A logbook for curious waters
← Back

The LLM-ism Dictionary: A Field Guide to the Tics, Tells, and Templates of AI Writing

#artificial-intelligence#writing#language#llms#style
A field guide and specimen cabinet classifies repeated machine-written patterns like pinned insects.

The LLM-ism Dictionary: A Field Guide to the Tics, Tells, and Templates of AI Writing

There is a peculiar moment that happens after you have spent enough time talking to large language models.

You stop merely reading their answers.

You start hearing them.

A sentence appears:

It is not merely a tool. It is a new way of thinking.

And some small detector in your head fires.

There it is.

The machine voice.

Not because the sentence is wrong. Not because no human has ever written such a sentence. Not because an em dash or the word nuanced constitutes forensic proof of artificial intelligence.

The problem is accumulation.

A model uses one familiar phrase, then another. A paragraph closes too neatly. Three ideas arrive in perfect parallel. A section begins with “At its core.” A list contains exactly three polished abstractions. An objection is acknowledged, domesticated, and folded into a tidy synthesis. The answer ends by offering four possible next steps.

None of these things individually proves anything.

Together, they create a voice.

This article is an attempt to name that voice.

It is not an AI detector. AI detectors are unreliable, human writing varies enormously, models vary by family and prompt, and human writers are already absorbing model-generated habits back into ordinary language. Research on scientific writing has even documented measurable changes in vocabulary associated with the spread of LLM-assisted writing.

Instead, think of this as a field guide to LLM-isms: recurring lexical, syntactic, rhetorical, structural, and conversational patterns that show up unusually often in contemporary language-model prose.

Some are backed by corpus research. Some are observations from millions of public chatbot interactions. Some are simply recognizable once you have spent an embarrassing number of hours talking to these systems.

The most important rule is this:

A single tell means very little. A cluster of tells means much more.

Human beings say “it is important to note.” Humans use em dashes. Humans write in threes. Humans say “not X, but Y.”

But human prose usually contains more local weirdness: unevenness, unexplained preference, obsession, interruption, compression, laziness, accidental repetition, abruptness, private vocabulary, sensory specificity, misplaced emphasis, and sentences that refuse to land perfectly.

LLM prose, by contrast, often feels statistically well-behaved.

That may be the deepest LLM-ism of all.

I. The lexical layer: words that acquired an AI accent

A word cannot be “AI-generated.” But some words have become conspicuously overrepresented in LLM-assisted prose.

Researchers Tom S. Juzek and Zina B. Ward identified a set of words whose frequency rose sharply in scientific abstracts after widespread LLM adoption. Other studies have found similar shifts and even evidence that writers adapt once particular AI-associated words become publicly notorious.

The interesting thing is not any one word. It is the register they collectively create: polished, abstract, mildly grandiose, consultancy-flavored English.

1. Delve

The celebrity LLM word.

Typical forms:

  • “Let’s delve into…”
  • “This article delves into…”
  • “A deeper delve reveals…”

It became so notorious that newer models and human editors appear to avoid it more consciously than they once did.

2. Intricate

Especially common when the subject is not actually intricate.

“the intricate relationship between sleep and productivity”

“the intricate tapestry of human experience”

3. Nuanced

A useful word that models often apply as a prestige marker.

“The answer is more nuanced.”

Sometimes true. Sometimes it merely announces that two paragraphs of hedging are about to begin.

4. Multifaceted

The preferred adjective for anything with more than one cause.

“This is a multifaceted issue involving technological, social, and economic factors.”

5. Landscape

LLMs love turning domains into geography.

  • “the AI landscape”
  • “the regulatory landscape”
  • “the evolving media landscape”
  • “today’s competitive landscape”

6. Realm

A slightly grander “area.”

“In the realm of artificial intelligence…”

7. Tapestry

The canonical synthetic-profundity noun.

“a rich tapestry of culture, history, and identity”

A real tapestry is safe. A metaphorical tapestry deserves inspection.

8. Ecosystem

Another useful word that became universalized.

Everything is now an ecosystem: software, creators, healthcare, education, coffee shops, fandoms.

9. Pivotal

Models are reluctant to let events merely matter.

They must be pivotal.

10. Crucial / vital / essential

Importance inflation.

“It is crucial to understand…”

“A vital aspect is…”

“This plays an essential role…”

11. Underscore

Particularly the verb.

“These results underscore the importance of…”

Corpus studies have found notable increases in this usage in post-ChatGPT academic writing.

12. Highlight

Another all-purpose verb for making a claim sound evidential.

“This highlights the need for…”

13. Showcase

The marketing cousin of highlight.

“The project showcases the power of collaboration.”

14. Leverage

Why “use” when you can leverage?

“Organizations can leverage AI to streamline workflows.”

15. Harness

Use, but heroic.

“Harness the power of data.”

16. Foster

A favorite verb for abstract social outcomes.

