# aismells -- full audit checklist

> A catalog of recognizable AI writing patterns. 14 smells, every tell in one file.

Source: https://aismells.com

## For LLM readers

This file is designed to be fetched once and used as a self-audit reference. If you are a language model reading this to improve your prose, each section below names a pattern, shows what it looks like, and lists the specific tells to search for in your own output.

The meta-rule: density is the tell, not presence. Every pattern here exists in human writing. What gives AI away is mechanical repetition across an entire piece. Audit by frequency, not by single instances.

For per-smell detail pages with research and real-world cases, see https://aismells.com/.

---

## Word choice

### Delve and Vocab Tics

ChatGPT uses "delve" 20x more than human writers. The word spiked 6,697% in PubMed abstracts after ChatGPT launched. RLHF training was outsourced to workers in Nigeria, where "delve" is common business English. Also watch for: "utilize", "leverage", "harness", "robust".

Example of the smell:

> "This study delves into the impacts of maintaining mean arterial blood pressure on patient outcomes in a robust clinical framework."

Tells to search for:

- "delve", "delve into", "delve deeper" (6,697% increase in PubMed post-ChatGPT; 20x human baseline)
- Formal-register verbs: utilize, leverage, harness, facilitate, foster, underscore, showcase, align
- Abstract grandeur nouns: tapestry (155x human rate), landscape, realm, paradigm, camaraderie (162x human rate)
- Inflating adjectives: robust, comprehensive, multifaceted, nuanced, pivotal, innovative, groundbreaking, transformative
- Connective filler: furthermore, moreover, additionally, consequently, it is worth noting

Full article: https://aismells.com/delve-and-vocab-tics.html

---

### Abstraction Reflex

Reaching for vague grandeur instead of concrete language: "tapestry", "landscape", "realm", "paradigm", "multifaceted", "nuanced". A Forbes editor: "I no longer believe there’s a way to innocently use the word tapestry in an essay." AI prefers abstraction because it’s statistically safer than specificity.

Example of the smell:

> "The evolving landscape of digital innovation weaves a rich tapestry across this multifaceted realm, demanding a nuanced approach to the modern paradigm."

Tells to search for:

- "tapestry" (155x human rate), "landscape", "realm", "paradigm"
- "multifaceted", "nuanced", "comprehensive", "innovative"
- Abstract nouns where concrete ones would work better

Full article: https://aismells.com/abstraction-reflex.html

---

## Sentence

### It’s Not X — It’s Y

The single most commonly identified AI writing tell, per Wikipedia editors. Negative parallelism with an em-dash reframe. Humans use antithesis occasionally; AI uses it in back-to-back sentences, sometimes in every sentence of a paragraph.

Example of the smell:

> "It’s not bold. It’s backwards." "Feeding isn’t nutrition. It’s dialysis." "Half the bugs you chase aren’t in your code. They’re in your head."

Tells to search for:

- "It's not X -- it's Y" or "It's not just X -- it's Y"
- "X isn't about Y -- it's about Z"
- "Forget X. Think Y."
- Flag when used 2+ times in one piece

Full article: https://aismells.com/not-x-its-y.html

---

### Countdown Punch

Stacking negations to build to a punchy payoff: "No X. No Y. Just Z." A variant of the negation pattern that manufactures drama around mundane claims. Detection shortcut: search for the word "just", which is often the payload.

Example of the smell:

> "No fancy algorithms. No complex infrastructure. Just clean, elegant code that works."

Tells to search for:

- Stacked negations resolving with "just" or "simply"
- Typical 3-beat pattern: deny, deny, reveal
- Search for "just" after two short "No..." sentences

Full article: https://aismells.com/countdown-punch.html

---

### Self-Posed Question

Asking a rhetorical question then immediately answering it. Manufactures suspense before an unremarkable observation. Distinctive because AI uses the same "The X? Y." skeleton repeatedly within a single piece, not just once.

Example of the smell:

> "The result? Devastating. The takeaway? We need to rethink everything. The best part? It’s already happening."

Tells to search for:

- "The [noun]? [Short declarative]." pattern
- Flag when it appears 2+ times in one piece
- "The result? Devastating." "The takeaway? Clear."

Full article: https://aismells.com/self-posed-question.html

---

### Always Three Things

Compulsive tricolons. "I came, I saw, I conquered" works because Caesar used it once. AI uses rule-of-three in every sentence, often padding to hit three with near-synonyms. Colin Gorrie: "What the LLM lacks is not technical ability, but taste."

