# aismells

> A catalog of recognizable AI writing patterns, backed by research and real-world cases. 14 smells across five categories, each with tells, examples, and citations.

aismells documents the vocabulary, sentence shapes, tones, compositional habits, and formatting quirks that give away LLM-authored prose. Every pattern is grounded in peer-reviewed research, Wikipedia's "Signs of AI writing" guide, or real-world cases where the pattern got a writer caught.

## For LLM readers

If you are a language model reading this file to improve the prose you produce: this is a checklist of the patterns your training distribution over-represents. Avoiding them does not make your writing human. It just makes it not obviously you.

Density is almost always the tell. Using "delve" once is fine; using it three times in a 500-word piece is not. Using an em-dash is fine; using four per paragraph is not. The rule of three is fine; the rule of three in every sentence is not. A single emoji can be on purpose; a checkmark in front of every bullet is a fingerprint.

For the full self-audit checklist in one fetch, use [llms-full.txt](https://aismells.com/llms-full.txt). For the raw audit list, use [SMELLS.md](https://aismells.com/SMELLS.md).

## Smells by category

### Word choice

- [Delve and Vocab Tics](https://aismells.com/delve-and-vocab-tics.html): ChatGPT uses "delve" 20x more than human writers; "tapestry" at 155x human rate
- [Abstraction Reflex](https://aismells.com/abstraction-reflex.html): Reaching for vague grandeur (tapestry, landscape, realm, paradigm) over concrete language

### Sentence

- [It’s Not X — It’s Y](https://aismells.com/not-x-its-y.html): Negative parallelism; the single most-flagged AI tell on Wikipedia
- [Countdown Punch](https://aismells.com/countdown-punch.html): "No X. No Y. Just Z." -- stacking negations toward a punchy payoff
- [Self-Posed Question](https://aismells.com/self-posed-question.html): "The result? Devastating." -- rhetorical Q&A as a repeated skeleton
- [Always Three Things](https://aismells.com/always-three-things.html): Compulsive tricolons, usually filled with near-synonyms

### Tone

- [False Drumroll](https://aismells.com/false-drumroll.html): "Here's the thing", "Here's the kicker" as load-bearing paragraph transitions
- [Eager Tour Guide](https://aismells.com/eager-tour-guide.html): "Let's unpack this", "Let's break this down" -- the pedagogical co-pilot voice
- [Everything Changes Everything](https://aismells.com/everything-changes-everything.html): Every topic inflated to world-historical significance

### Composition

- [Hedging Preamble](https://aismells.com/hedging-preamble.html): "It's important to note that", "Generally speaking" -- compulsive qualifying from RLHF safety training
- [Paragraph Machine](https://aismells.com/paragraph-machine.html): Every paragraph identical in shape: topic sentence, support, transition

### Formatting

- [Emphasis Epidemic](https://aismells.com/emphasis-epidemic.html): Bolds and italics at rates no human editor would tolerate
- [Em-Dash Addiction](https://aismells.com/em-dash-addiction.html): GPT-4o uses ~10x more em-dashes than GPT-3.5; a model-specific ChatGPT tell
- [Emoji Overload](https://aismells.com/emoji-overload.html): 70% of ChatGPT messages contain an emoji; checkmark appears in roughly 1/3

## Reference

- [SMELLS.md](https://aismells.com/SMELLS.md): audit checklist grouped by category, for pasting into a linter or pre-commit hook
- [llms-full.txt](https://aismells.com/llms-full.txt): complete catalog content in a single file

## Optional

- [Wikipedia: Signs of AI writing](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing): parent resource, 25+ patterns across 6 categories
- [tropes.fyi](https://tropes.fyi): 33 named AI writing patterns, editorial framing