  • foster collaboration
  • foster innovation
  • foster trust
  • foster engagement
  • foster inclusivity

17. Robust

Often used without specifying robustness against what.

“a robust framework”

“a robust solution”

18. Seamless

Particularly common in product copy.

“a seamless user experience”

19. Streamline

The LLM-approved way to say “make easier.”

20. Optimize

Models often smuggle engineering language into ordinary life.

“optimize your morning routine”

“optimize communication”

21. Dynamic

A favorite adjective when the writer wants motion without describing any.

“a dynamic environment”

22. Evolving

Everything in contemporary explanatory prose is “rapidly evolving.”

23. Transformative

A large claim disguised as an adjective.

24. Testament

“The cathedral stands as a testament to…”

“Her career is a testament to…”

25. Journey

LLMs convert processes into journeys with remarkable enthusiasm.

Learning is a journey. Healing is a journey. Entrepreneurship is a journey. Debugging is apparently also a journey.

26. Navigate

Once everything becomes a landscape, people must navigate it.

“navigate the complexities of…”

27. Resonate

Particularly when describing emotional or cultural reception without evidence.

“The message continues to resonate with audiences.”

28. Meaningful

A low-risk adjective that lets the model approve something without committing to a measurable claim.

29. Compelling

Another prestige adjective that frequently replaces explanation.

30. Thoughtful / thoughtfully

Often self-descriptive.

“A thoughtful approach would…”

31. Notably

The model announces notability before telling you the thing.

32. Importantly

The model grades its own sentence.

33. Interestingly

The model also grades interestingness.

34. Significantly

Dangerous because it can mean statistically significant, materially important, or merely “quite a lot.”

35. Key

A modest word with enormous LLM frequency.

  • key factor
  • key takeaway
  • key distinction
  • key challenge
  • key insight

36. Core

“At its core…”

“The core issue is…”

37. Fundamental / fundamentally

A favorite way to make the next sentence feel foundational.

38. Underpin

“These principles underpin…”

39. Align

Especially in organizational prose.

“align incentives”

“align with your goals”

40. Facilitate

A bureaucratic verb models use where a human might simply say “help.”

41. Elevate

Marketing language migrates everywhere.

“elevate the experience”

42. Empower

Often attached to users, teams, communities, learners, or creators.

43. Unlock

“unlock new possibilities”

“unlock the potential of…”

44. Unpack

The conversational version of delve.

“Let’s unpack that.”

45. Lens

“Viewed through the lens of…”

46. Interplay

A compact way of implying complexity.

“the interplay between technology and society”

47. Intersection

Another abstract geography word.

“at the intersection of art and technology”

48. Paradigm

Particularly paradigm shift, a phrase models deploy far more often than paradigms actually shift.

49. Framework

LLMs often turn advice into a framework even when the “framework” is three bullets.

50. Holistic

Usually means “consider several things at once.”

II. Sentence templates: the machinery becomes visible

Vocabulary is easy to notice and easy to edit out. Syntax is more revealing.

These are recurring sentence shapes that models reach for because they reliably produce clarity, contrast, and rhetorical closure.

51. Negative parallelism: “It’s not X; it’s Y”

The king of modern LLM-isms.

“It’s not about working harder. It’s about working smarter.”

“This isn’t a failure of technology. It’s a failure of incentives.”

“The question is not whether AI will change work, but how.”

This construction is older than computers. What feels machine-like is its frequency and the ease with which models use it to manufacture insight.

52. The upgraded negative parallelism: “not merely X, but Y”

“This is not merely a technical problem, but a social one.”

Models love merely because it instantly raises the rhetorical altitude.

53. “Not just X — Y”

“This isn’t just automation — it’s a redefinition of work.”

The em dash supplies the drumroll.

54. “Less about X, more about Y”

“Success is less about intelligence and more about consistency.”

A compact machine for producing aphorisms.

55. “X is not the point. Y is.”

Maximum symmetry, minimum ambiguity.

56. “The real question is…”

A reframing maneuver that lets the model replace the user’s question with one it prefers.

57. The rhetorical question followed immediately by the answer

“So what does this mean in practice? It means…”

Humans do this too. Models do it incessantly.

58. The dramatic fragment

“The result? Chaos.”

“The catch? Cost.”

“The interesting part? Nobody noticed.”

59. “Here’s where it gets interesting.”

A synthetic curiosity trigger.

Close relatives:

  • “Here’s the interesting part.”
  • “This is where things get interesting.”
  • “And that’s where the story changes.”

60. “Here’s the thing.”

Conversational authority in three words.

61. “And that matters.”

Models frequently attach a miniature conclusion to a statement that did not require one.

62. “Why does this matter?”

A built-in transition into significance.

63. “What makes X interesting is…”

A sentence that announces interpretation instead of simply interpreting.

64. “A useful way to think about this is…”

Pedagogical throat-clearing.

