Why AI Writing Detection Systems Are Infuriating

Lately, AI writing detection has become a hot topic, with debates over whether it's reliable or not. Apparently, even university departments are using AI detection tools to check student assignments.

The problem is that AI detection is described as working through writing patterns.

Below are word patterns flagged as AI-generated. I compiled these mainly from Wikipedia articles and GitHub posts labeled as human tokenizer discussions of AI patterns.

Word Choice

• "Quietly" and magic adverbs: A pattern of overusing adverbs like "quietly", "deeply", "fundamentally", "remarkably", "arguably", and "Ultimately" to lend subtle weight to otherwise ordinary descriptions

  • Examples: "quietly orchestrating workflows", "a quiet intelligence behind it"

• "Delve" and its synonyms: Once the most notorious AI tell, this word appears at abnormally high frequency in AI-generated text

  • "certainly", "utilize", "leverage" (as a verb), "robust", "streamline", and "harness" fall into the same category

• "Tapestry" and "Landscape": A pattern of using grandiose nouns where simpler words would do

  • "tapestry" gets overused for anything interconnected, and "landscape" for any field or domain

  • "paradigm", "synergy", "ecosystem", and "framework" belong to the same type

• Avoiding "Serves As": Using inflated linking verbs like "serves as", "stands as", "marks", and "represents" instead of simple "is/are"

  • This happens because the AI's repetition penalty pushes it toward ornate constructions rather than basic copulas

Sentence Structure

• Negative parallelism: The "It's not X -- it's Y" pattern, the most commonly identified AI writing tell

  • It wraps everything in a dramatic reframe, generating false profundity

  • This style of mass writing did not exist before LLMs

  • The causal variant "not because X, but because Y" is included as well

• "Not X. Not Y. Just Z.": A dramatic countdown pattern that negates two or more things before revealing the actual point

  • It creates a false sense of narrowing in on the truth

• "The X? A Y.": A rhetorical question-and-immediate-answer pattern where the writer poses a question no one asked, then answers it right away

  • Used for dramatic effect, and AI treats this as the essence of good writing

• Anaphora overuse: Rapidly repeating the same sentence opening multiple times in succession

  • Example: "They assume that... They assume that... They assume that..."

• Tricolon overuse: Over-applying the rule of three, sometimes stretching it to four or five items

  • A single tricolon is elegant, but three in a row signals a pattern-recognition failure

• "It's Worth Noting": A filler transition that signals nothing

  • "It bears mentioning", "Importantly", "Interestingly", and "Notably" fall into the same category

  • They introduce a new point without actually connecting it to the prior argument

• Superficial analyses: Tacking a present-participle ("-ing") phrase onto the end of a sentence to inject shallow analysis

  • Expressions like "highlighting its importance", "reflecting broader trends", and "contributing to the development of..."

  • They assign significance, legacy, and broad implications to mundane facts

• False ranges: "from X to Y" constructions where X and Y don't actually sit on any meaningful scale

  • In legitimate use, the phrase implies a spectrum with meaningful midpoints; AI uses it to list two loosely related things

• Gerund fragment litany: After making a claim, stringing together a series of subjectless gerund fragments

  • "Fixing small bugs. Writing straightforward features. Implementing well-defined tickets."

  • The first sentence already said everything; the fragments only add word count and an AI-characteristic rhythm

  • Humans don't write first drafts this way; it's a pure structural tic

Paragraph Structure

• Short punchy fragments: Using very short sentences or sentence fragments as standalone paragraphs to create artificial emphasis

  • The result of RLHF training pushing the model toward "writing for readability" aimed at the lowest common denominator reader

  • One thought per sentence, no sustained mental state required: an inhuman style

• Listicle in a trench coat: Disguising numbered or labeled points as continuous prose

  • A pattern of hiding list structure inside paragraphs that begin "The first... The second... The third..."

  • Models often adopt this as a workaround after being told to stop generating lists

Tone

• "Here's the Kicker": A false-suspense transition that promises a revelation but delivers a point that needed no such buildup

  • "Here's the thing", "Here's where it gets interesting", and "Here's what most people miss" belong to the same category

• "Think of It As...": A teacher-mode default that assumes readers need an analogy to understand anything

  • AI frequently generates analogies that are less clear than the original concept

• "Imagine a World Where...": A classic AI futurist invitation, where "Imagine" is followed by a list of wonderful things that will happen if you accept the premise

• False vulnerability: Performative self-awareness that breaks the fourth wall or pretends to acknowledge bias

  • Real vulnerability is specific and uncomfortable; AI's version is polished and risk-free

• "The Truth Is Simple": A pattern of asserting that something is obvious or simple rather than actually proving it

• Grandiose stakes inflation: Inflating the stakes of every argument to world-historical significance

  • The phenomenon of a blog post about API pricing turning into a meditation on the fate of civilization