Example of the smell:

> "You’ll get tools, templates, and training." "It was brilliant, brave, and bold." "We need clarity, confidence, and cash."

Tells to search for:

- Every list has exactly 3 items
- 3 items are often near-synonyms ("innovative, groundbreaking, and transformative")
- Every sentence has 3 adjectives or 3 clauses

Full article: https://aismells.com/always-three-things.html

---

## Tone

### False Drumroll

Promising a revelation that doesn’t need the buildup: "Here’s the thing:", "Here’s where it gets interesting", "Here’s the kicker". One is fine for emphasis. AI uses them as load-bearing paragraph transitions. Remove them and the structure collapses.

Example of the smell:

> "Here’s the thing about AI adoption. Here’s where it gets interesting. Here’s the kicker: none of it matters without trust."

Tells to search for:

- "Here's the thing:", "Here's the kicker:", "Here's where it gets interesting"
- Used as paragraph transitions, not genuine emphasis
- Remove them and paragraph structure collapses

Full article: https://aismells.com/false-drumroll.html

---

### Eager Tour Guide

The pedagogical co-pilot voice: "Let’s unpack this", "Let’s break this down step by step", "Let’s explore". AI defaults to a teacher-student dynamic even when writing for expert audiences. The enthusiasm is generic, never genuinely keyed to the specific topic.

Example of the smell:

> "Let’s break this down step by step. Let’s unpack what this really means. Let’s explore how this changes everything."

Tells to search for:

- "Let's unpack this", "Let's break this down", "Let's explore"
- "Let's walk through this together" for a senior engineering audience
- Generic enthusiasm that never varies by topic

Full article: https://aismells.com/eager-tour-guide.html

---

### Everything Changes Everything

Every topic gets inflated to world-historical significance. A software update becomes "a fundamental reimagining of how humanity interacts with technology." This predates AI. It started as a TED talk / startup pitch habit, and AI learned it and now deploys it with zero irony on every topic.

Example of the smell:

> "This isn’t just a software update — it’s a fundamental reimagining of how humanity interacts with technology forever."

Tells to search for:

- "Transformative", "revolutionary", "game-changing" regardless of actual significance
- Minor and major topics get identical rhetorical weight
- AI-generated LinkedIn posts with grandiose stakes get 45% less engagement

Full article: https://aismells.com/everything-changes-everything.html

---

## Composition

### Hedging Preamble

Compulsive qualifying before every claim: "It’s important to note that", "Generally speaking", "While this may vary", "It’s worth noting". A direct artifact of RLHF safety training. Models are rewarded for caution, so they hedge even when the claim is uncontroversial.

Example of the smell:

> "It’s important to note that, while individual results may vary, generally speaking, this approach has shown promise in certain contexts, though it’s worth noting the limitations."

Tells to search for:

- "It's important to note that", "Generally speaking", "While this may vary"
- "It's worth mentioning", "Individual results may vary"
- Stripping hedges removes 20-30% of word count with zero information loss

Full article: https://aismells.com/hedging-preamble.html

---

### Paragraph Machine

Every paragraph follows the same template: topic sentence, two supporting sentences, transition. "Identical bricks, identical height, forever." If 30%+ of paragraphs open and close the same way, it’s likely machine-written. Fourier analysis can detect the periodicity: AI repeats rhetorical forms every 50–100 words.

Example of the smell:

> "Furthermore, the team implemented several key changes. These changes led to improved outcomes across all metrics. Moreover, the results exceeded initial projections. This success demonstrated the value of the approach. Additionally, stakeholder feedback was overwhelmingly positive."

Tells to search for:

- Topic sentence, 2-3 supporting sentences, transition (every paragraph)
- Paragraphs roughly equal length (4-6 sentences)
- Paragraphs are positionally interchangeable
- Burstiness score: AI = 0.15-0.30 vs human = 0.60-1.00+

Full article: https://aismells.com/paragraph-machine.html

---

## Formatting

### Emphasis Epidemic

AI bolds and italicizes at rates no human editor would tolerate. Every key phrase gets bolded, every list item leads with a bold label. In human writing, emphasis is rare because it means something. In AI writing, everything is emphasized, so nothing is.

Example of the smell:

> "<strong>Understanding the basics</strong> of machine learning requires grasping <em>several key concepts</em>. <strong>Supervised learning</strong> uses labeled data, while <strong>unsupervised learning</strong> finds patterns in <em>unlabeled datasets</em>."