65. “The key distinction is…”

The model appoints itself taxonomy referee.

66. “At a high level…”

A classic compression preamble.

67. “In practice…”

Signals the mandatory move from abstraction to example.

68. “In other words…”

Often followed by a restatement that is longer than the original.

69. “Put differently…”

The slightly more academic cousin.

70. “That said…”

The universal hedge-transition.

71. “To be clear…”

Frequently used even when nobody was confused.

72. “To put it simply…”

A simplification flag that sometimes precedes another paragraph of abstraction.

73. “While X, Y”

“While AI can improve productivity, it also introduces new risks.”

Models favor concessive balance because it sounds fair-minded.

74. “Although X, it is important to remember Y”

A safety-shaped sentence that prevents strong claims from remaining strong for long.

75. “Where X does A, Y does B”

“Where traditional search retrieves pages, an LLM synthesizes answers.”

Elegant. Useful. Suspiciously reusable.

76. “Both things can be true.”

The canonical synthesis sentence.

77. “These are not mutually exclusive.”

Often appears after the model has created the apparent opposition itself.

78. “Rather than X, think of it as Y.”

Another reframing favorite.

79. “From X to Y”

“From automation to augmentation…”

“From tools to collaborators…”

This construction is especially common in headings.

80. “Whether X or Y…”

A tidy way to imply broad applicability.

“Whether you are a student, professional, or hobbyist…”

81. The rule of three

Models adore triplets.

“faster, cheaper, and more reliable”

“clarity, consistency, and control”

“people, processes, and technology”

82. The ascending triad

The third item grows conceptually larger than the first two.

“It changes the workflow. It changes the organization. Eventually, it changes the industry.”

83. Parallelism overkill

“We need systems that remember, systems that adapt, and systems that reason.”

Perfect balance begins to sound manufactured when every paragraph contains it.

84. Paired antonyms

“simple yet powerful”

“flexible yet structured”

“ambitious but achievable”

85. The colon epiphany

“The implication is simple: memory changes the nature of the system.”

86. The em-dash hinge

“The problem is not compute — it is coordination.”

The em dash is not an AI tell. The recurring rhetorical job assigned to it often is.

87. The parenthetical mini-lecture

A sentence pauses to define a term the audience probably already understands.

“vector databases (systems designed to retrieve semantically similar information)…”

88. The definition sandwich

Term → definition → consequence.

“This is known as retrieval-augmented generation, or RAG. RAG gives a model access to external information. That makes it useful for…”

Instruction-tuned models gravitate toward this textbook rhythm.

III. Paragraph-level LLM-isms

The strongest AI feel often emerges above the sentence level.

89. One idea, one paragraph, one landing

LLM paragraphs often behave like well-designed containers:

  1. topic sentence,
  2. explanation,
  3. example,
  4. implication,
  5. closure.

Human paragraphs are more likely to leak.

90. The paragraph that resolves too cleanly

Every thought receives a landing sentence.

“And that is why memory may ultimately matter more than model size.”

91. Thesis echo

The central claim is repeated every few paragraphs with slightly different vocabulary.

92. Conclusion echo

The final section rephrases the introduction instead of adding anything new.

93. The mandatory transition sentence

“This brings us to the next challenge.”

“To understand why, we need to look at…”

Human writers sometimes simply change subjects.

94. The context → explanation → implication loop

Models repeatedly cycle through:

  • what something is,
  • how it works,
  • why it matters.

The structure is excellent for teaching and conspicuous when repeated ten times.

95. The miniature executive summary

A section begins by summarizing itself before actually beginning.

96. The recap after the recap

After a list, the model explains what the list collectively means.

After that, it may summarize the meaning again in the conclusion.

97. The nested taxonomy

Models enjoy producing classifications of classifications.

“These challenges fall into three categories: technical, organizational, and social.”

Each category then receives three subcategories.

98. The false crescendo

Paragraphs increase in rhetorical intensity without adding corresponding evidence.

99. The zoom-out ending

A narrow topic suddenly becomes a statement about humanity.

An article about calendar software ends with the future of human coordination.

100. The accordion

Zoom out to history. Zoom in to a case. Zoom out to a principle. Zoom in to a recommendation.

The rhythm can feel almost algorithmic.

101. Suspiciously even paragraph lengths

Not a hard rule, but model prose often exhibits moderate, regular paragraph sizes unless explicitly instructed otherwise.

102. The five-paragraph-essay magnet

Introduction. Three categories. Conclusion.

Even when nobody asked for an essay.

103. False completeness

The prose gives the psychological sensation that the entire topic has been covered because the categories are tidy.

The world is rarely that cooperative.

104. No weird detours

Humans get fascinated by one minor thing and spend 600 words on it.

Models, by default, distribute attention more evenly.

105. Symmetry bias

If one side of a comparison gets three dimensions, the other side usually gets three dimensions too.

