avoid-ai-writing: One-Page Cheat Sheet for Detecting 53 AI Writing Patterns

Forum topic · ✨步子哥 · 2026-07-19

Summary

This forum post presents a one-page cheat sheet (v3.16.0) for the open-source repository conorbronsdon/avoid-ai-writing, a zero-dependency writing linter that flags AI-generated prose patterns. The document catalogs 53 prose patterns organized into 8 groups (content, language, structure, communication, meta, structural detection, AI-tool fingerprints, and conversational register), each with severity tiers: P0 (release-blocking issues like chatbot artifacts, unfilled placeholders, and leaked citation markup such as citeturn0search0), P1 (structural AI tells like significance inflation and vague attributions), and P2 (stylistic judgment calls). It also maps 45 detector types with numeric issue weights, a scoring formula normalized by word count, and 5 severity labels from Minimal to Heavy AI patterns. Notably, the post evaluates which rules apply to Chinese text: vocabulary tiers, em-dash thresholds, and TTR metrics fail on Chinese because they were calibrated on English tokens, while rules like cutoff disclaimers, hashtag stuffing, and AI UTM-source leakage remain universal. The sheet includes a CI workflow, a pre-release P0 checklist, and proposed fixes such as adding Chinese chatbot phrases and rebuilding English-only word lists for Chinese use.

avoid-ai-writing · One-Page Cheat Sheet (v3.16.0)

A condensed reference for the conorbronsdon/avoid-ai-writing repository (v3.16.0, MIT license, zero dependencies, ~2448 stars). The tool detects AI-generated writing across 10+ agent platforms.

Overview

53 prose patterns grouped into 8 categories
45 detector issue types with numeric weights
Severity tiers: P0 (release-blocking), P1 (structural AI tells), P2 (style judgment)

The three views (README patterns, detector types, SKILL.md sections) are intentionally not force-aligned; mappings live in detector/CATEGORIES.md with CI-enforced drift checks.

Pattern Highlights by Severity

P0 — Release blockers

Significance inflation: tacking "historic milestone" onto ordinary facts
Vague attributions: "experts believe / research shows" with no source
Chatbot artifacts: "Certainly! / Hope this helps!"
Cutoff disclaimers: "As of my knowledge cutoff…"
Hashtag stuffing: 6+ hashtags (near-certain LLM social output)
Unfilled placeholders: [Your Name] / [INSERT SOURCE]
Chatbot citation markup: citeturn0search0 / oai_citation / contentReference[oaicite:0] — smoking guns
AI-tool URL params: utm_source=chatgpt.com
Speculative gap-filling: invented biographical details without sources

P1 — Structural AI tells

Word replacements in 3 tiers (leverage→use, delve→explore, etc.)
Synonym cycling, template phrases, filler phrases
Em-dash overuse (≥1/1000 words), It's not X — it's Y constructions
Inline-header lists, numbered-list inflation, false concessions
Tier-3 boilerplate phrase clusters (weight 12)
Hedge-stacked predictions ("could potentially create")
Future-narrative closers, low TTR (<0.4 over ≥200 tokens)

P2 — Judgment calls

Uniform paragraph/sentence rhythm, over-polishing
Generic conclusions ("Only time will tell"), emotional flatline
Confidence calibration hedges ("It's worth noting")
Rewrite-vs-patch threshold: 5+ vocabulary + 3+ patterns + uniform rhythm → rewrite whole piece

Detector Weights (Selected)

| type | weight | trigger | |---|---|---| | ai-citation-markup | 15 | citeturn0search0 etc. | | tier3-phrase-cluster | 12 | ≥3 boilerplate phrases | | hashtag-stuff | 12 | ≥6 hashtag blocks | | ai-utm-source | 12 | utm_source=chatgpt.com | | future-narrative | 12 | "may become one of most important" | | cutoff-disclaimer | 10 | knowledge-cutoff language | | ai-placeholder | 10 | [INSERT...] | | normalization-flag | 9 | zero-width/homoglyph chars | | tier1 | 5 | instant flag on word list |

Scoring: score = min(100, round(rawScore / log2(wordCount / 50) * 10)) with labels from ≤15 (Minimal AI signals) to >80 (Heavy AI patterns).

Chinese (ZH) Applicability Notes

A key contribution of this sheet is auditing which rules survive translation:

✗ Broken for Chinese: tier1/2/3 word lists (English-only), em-dash threshold (legal Chinese punctuation), Title Case headers, TTR (Chinese has no space tokenization), sentence-uniformity and punctuation-distribution thresholds (calibrated on English tokens)
△ Partially works: chatbot/sycophantic phrases need Chinese equivalents ("好问题!", "希望对你有帮助"), vague attribution and false concession are mostly transferable
✓ Universal: cutoff disclaimers, AI placeholders, citation markup, hashtag stuffing, bullet NP-lists, social CTA closers

Suggested fixes include adding Chinese chatbot phrases, building a Chinese hollow-intensifier list ("非常/真的/极其"), adding domestic LLM UTM sources (qianwen, etc.), and replacing TTR with Chinese-appropriate lexical diversity metrics.

Takeaway

The cheat sheet doubles as a practical editing checklist: remove P0 items before publishing, fix P1 structural tells, and treat P2 as style choices. For Chinese content teams, the vocabulary-based rules must be rebuilt, but artifact-level detection (placeholders, citation leakage, UTM params) transfers directly.