, and it’s lunchtime when they finish? LLMs can’t track the passage of time any more than they can understand the laws of physics or track where a character is at any given point in a fic.
Monotonous emotional registers
This one requires a substantial chunk of text to check, but does the story feel strangely stressful and numbing at the same time, because every scene is written at the same heightened emotional pitch? LLM prose, especially when given a prompt for a romance plotline, is prone to being overwrought and maintaining that tone throughout, whether or not it’s called for.
Uniform sentence rhythm and scene pacing
This one also needs a substantial amount of text to check, but AI tends to maintain the same pacing regardless of what’s called for on-page. In particular, it tends to favor bursts of dialogue intercut with lengthy internal musing, even during moments that should be fast-paced, such as a fight scene. Stories play with time dilation vs. compression all the time, but when to use which is always a judgement call—something LLMs are unequipped to do.
Markedly different writing styles in the same work, or a fluency discrepancy between the work and the author’s notes and comments
One of the joys of writing is experimenting with voice and form. We’re not talking about those sorts of deliberate experiments here, though; we’re referring to the fics that abruptly switch from paragraphs of overwrought prose filled with metaphor to prose that’s much more rough and ready with completely different rhythms.
Sometimes it's not so much a style difference as it is a difference in fluency, especially between the author's notes and the fic proper. If the author's notes and comments display a notable fluency gap, not just in vocabulary, but syntactically, it's worth going back and looking for other AI tells. (Just to be extremely clear: this is not a slam on people writing in English when it's not their native language! One of the people writing this guide learned English as a second language. So, you know, fuck yeah write the fic of your dreams in whatever language you want. Just don't fucking use AI to do it.)
A distinctive AI rhythm to the sentences
LLMs lean on a few workhorse sentence structures. To be clear: these are all perfectly legitimate ways to structure sentences! They became popular for a reason; a lot of them provide a nice rhetorical flourish. LLMs, however, over-use these structures to the point of monotony. If you want a crash course in the specific rhythms and stylistic tics of AI, check out this video that breaks down why Shy Girl is AI-generated. It goes through the book in exhaustive detail, and by the end of the video, it should give you a pretty good ear for the rhythms of AI-generated prose.
“The thing about Tom's body was that it never lied to him. It woke at five. It ran its mile.”
Notice this setup: X was that phrase. Pronoun + verb + prepositional phrase. Pronoun + verb + prepositional phrase.
This long–short–short-short sentence pattern is common in Claude-generated prose.
“Practiced neutrality. Revealing nothing. He understood the machinery of it. He was part of the machinery of it.”
Notice how the sentences use this pattern of repetition? LLM prompters call it an “amplification echo” because it repeats previously established information.
“She'd be pleased about it in the way that meant she was not pleased about it.”
The false profundity in this one is a pretty big AI tell (if she’s obviously not pleased, then she wasn’t pleased to begin with!), but check out the setup: Description + in the way that + false negation.
Subject, prepositional phrase, prepositional phrase, predicate with false negation. Once you see the pattern, you can’t unsee it.
“The moment moved on the way it did in this house—slow, interrupted by small children, going on without anger.”
Here’s another “the way”, but now it’s embedded in another classic Claude pattern: Description + the way + monotone emotion phrase, em dash, part 1, part 2, part 3.
Read it out loud to yourself and feel that rhythm. Notice how it throws in a metaphor that feels tantalizingly profound, but that falls apart upon examination.
“Her hands worked Sam's belt. Button. Zipper.”
This is an extremely common fragment list pattern, often deployed whether or not it suits the emotional tone or pacing. Short subject. Part 1. Part 2. Part 3.
"Sam looked at him—that direct, unhurried way he had, the way that didn't apologize for itself."
Yet another “the way,” but hopefully you can begin to recognize the pattern. Short subject + prepositional phrase, em dash, predicate with a phrase with falsely profound negation or metaphor.
On display here is another popular LLM construction: This, not that. You will also see the obverse: Not that, this. And speaking of popular LLM sentence constructions:
"Not unkind. Not impatient. Just—waiting."
This construction has become one of the most famous AI tells, and is especially common with ChatGPT-generated prose; it’s all over places like LinkedIn and in the descriptions of AI slop videos on YouTube and Instagram. Not X, not Y, but secretly Z.
“He had a voice like a waiting room — neutral, practiced. Designed to contain things.”
Superficially profound metaphor spotted! But we also have this pattern: Short subject + metaphor, em dash, part 1, part 2. Part 3.
This is an abstract-noun character description, which is one of the hallmarks AI-generated prose. These nouns are often meaningless in context, and an example of the way AI-generated prose defaults to telling instead of showing.
Like we said, this list isn't comprehensive. And once again: it's not about a fic containing any one particular thing from this list, or even two or three. It's about a sustained and repeated pattern. Feel free to add other tells we might've missed in the reblogs. Hopefully this helped give you an idea of the sorts of things to look out for.