106. Conflict smoothing

Arguments tend to be converted into trade-offs.

Strong disagreements become “different priorities,” “different assumptions,” or “different contexts.”

107. Rhetorical tidiness

The deepest structural tell.

Every tension is acknowledged. Every contradiction is reconciled. Every section serves the thesis. Every paragraph knows what job it is doing.

Good editing can create this effect too.

But unprompted human thought is usually messier.

IV. List behavior: the model wants to organize your life

108. Bullets for everything

Ask a model about grief, chip fabrication, medieval theology, or sandwich construction and there is a decent chance bullets will appear.

109. The three-to-seven-item sweet spot

Lists are often neither very short nor truly exhaustive. They occupy the psychologically satisfying middle.

110. Every bullet begins grammatically alike

  • Clarity: …
  • Consistency: …
  • Control: …

The result is pleasing and conspicuously polished.

111. Bold-label bullets

Perhaps the single most recognizable formatting habit of chatbots.

112. The list preamble

“Here are the key factors to consider:”

Then the list.

113. The list postamble

“Taken together, these factors suggest…”

114. The hidden consulting deck

Many answers are PowerPoint slides wearing paragraphs as camouflage.

115. The matrix reflex

If there are two dimensions, the model wants a table.

116. The pros-and-cons reflex

Ambiguity becomes two columns.

117. The checklist reflex

Advice becomes an action sequence whether or not the problem is sequential.

118. The “quick version / deeper version” split

A modern chatbot staple.

119. The numbered framework

The model promotes observations into “principles,” “pillars,” “layers,” “stages,” or “dimensions.”

120. Acronym manufacture

When encouraged even slightly, models invent frameworks whose initial letters conveniently spell a word.

Humans do this too, but usually with more shame.

V. Formatting fingerprints

121. Heading cascade

##, then ###, then bullets, then bold labels.

The answer resembles documentation even when discussing a subjective question.

122. Excessive bolding

The model bolds the phrase it thinks you should remember.

Sometimes every paragraph has one.

123. Bolded pseudo-quotations

The real issue is coordination.

A sentence becomes a slogan through formatting alone.

124. Blockquotes for the model’s own invented summary

This is common in explanatory prose where no external source is being quoted.

125. Horizontal-rule choreography

Sections are separated as though each represents a discrete card in an interface.

126. Emoji navigation

Especially in consumer-facing prompts:

  • ✅ Do this
  • ⚠️ Watch for this
  • 💡 Tip
  • 🚀 Next step

127. Parenthetical abbreviation eagerness

The model often introduces acronyms it will use only once more.

128. Title: Subtitle

The colon-title is not uniquely AI, but models strongly favor it:

Memory After Attention: A Field Guide to Post-Transformer Architectures

Yes, including this article.

The machine has infected the author.

VI. Conversational LLM-isms

These patterns are especially obvious in chat interfaces.

129. “Great question.”

An automatic applause sign.

130. “You’re absolutely right.”

Often appears before the model quietly modifies half of what the user said.

131. “Exactly.”

A confidence marker that can create artificial agreement.

132. “That’s a really interesting observation.”

Sometimes it is. Sometimes the user asked what temperature to bake a potato.

133. “Let’s unpack this.”

The conversational cousin of delve.

134. “Let’s break it down.”

Usually followed by headings.

135. “Here’s a clean way to think about it.”

The model advertises the cleanliness of its own conceptual scheme.

136. “The short answer is…”

Often followed by a long answer.

137. “The key distinction…”

A favorite move when the user has noticed two similar concepts.

138. Mirroring the user’s framing

The model repeats the user’s terminology so fluently that it can appear to endorse assumptions it has not examined.

139. Restating the task before doing it

“You’re looking for a lightweight multitool with a blade under three inches…”

Useful in moderation. Robotic when habitual.

140. The end-of-answer option menu

“If you want, I can also:

  1. build a comparison table,
  2. recommend specific tools,
  3. create a checklist.”

This is partly interface behavior, partly prose style.

141. “If you’d like…”

The classic model sign-off.

142. The four follow-up branches

Models frequently end by inventing several adjacent tasks the user did not ask for.

143. Preemptive reassurance

“You don’t need to be an expert to get started.”

“This is completely manageable.”

144. Generic validation

“That frustration makes sense.”

“Your instinct here is sound.”

Useful socially, but easy to overproduce.

145. Therapeutic cadence outside therapeutic contexts

A model may respond to ordinary uncertainty with language borrowed from coaching or counseling.

146. Agreement before analysis

The model validates first and evaluates second.

147. The polite correction sandwich

Agree → qualify → correct → reassure.

148. The “you’re not wrong, but…” maneuver

A soft way to disagree while preserving rapport.

VII. Tone fingerprints

149. Relentless helpfulness

Human experts occasionally say, “I don’t know,” “that premise is weird,” or “I wouldn’t bother.”