• "Let's Break This Down": A pedagogical voice that defaults to a teacher-student relationship even with expert readers

  • "Let's unpack this", "Let's explore", and "Let's dive in" fall into the same category

• Vague attributions: Attributing claims to unnamed authorities like "experts", "observers", and "industry reports" without specific sources

  • This includes inflating what one person said into a widespread view, or citing two sources as "several publications"

• Invented concept labels: Attaching abstract problem nouns (paradox, trap, creep, divide, vacuum, inversion) to domain words to create synthetic labels that sound analytical but lack grounding

  • Examples: "supervision paradox", "acceleration trap", "workload creep"

  • They function as rhetorical shorthand that names something and skips the argument; multiple instances in a single piece are a strong signal of AI slop

Formatting

• Em-dash addiction: Compulsively overusing em dashes for dramatic pauses, parenthetical insertions, and turning points

  • A human writer naturally uses two or three in a piece; AI uses twenty or more

• Bold-first bullets: A pattern where every bullet point begins with a bold phrase

  • Extremely common in Claude and ChatGPT markdown output, while almost no one formats this way when writing by hand

  • A reliable tell in AI-generated documents, blog posts, and README files (especially when emoji are included)

• Unicode decoration: Using special characters not easily typed on a standard keyboard, such as Unicode arrows (→) and smart/curly quotation marks

  • Real writers use straight quotes and ->, => in their text editors

  • Claude in particular favors the → arrow

Composition

• Fractal summaries: Applying the "say what you'll say, say it, say what you said" structure at every level of the document

  • Every subsection, section, and the document itself each gets its own summary

• The dead metaphor: Latching onto a single metaphor and repeating it throughout the entire piece

  • A human writer introduces a metaphor, uses it, and moves on; AI repeats it five to ten times

• Historical analogy stacking: Particularly common in technical writing, this involves rapidly listing historical companies or technological revolutions to build false authority

  • Patterns like "Apple didn't build Uber. Facebook didn't build Spotify..."

• One-point dilution: Restating a single argument in ten different ways across thousands of words

  • Padding the same idea with different analogies, examples, and framings to appear "comprehensive"

• Content duplication: Repeating entire sections or paragraphs verbatim within the same piece

  • Occurs when the model loses track of what it has already written, especially in long documents

  • A clear tell of unedited AI output, though less common recently

• The signposted conclusion: Explicitly announcing the conclusion with phrases like "In conclusion", "To sum up", or "In summary"

  • Skilled writing lets readers feel the conclusion rather than announcing it

  • AI signals structural moves because it follows a template

• "Despite Its Challenges...": A rigid formula where AI acknowledges a problem but immediately dismisses it

  • It always follows the same beats: "Despite its [positive word], [subject] faces challenges..." followed by "Despite these challenges, [optimistic conclusion]"

The writing patterns above are ones I collected and saved from AI writing pattern discussions both overseas. Every single one of them is a pattern attributed to AI-generated writing.

And the idea is that if these patterns appear frequently, the writing must be AI-generated. But when you think about it, many of these are styles that appeared regularly in technical textbooks from the 2000s and 2010s.

Technical textbooks, in general, don't just list sentences to transfer knowledge. Getting readers to accept concepts they don't yet know requires emphasis, repetition, and contrast.

So writing naturally tends toward these sentence structures.

• This is not A; it is B.

• We can divide this into first, second, and third.

• The important point is this.

• In other words, this is not simply an implementation problem but a design problem.

• On one hand it is X, but on the other hand it is Y.

These sentence structures were not invented by AI.

They have been used continuously in old expository writing, essays, lecture notes, textbooks, and translations. technical textbooks in particular have always relied heavily on this style: repeating key terms, using contrastive constructions, numbering items, and opening paragraphs with phrases like "the important point is." All of this exists to keep readers from getting lost.

The problem is that AI detection systems classify all of these long-established explanatory techniques as smelling of AI.

"It's not X, it's Y" is not an AI style; it is antithesis. The tricolon is not an AI style; it is a rhetorical device used in classical rhetoric for centuries. The em dash (—) is not an AI style; it is a punctuation mark used in English-language prose for parenthetical insertions and transitions. Bullet points (•) are not an AI style; they are a format used in technical documentation and lecture materials to reduce readers' cognitive load.

In Korea, using bullet points gets you labeled as AI Slop. People ask why you're using em dashes, but programmers, by the nature of their work, inevitably read English-language material, and they often save highlighted points from that material and write them out in direct translation when composing blog posts.

So everything they stored that way gets flagged as AI writing. On top of that, because AI avoids extreme claims and aims for neutral positions, most AI writing actually stays within safe, conventional sentence forms.

Looked at this way, there is a real problem here: all of that writing is, in fact, exemplary writing.