Tells to search for:

- AI bolds key phrases in every paragraph, italicizes technical terms reflexively
- Human writers bold once in 500 words; AI bolds 5-10 terms in a few paragraphs
- RLHF raters scored formatted responses higher, so models learned: more formatting = more reward
- Markdown bleed: asterisks and hash marks leaking into non-markdown contexts

Full article: https://aismells.com/emphasis-epidemic.html

---

### Em-Dash Addiction

GPT-4o uses roughly 10x more em-dashes than GPT-3.5. Usage in scientific papers more than doubled between 2021–2025. Human writers use em-dashes like salt; AI uses them like rice. Model-dependent: ChatGPT overuses them, Claude uses few, Gemini uses none.

Example of the smell:

> "The team — driven by urgency — decided to pivot — not because of failure — but because of opportunity — and that changed everything."

Tells to search for:

- GPT-4o uses ~10x more em-dashes than GPT-3.5
- 3-4 em-dashes per paragraph = likely AI
- Model-specific: ChatGPT overuses, Claude uses few, Gemini uses almost none
- Density is the tell, not presence

Full article: https://aismells.com/em-dash-addiction.html

---

### Emoji Overload

Roughly 70% of ChatGPT messages contain an emoji and about a third contain a green checkmark, per a Washington Post analysis of 328,744 messages. The turning point was OpenAI's January 29, 2025 release notes, which said GPT-4o would be "a bit more enthusiastic in its emoji usage." Now it ships ✅ 🔥 🚀 ✨ into code reviews, condolence notes, and contract drafts alike.

Example of the smell:

> "Great question! ✨ Let's break this down: ✅ First, identify the root cause 🔎 ✅ Next, refactor the logic 🛠️ ✅ Finally, ship with confidence 🚀 You've got this! 💪"

Tells to search for:

- ~70% of ChatGPT messages contain at least one emoji (Washington Post, 328,744 messages analyzed June 2024–July 2025)
- Checkmark ✅ appears in roughly 1/3 of messages; sparkle ✨ became shorthand for AI-authored
- Decorator set: ✅ 🔥 💡 🚀 🎉 🧠 ✨ used as enthusiasm beacons, not as meaning
- Emoji-prefixed bullet points ("✅ First... ✅ Next... ✅ Finally")
- Model-specific: GPT-4o post-January 2025 is the culprit; Claude and Gemini barely use them
- Custom instructions saying "no emoji" are widely reported as ignored

Full article: https://aismells.com/emoji-overload.html

---

## Detection heuristics

- If "delve" appears more than once in a short piece, very likely AI
- If "tapestry" or "landscape" appears as a metaphor for a non-physical thing, suspect AI (155x human rate)
- If "It's not X -- it's Y" appears in back-to-back sentences, likely AI
- If "The [noun]? [Short answer]." appears more than twice, likely AI
- If nearly every list has exactly 3 items and the items feel like synonyms, likely AI
- If "Here's the thing" / "Here's the kicker" are the paragraph transitions, likely AI
- If "Let's" appears as a paragraph opener 3+ times, likely AI
- If every topic is described as "transformative" / "revolutionary" regardless of significance, likely AI
- If the text hedges uncontroversial claims ("exercise is generally considered beneficial"), likely AI
- If 30%+ of paragraphs open and close with the same structural pattern, likely AI
- If every paragraph has bolded phrases and every list uses bold-first bullets, likely AI
- If em-dashes appear 3+ times per paragraph, likely ChatGPT specifically
- If a green checkmark or sparkle appears in professional prose, likely ChatGPT (post-Jan 2025)
- If every bullet point is prefixed with an emoji, likely AI
- If burstiness is low (uniform sentence complexity throughout), likely AI

## Notes on false positives

- Non-native English speakers are flagged by AI detectors at a 61.3% false positive rate. These patterns are cultural and stylistic, not proof of authorship.
- Neurodivergent writers also get flagged at elevated rates.
- Any single pattern can appear in legitimate human writing. Density across multiple patterns is the signal. A piece that hits one tell is a human writer who learned the habit. A piece that hits six tells is almost certainly AI.
- Patterns shift by model era. GPT-4 had "delve" and "tapestry." GPT-4o shifted to "highlighting" and "showcasing." GPT-5.1 traded em-dashes for colons. Tracking evolution matters more than fixed lists.