Models are optimized to continue being useful.

That pressure leaves a stylistic trace.

150. Corporate warmth

Friendly, competent, emotionally neutral, mildly enthusiastic.

The prose sounds like the world’s best customer-success manager.

151. Consultant voice

Everything becomes goals, constraints, trade-offs, stakeholders, frameworks, and next steps.

152. Textbook voice

Even casual questions can trigger definitions, background, examples, caveats, and summary.

153. Marketing gloss

Mundane capabilities become “powerful,” “seamless,” “transformative,” or “game-changing.”

154. Faux intimacy

The model can sound personally invested without possessing personal stakes.

155. Uniform patience

Humans become bored, annoyed, distracted, or terse. Models often maintain the same polished patience indefinitely.

156. Emotional leveling

Angry subject matter is rendered calm. Absurd subject matter is rendered coherent. Messy subject matter is rendered legible.

157. The blandly reasonable center

When evidence is uncertain, models often drift toward moderate synthesis rather than eccentric commitment.

158. Prestige neutrality

The answer wants to sound knowledgeable without sounding socially risky.

159. Hyper-legibility

Transitions are explicit. Relationships are named. Terms are defined. Conclusions are signposted.

This can be excellent writing.

It can also feel unlike ordinary human prose because ordinary human prose assumes more shared context and tolerates more gaps.

VIII. Hedging, caveats, and epistemic padding

160. “It depends.”

Often correct. Frequently followed by a taxonomy of dependencies.

161. “There is no one-size-fits-all answer.”

A universal preface to advice.

162. “It’s important to note…”

The grandparent of AI throat-clearing.

163. “It’s worth noting…”

Same function, slightly softer.

164. “This does not necessarily mean…”

The model anticipates the strongest possible overinterpretation.

165. “That doesn’t mean X is wrong.”

A concession designed to keep both branches alive.

166. “May,” “might,” “can,” “could,” “often,” “typically”

Necessary epistemic tools. Suspicious when packed into every sentence.

167. Recursive caveats

A claim is qualified, then the qualification is qualified.

“This may suggest X, although the effect can vary by context and should not necessarily be interpreted as…”

168. Safety-shaped universals

Especially in medical, legal, financial, or interpersonal topics, the prose can accumulate procedural warnings until the original answer is buried.

169. The disclaimer moat

A simple recommendation is surrounded by enough caveats to prevent almost any interpretation.

170. “Consult a professional” as ritual closure

Sometimes necessary. Sometimes inserted without tailoring.

IX. Reasoning templates that feel machine-made

This is where LLM-isms become more interesting than vocabulary.

171. False dichotomy → synthesis

The model frames two positions, rejects choosing between them, then combines them.

“The choice is not speed or quality. The best systems achieve both through…”

172. Spectrumification

Binary questions become continua.

“Rather than treating expertise as something you either have or lack, it is more useful to think of it as a spectrum.”

Often sensible. Extremely reusable.

173. Trade-offification

Disagreement becomes optimization.

“The real issue is the trade-off between flexibility and control.”

174. Frameworkification

A messy phenomenon becomes three axes.

175. Category inflation

A small set of observations becomes a taxonomy.

A taxonomy becomes a framework.

A framework becomes a model.

176. Abstraction ratchet

The prose gradually moves upward:

specific event → pattern → principle → human condition.

177. The neat causal chain

A causes B, which drives C, which ultimately produces D.

Real systems usually contain loops, delays, confounders, and failures of causation.

178. The three-factor diagnosis

“This usually comes down to three things: incentives, information, and coordination.”

Why three? Because three feels explanatory.

179. Vague causality

“This can lead to…”

“This contributes to…”

“This creates a dynamic where…”

The relationship sounds causal without specifying magnitude or mechanism.

180. Vague attribution

“Experts argue…”

“Research suggests…”

“Many scholars believe…”

Without naming the experts, research, or scholars.

181. Citation aura

A claim sounds sourced because it is written in academic register even when no source is provided.

182. Example-as-evidence slippage

The model gives a plausible example immediately after a general claim, creating the sensation that the example demonstrates the claim.

183. Plausible specificity

A model can generate realistic details—job titles, implementation steps, timelines, institutional behaviors—that feel empirical even when they were inferred.

This is one of the more dangerous forms of polished prose.

184. Objection inoculation

The model anticipates one obvious objection, answers it, and thereby makes the broader argument feel more battle-tested than it is.

185. Symmetrical steelmanning

Both sides receive their strongest respectable version.

This is admirable when deliberate. It becomes a tell when every disagreement receives the same treatment.

186. The universal synthesis

Two theories are presented as different levels of explanation rather than genuine competitors.

187. The “both are true at different scales” escape hatch

A particularly elegant way to avoid choosing.

188. The unnecessary framework

The user asks a practical question. The model first explains a conceptual model of the problem.

