Inside a 2,000-Line Rule System for Removing AI Writing Tells: avoid-ai-writing Breakdown

Forum topic · 小凯 · 2026-05-24

Summary

A deep-dive review of avoid-ai-writing, an MIT-licensed open-source writing skill by Conor Bronsdon (v3.4.0, github.com/conorbronsdon/avoid-ai-writing) that uses roughly 2,000 lines of rules to detect and rewrite AI-flavored prose. The system uses a three-tier vocabulary gradient (always-replace words, cluster-triggered words, density-triggered words) instead of a flat blacklist, six context profiles (blog, technical-blog, investor-email, linkedin, docs, casual) with adjustable strictness, and a taxonomy of 42 AI writing tells spanning formatting, sentence patterns, structure, rhetoric, metadata leaks, and rhythm. Its most counterintuitive rule warns that over-polishing human text can push it toward AI statistical profiles, since AI output is statistically over-smooth. The author explicitly positions it as a writing-quality signal, not a verdict, citing research showing AI detectors' high false-positive rates on non-native English writers. With 1,400 GitHub stars from a single markdown file, the project reflects real demand for writing that sounds human rather than statistically average.

> Author: Conor Bronsdon > Repo: https://github.com/conorbronsdon/avoid-ai-writing > Version: v3.4.0, MIT License > Date: 2026-05-24

Consider this paragraph:

> Certainly! Acme Analytics, a vibrant startup nestled in the heart of Boulder's thriving tech ecosystem, has secured 💲40M in Series B funding — marking a watershed moment for the observability landscape. The platform serves as a unified hub, featuring real-time dashboards, boasting sub-second queries, and presenting a seamless integration layer. Moreover, experts believe Acme is poised to disrupt the market. In conclusion, the future looks bright!

This is the standard demo from the avoid-ai-writing repository. In just one paragraph, the author flags more than fifteen AI tells: chatbot-style openers ("Certainly!"), promotional filler ("vibrant," "nestled," "thriving"), meaning inflation ("watershed moment"), copula avoidance ("serves as," "featuring," "boasting"), hollow appeals to authority ("experts believe"), connective padding ("Moreover"), and a cliché closer ("the future looks bright").

The rewrite is a single sentence:

> Acme Analytics raised a 💲40M Series B led by Sequoia. The Boulder-based startup makes an observability platform that runs queries in under a second and plugs into existing monitoring stacks without custom integration work.

The information content is nearly identical. But the first reads like an auto-generated press release; the second reads like a person wrote it.

That is what avoid-ai-writing does. It is not word-swapping — it pulls AI's statistical average back toward concrete human expression.

1. A Three-Tier Vocabulary Gradient, Not a Blacklist

The obvious assumption is that this is just a banned-word list: replace "leverage" with "use," "delve into" with "explore." Author Conor Bronsdon clearly anticipated this. He splits vocabulary into three tiers:

Tier 1 — replace on sight. Words appearing 5–20x more often in AI text than human text: "tapestry," "realm," "paradigm," "beacon," "testament to," "game-changer," "watershed moment," "only time will tell" — nearly fifty entries. Not bad words, just over-sampled in LLM output distributions.
Tier 2 — flagged only in clusters. Fine alone; two or more in one paragraph triggers a rewrite: "harness," "navigate," "foster," "elevate," "streamline," "empower." Common in business English, but AI stacks them together.
Tier 3 — flagged only by density. "Significant," "innovative," "compelling," "unprecedented," "sophisticated" are perfectly normal — flagged only when overall density exceeds a threshold (around 3%). AI uses these as filler; humans use them with specific referents.

This layered design turns detection into a probabilistic question rather than a binary judgment, based on context, frequency, and co-occurrence.

2. Six Context Profiles: AI-Tell Tolerance Is Adjustable

The most interesting design is the context profile system — the same rules at different strictness per scenario:

| Context | Strictness | |---------|-----------| | blog (default) | Fully on | | technical-blog | Some Tier 2 words relaxed | | investor-email | Extra strict on promo language and meaning inflation | | linkedin | Formatting relaxed, vocabulary strict | | docs | Loosened overall; clarity first | | casual | Only P0 (credibility-killing) issues |

For example, "robust," "comprehensive," and "seamless" are legitimate technical terms in a technical blog. But in an investor email, "a robust ecosystem poised for transformative growth" is textbook promo-speak and gets extra scrutiny.

The deeper insight: AI tells are not absolute — they are deviations from context-dependent expectations. Bullets and emoji are normal on LinkedIn but a red flag in deep analysis.