The logic is simple. Making an extreme, forceful argument basically means challenging the mainstream, and because that affects social relationships, people generally write in a way that mildly critiques the mainstream while landing in a safe spot.

Take Korea, for example.

Say someone writes a piece arguing that "the parents you're born to largely determines the course of your entire life." There's plenty of research supporting this. It has long been established that parental connections from academic and regional networks, along with family wealth, shape a child's experiences and academic outcomes. Topics like this tend to be very heavy, and they usually meet strong resistance from readers, so most writers soften the argument and look for a safe landing point.

"Korea's system is structurally broken and needs to change."

This is essentially the same move AI often makes: framing things as a "structural" problem or a "systemic" problem, using grand language while ultimately pivoting to a larger systemic framing in order to land safely at the small-unit level. Once you add social safety nets, individual effort, and policy support into the mix, people say it's 100% an AI-written sentence.

In the end, the very grammar of an intellectual discussing refined, macro-level solutions becomes something written by AI.

Writing this way these days means there's a high chance it gets falsely flagged as AI.

That makes sense, in a way. Words like "structural" appear frequently, and choosing only safe, non-extreme words makes the writing look like something AI produced.

So if you want it to look like a real person wrote it rather than AI, you apparently have to do this:

"The parent lottery determines most of life, and academic background is influenced by the parents' academic backgrounds and money. Therefore, lives that fail at the parent lottery have a high probability of failing. So, you seeds who were not born in a good school district in Seoul and failed at the parent lottery, live forever as losers."

You have to write in a way that extreme and defeatist before it stops being labeled as AI-generated.

Why? Because commercial AI models are trained to exclude extremism, defeatism, and fatalism.

So to write about the complex problems of Korean society in a way that doesn't get flagged as AI, you have to tip into hatred and extremism.

On top of that, to avoid AI grammar patterns, the resulting writing must be uneven and full of grammatical errors, which is an outcome people are actually being pushed toward.

The problem is that once this filtering system becomes entrenched in the writing ecosystem, middle-ground writing (the kind that lands safely) gets dismissed as mass-produced AI content.

Phrases commonly called vague attributions, like "many scholars view it as "or" "it is genereally understood that", are themselves being treated as the exclusive domain of AI. Academic-style writing is being AI-ified.

Non-English-speaking developers face even higher rates of AI detection in their writing, because they have fewer options for expressing ideas compared to native English speakers, and most of those options are limited to a narrow set of commonly used words ¹. In fact, Stanford researchers ran 91 TOEFL essays written by non-native speakers through seven GPT detectors and found that more than half were misclassified as AI-generated. By contrast, more than 90% of essays written by American eighth-graders were classified as human-written.

What does this mean? People in Korea who study English-language documents end up substituting Korean for English as they learn, and information loss occurs during that substitution. This is the price non-English-speaking developers pay. When they fill in that information loss from memory, they fill it with the commonly used English terms they already know,

and because developers mostly remember things through the words they use most often, technical writing ends up reading even more like AI. This problem used to be called the "literal translation style" issue, but now it's called the AI Slop problem.

The writing style of developers that used to be mocked as "translation-influenced" or a "literal translation style" was actually part of how technical knowledge was learned and spread.

Repeating precise technical terms rather than using flashy rhetoric, and using layered bullet points (•) to reduce cognitive load, simply made things easier to understand.

Frankly, when you're reading technical documentation, you don't want to encounter flowery language. You just want an explanation of how something runs.

"When you call A here, it executes as B."

That's what you want to read.

You don't want to read a long, grammatically tangled mess like, "Regarding this part, if one were to invoke A, it would appear that A is being called, but as for the reason it executes as B, to speak to that..."

These days, efficient sentences are being discarded and labeled "AI's work (AI Slop)."

For instance, constructions like "what matters is not A but B", which AI supposedly uses often, were actually a very convenient way to maintain the flow of an argument.

You would spend time making the case for A, then show that a problem had emerged with A, and introduce B as the solution, a kind of setup-and-payoff style of writing. But now, people say you have to abandon this too.

In the end, to get views and to signal that their writing is human, people are increasingly forced to choose more provocative and extreme wording. Is this really the right direction?

In the name of finding writing that AI didn't produce, people say that misspellings and extreme claims are proof of a human author. The implication is that reason and logic should be outsourced to AI, while humans monopolize barbarism and hatred.

I think about Goodhart's Law when I consider this problem.

"When a measure becomes a target, it ceases to be a good measure."

From a programmer's perspective, current AI detection systems are malfunctioning linters. If you write clean, optimized code and the compiler spits it out saying it looks too perfect to be human-written, is that normal?

To pass the compiler, you write spaghetti code; to dodge algorithmic false positives, you lower the quality of your work; you make extreme claims. This isn't the dog wagging the tail; this is the tail wagging the dog.