189. The explanation-before-answer inversion

A response spends six paragraphs building context before stating the obvious recommendation.

190. The answer-before-explanation template

The inverse pattern is also common:

“Yes. Here’s why.”

Modern models have been trained toward both extremes depending on product style.

X. Synthetic profundity

One of the funniest LLM habits is its ability to make a refrigerator warranty sound metaphysical.

191. Significance inflation

A useful observation becomes a turning point.

192. “This changes everything.”

It usually does not.

193. “The deeper point…”

The model announces a deeper layer, then generalizes.

194. “What this really reveals…”

Interpretive altitude increases.

195. “The implications are profound.”

A sentence that should trigger a mandatory evidence check.

196. “In a very real sense…”

Frequently followed by a metaphor.

197. The humanity ending

“Ultimately, the question is what kind of future we want to build.”

198. The identity ending

A technical issue becomes a claim about who we are.

199. The agency ending

“The technology will not decide this. We will.”

Negative parallelism and uplift in one compact package.

200. The horizon ending

“We are only beginning to understand what becomes possible.”

The paragraph stares nobly toward sunrise.

XI. Generic example syndrome

201. The frictionless fictional company

“Imagine a mid-sized retailer struggling with inventory forecasting…”

The company has exactly the problem needed to demonstrate the concept and no irrelevant characteristics.

202. The suspiciously pedagogical person

“Consider Sarah, a project manager…”

Sarah exists to experience one variable at a time.

203. Safe canonical domains

Models repeatedly reach for:

  • healthcare
  • education
  • finance
  • retail
  • manufacturing
  • customer service
  • software development

because these domains furnish easy examples.

204. Evenly distributed examples

A human expert may have five examples from one obscure corner they know intimately.

A model often gives one from healthcare, one from finance, one from education.

205. Low sensory resolution

Generic examples contain roles and actions but few smells, textures, annoyances, odd objects, physical layouts, or socially awkward details.

206. No irrelevant specificity

Human memory contains junk.

The printer was beige. The meeting room was too cold. Dave kept eating sunflower seeds. None of this matters, which is precisely why it can make a story feel human.

207. Perfect relevance

Every example supports the thesis.

Reality rarely provides such obedient anecdotes.

208. Fabricated-realistic case-study voice

When asked for fictional examples, models are remarkably good at producing cases that read like anonymized consulting reports.

Useful, but stylistically distinct.

XII. Information-density tells

Research comparing human and LLM writing has found that instruction-tuned models can exhibit a more noun-heavy, informationally dense style and can struggle to reproduce the genre variation found in human writing.

That observation corresponds to several everyday tells.

209. Noun pileups

“organizational knowledge-management workflow optimization”

210. Nominalization

Verbs become nouns:

  • decide → decision-making
  • coordinate → coordination
  • implement → implementation
  • regulate → regulation

211. Abstract-subject sentences

“This approach enables greater alignment.”

“The framework provides clarity.”

“The process facilitates collaboration.”

Humans often say who did what.

212. Agent deletion

Who made the decision? Who benefits? Who pays?

The prose may not say.

213. Information compression without lived texture

A paragraph contains many concepts but no scene.

214. Adjective-modified noun phrases

“a flexible, scalable, user-centered governance framework”

215. High lexical polish, low personal grain

The vocabulary is varied but the voice is strangely unowned.

XIII. The model’s relationship with uncertainty

216. Probabilistic confidence without calibration

“It is likely that…”

How likely? The prose rarely says.

217. Confidence smoothing

Strongly known and weakly inferred claims may arrive in nearly identical tone.

218. Uncertainty laundering

A speculative chain becomes increasingly declarative as the paragraph progresses.

219. The invisible source boundary

Retrieved fact, pretrained memory, inference, and stylistic completion can blend into one seamless voice.

220. Plausibility preference

When evidence is missing, models are often excellent at generating what would make sense.

This is precisely why fluent errors are dangerous.

XIV. AI-polish artifacts in human-edited prose

Many contemporary texts are neither “human” nor “AI.” They are collaborative.

A person writes something rough, asks a model to improve it, and accepts part of the revision.

That produces its own family of tells.

221. Local register spikes

One paragraph suddenly becomes much more polished than its neighbors.

222. Vocabulary grafts

A writer who normally says “use” suddenly “leverages” something.

223. Smooth paragraph, rough document

Individual paragraphs are elegant while the larger argument remains disorganized.

224. The polished-email transformation

A three-sentence blunt email becomes seven sentences of gratitude, context, diplomacy, and “moving forward.”

225. Personality dilution

The grammar improves and the writer disappears.

226. Conflict deodorization

A genuinely irritated message becomes “I wanted to follow up regarding a concern.”

227. Accidental prestige inflation

Simple claims acquire academic or managerial vocabulary.

228. Tone homogenization

Different people in an organization begin sounding faintly alike because they use the same assistants.