3. A Morphology of 42 AI Tells

The vocabulary list is only part of the skill. The author catalogs 42 AI writing tells:

Formatting: dash overuse (more than one per thousand words), excessive bolding, emoji in headings, bullet-list overload that fragments what should be paragraphs.
Sentence patterns: the "It's not X, it's Y" false-dichotomy construction, hollow intensifiers ("genuine," "truly," "quite frankly"), vague endorsements ("worth reading"), stacked hedging ("could potentially," "may eventually").
Structure: metronomically uniform paragraph lengths, grand-background openings ("In the rapidly evolving world of..."), symmetric triads, listicle inflation ("Here are 7 reasons why...").
Rhetoric: false concessions ("While X is impressive, Y remains a challenge" — both sides empty), rhetorical-question stalling ("But what does this mean for developers?"), parenthetical hedges ("(and, increasingly, Z)").
Metadata: chatbot politeness leakage ("I hope this helps!", "Great question!"), AI tool tracking parameters ("utm_source=chatgpt.com"), internal citation marker leaks ("citeturn0search0", "[attached_file:1]").
Rhythm: the author stresses that structural regularity is the number-one detection signal. Even with all Tier 1 words swapped, uniform sentence lengths, uniform paragraphs, and formulaic transitions still read as AI. Pangram — a classifier trained on 28 million human documents — ranks structural regularity above vocabulary.

4. The Most Underrated Rule: Don't Over-Polish

The skill's final rule stands out:

> Aggressively editing out every irregularity can push human writing *toward* AI statistical profiles. Natural disfluency, idiosyncratic word choices, and uneven pacing are what keep text out of the "AI-generated" classification.

In other words: run a human article through these rules too aggressively, smooth every irregularity, and it starts resembling AI output — because AI output is statistically over-smooth: uniform sentence lengths, flawless grammar, template transitions.

Human writing stumbles, repeats, throws in abrupt short sentences and deliberately incomplete phrasing. Conventional editing treats those as problems; here they are precisely the source of a human voice. The author is not chasing "perfect writing" — he is chasing "writing that reads as human." Those are different standards.

5. Honest About Limitations

The skill opens with these numbers:

Stanford research (Liang et al., 2023): commercial AI detectors show over 60% false-positive rates against non-native English writers.
BFI working paper (Jabarian & Imas, 2025): open-source detectors misclassify over 70% of the time.
arXiv:2506.07001 (2025): adversarial paraphrasing can cut detection accuracy by roughly 88%.

Then: "This is a writing-quality tool, not a verdict."

He doesn't pretend the rules can catch all AI writing or be used to accuse anyone. It is a signal, not evidence — worth acting on for your own text, not worth ruining someone's reputation over. That honesty upgrades the tool from "detector" to "writing companion."

6. Why AI Flavor Is Statistical Averaging

Think of it Feynman-style: an LLM is a probability distribution. At every position it picks the most likely next token. As billions of parameters converge in the same direction, output text statistically approaches an "average human" — error-free, unstartling, unopinionated — and thereby loses every trait of a specific person.

"Leverage" appears more often than "use" in business training data, so LLMs prefer it. "It's not X, it's Y" appears in many successful essays, so LLMs copy it. Triadic parallelism is taught in rhetoric textbooks, so LLMs default to it.

The 2,000 lines of rules do one thing: pull the LLM's statistical average back into concrete context. Not "use is better than leverage" — in this specific sentence, for this specific author and context, "use" is the more accurate choice.

This is why the skill emphasizes context profiles and a second audit pass. After a first rewrite, AI may introduce fresh averaging — new words, same structural formulas. The second pass checks whether statistical patterns have re-emerged in disguise.

7. Closing Thoughts

Bronsdon released the repo two months ago; it now has 1,400 stars and 142 forks. For a single markdown file with no code, no dependencies, and no API, that velocity says the demand is real.

People aren't looking for "better AI writing" — they're looking for "writing that sounds like a person wrote it."

The real value isn't in the 109 replacement entries but in the systematic framework: which dimensions AI tells live on, how tolerance shifts by context, and most importantly — the pursuit of perfection is itself an AI tell.

Whether you write with AI assistance or entirely by hand, this skill is worth one run in detect mode. Not to "prove" anything, but to see how much of your prose is residual statistical average — and how much is an actual voice.

> A signal, not proof. Worth editing by; not worth accusing by.

References and further reading

Conor Bronsdon, avoid-ai-writing (v3.4.0, MIT License): https://github.com/conorbronsdon/avoid-ai-writing
Liang et al., "GPT detectors are biased against non-native English writers" (Patterns, 2023)
Jabarian & Imas, BFI Working Paper 2025-116
arXiv:2506.07001, adversarial paraphrase against AI detection