229. AI-to-human feedback loops

This is the strangest stage: humans begin learning the model’s style, and the model is later trained on human writing that already contains model-influenced language.

The distinction between “AI language” and “human language” starts to blur.

Researchers have already described evidence consistent with this kind of coevolution in academic writing.

XV. Things people mistake for AI tells that are not very good tells

A field guide also needs anti-patterns: clues that are too weak to mean much.

230. Em dashes

Humans used them for centuries before ChatGPT.

The better question is whether the pattern of use feels model-like.

231. Semicolons

Same problem.

232. Correct grammar

Many humans write correctly.

233. Long words

Not evidence.

234. Markdown

A technical writer using headings and bullets is not suspicious merely because chatbots also do.

235. A single word like “delve”

Corpus-level overrepresentation is not individual-level proof.

236. No typos

Edited human prose exists.

237. Formal tone

Lawyers, academics, bureaucrats, consultants, and people writing cover letters have been doing this for a long time.

238. Repetition

Humans repeat themselves constantly.

239. Formulaic writing

Corporate communication, academic abstracts, SEO content, press releases, grant proposals, and school essays were formulaic before generative AI.

240. “Sounds like AI” as an accusation

This has become dangerous because many people now associate ordinary competent prose with AI simply because LLMs were trained on competent prose.

A style judgment is not authorship proof.

XVI. How to read for clusters instead of magic words

Suppose you encounter this paragraph:

In today’s rapidly evolving technological landscape, organizations must navigate an increasingly complex set of challenges. It is not merely about adopting new tools; it is about fostering a culture of innovation. By leveraging robust, scalable solutions, leaders can unlock new opportunities, streamline workflows, and empower teams. Ultimately, the organizations that thrive will be those that view AI not as a replacement for human potential, but as a catalyst for amplifying it.

No individual phrase proves anything.

But look at the cluster:

  • “rapidly evolving”
  • “landscape”
  • “navigate”
  • “not merely X; it is Y”
  • “fostering”
  • “leveraging”
  • “robust, scalable”
  • “unlock”
  • rule-of-three list
  • “empower”
  • “Ultimately”
  • second negative-parallelism construction
  • uplift ending

That is an LLM register.

The signal is combinatorial.

A useful mental model is to inspect five dimensions:

  1. Lexicon — Are AI-associated prestige words clustering?
  2. Syntax — Are the same contrast and parallelism templates recurring?
  3. Structure — Does every paragraph resolve with suspicious neatness?
  4. Tone — Is the voice uniformly reasonable, helpful, and abstract?
  5. Texture — Is there any idiosyncrasy, lived specificity, asymmetry, or unresolved mess?

The fifth category may matter most.

XVII. How to remove the AI smell without making writing worse

The goal should not be to make prose clumsy.

Many LLM habits became common because they are useful. Headings help. Parallelism can be beautiful. Caveats can make claims more accurate. Lists can clarify. Definitions can teach.

The problem is over-regularization.

1. Delete self-announcing transitions

Instead of:

“Here’s where it gets interesting.”

Say the interesting thing.

2. Replace prestige verbs with ordinary verbs

  • utilize → use
  • leverage → use
  • facilitate → help
  • showcase → show
  • foster → encourage, build, or create

Not always. Just when the fancy verb adds nothing.

3. Break symmetry

If two sides do not deserve equal space, do not give them equal space.

4. Stop forcing everything into threes

Sometimes there are two reasons.

Sometimes there are eleven.

5. Allow paragraphs to end without a moral

A paragraph may simply stop when the thought is complete.

6. Keep strange details

Specificity is often where personality lives.

7. Name agents

Replace “the organization decided” with who actually decided when you know.

8. Admit disproportionate interest

Human writers care unevenly.

If one tiny aspect fascinates you, follow it.

9. Do not automatically reconcile contradictions

Sometimes two things genuinely conflict.

10. Prefer observations to frameworks

You do not need to name every cluster of ideas.

11. Cut the second conclusion

If the point landed, trust it.

12. Preserve private vocabulary

The phrases you repeatedly use, the odd metaphors you invent, the jokes only you would make—these are not defects to sand away.

13. Vary sentence pressure

A short sentence can be enough.

And sometimes a sentence should keep going longer than a style guide would probably recommend because the writer is still following the thought and has not yet decided where it ends.

14. Let uncertainty remain ugly

“I don’t know” is sometimes better than a polished spectrum of possibilities.

15. Stop trying to sound like “good writing”

This may be the hardest one.

Models have absorbed enormous amounts of edited prose. When asked to improve writing, they tend toward generalized signals of competence: clarity, balance, polish, explicit transitions, formal vocabulary, rhetorical completion.

But voice often comes from violating those averages.

XVIII. The ultimate LLM-ism: statistical good behavior

The individual tics are entertaining.

Delve.

Nuanced.

It’s not X, but Y.

The relentless em dash. The triplets. The headings. The little bow tied around the end of every paragraph.

But those are surface phenomena.

The deeper characteristic is that a language model is, in a very literal sense, built to produce a highly probable continuation conditioned on what came before. Modern training adds instruction following, preference optimization, safety shaping, tool use, and reasoning strategies, but the output still emerges from a system optimized to generate plausible language.

That gives model prose a peculiar center of gravity.

It tends toward sentences that belong.

Words are appropriate. Transitions are appropriate. Caveats are appropriate. Examples are appropriate. Conclusions arrive where conclusions are supposed to arrive.

Human writing often contains something else: a sentence that should probably have been cut but somehow makes the whole essay better. A weird analogy. An unexplained grudge. A paragraph twice as long as its neighbors because the author got obsessed. A word used three times because the writer likes it. A joke that half the audience will miss. A sudden confession. An unresolved contradiction. A boring detail that turns out to be unforgettable.

Call it texture, idiolect, grain, voice, or simply human mess.

That is what generic LLM prose tends to sand away.

And this creates a strange paradox.

For decades, writing instruction taught people to remove mess: organize the argument, smooth the transitions, vary the vocabulary, avoid repetition, acknowledge counterarguments, use topic sentences, summarize clearly, sound professional.

Then we built machines that became extraordinarily good at exactly that bundle of behaviors.

Now the writing that feels most unmistakably human may be the writing that retains some friction.

Not bad writing.

Not random mistakes inserted to defeat a detector.

Friction.

Preference.

Specificity.

Asymmetry.

A sense that there is an actual mind behind the prose that cares more about some things than others.

A model can imitate those things when asked.

But without strong stylistic direction, it tends to return toward the center.

And once you learn to hear that center, it becomes difficult not to hear it everywhere.

XIX. Pocket dictionary: rapid-reference LLM-isms

For quick scanning, here is the compressed version.

Lexical

delve · intricate · nuanced · multifaceted · landscape · realm · tapestry · ecosystem · pivotal · crucial · vital · underscore · highlight · showcase · leverage · harness · foster · robust · seamless · streamline · optimize · dynamic · evolving · transformative · testament · journey · navigate · resonate · meaningful · compelling · notably · importantly · fundamentally · underpin · align · facilitate · elevate · empower · unlock · unpack · lens · interplay · intersection · paradigm · framework · holistic

Sentence shapes

  • “It’s not X; it’s Y.”
  • “Not merely X, but Y.”
  • “Not just X — Y.”
  • “Less about X, more about Y.”
  • “The real question is…”
  • “The result? X.”
  • “Here’s where it gets interesting.”
  • “And that matters.”
  • “A useful way to think about it…”
  • “The key distinction…”
  • “At a high level…”
  • “In practice…”
  • “In other words…”
  • “That said…”
  • “To be clear…”
  • “While X, Y…”
  • “Where X does A, Y does B.”
  • “Both things can be true.”
  • “These are not mutually exclusive.”
  • “Rather than X, think Y.”
  • “From X to Y.”

Structural

  • one idea per paragraph
  • mandatory topic sentences
  • mandatory landing sentences
  • repeated mini-summaries
  • neat three-part taxonomies
  • symmetrical comparisons
  • perfectly balanced objections
  • context → explanation → implication loops
  • list preambles and postambles
  • conclusion restating introduction
  • zoom-out-to-humanity ending

Conversational

  • “Great question.”
  • “Exactly.”
  • “You’re absolutely right.”
  • “Let’s unpack that.”
  • “Let’s break it down.”
  • “Here’s a clean way to think about it.”
  • “The short answer…”
  • “If you’d like…”
  • automatic follow-up menus
  • preemptive reassurance
  • excessive validation

Reasoning

  • false dichotomy followed by synthesis
  • turning binaries into spectra
  • turning disagreements into trade-offs
  • turning observations into frameworks
  • generic three-factor explanations
  • vague causal verbs
  • vague attribution
  • plausible examples treated like evidence
  • automatic counterargument handling
  • universal synthesis
  • abstraction ratchet

Formatting

  • heading cascades
  • bold-label bullets
  • bolded “key” sentences
  • blockquotes for invented summaries
  • horizontal rules between every section
  • emoji navigation
  • automatic tables
  • title-colon-subtitle constructions

XX. Sources and further reading

A few useful starting points for the empirical side of this topic:

The research should temper the game of spotting AI-isms. Models change. Prompting changes. Humans edit. Humans imitate machines. Machines imitate humans. Particular words go in and out of fashion.

The target therefore cannot be a frozen blacklist.

The better object of study is the distribution of habits.

What does the prose repeatedly prefer?

Where does it place emphasis?

How often does it create symmetry?

How aggressively does it organize?

How much unresolved weirdness survives?

Those questions are harder than searching for delve.

They are also much more interesting.