Overview
The zhichai.net forum post introduces lieflat-less-ai-tone, an open-source linguistics study published on GitHub (repo) that empirically tests which stylistic features actually distinguish AI-generated Chinese text from human writing.
The study built a large parallel corpus:
- 629 articles
- 2,826,972 Chinese characters
- ~95,000 sentences
- ~45,000 paragraphs
- Five models compared: Claude, DeepSeek, Gemini, GPT, Kimi
- 26 stylistic features measured, yielding 11 statistically significant discriminative markers
- DeepSeek & Claude: em-dash overuse, heavy lead-in colons
- Gemini: question-form section headings
- GPT: heavy "not A but B" constructions, almost no em-dashes
- Kimi: dense ordinal headings (first, second, third)
- Restore referents: "Notably, this brought huge changes" → "This architecture change brought huge changes"
- Concretize metaphors: "like a guide versed in past and present" → "like a lathe operator who has repaired engines for 30 years"
- Restore real numbers: "dramatic performance leap" → "throughput up 2.59x, latency down 42%"
- Larashero & Moxt Open-Source Community (2026). *Lieflat Less AI Tone*. GitHub:
https://github.com/larashero3-dotcom/lieflat-less-ai-tone - Biber, D. (1988). *Variation across speech and writing*. Cambridge University Press.
Key Findings: Popular Beliefs vs. Measured Reality
| Feature | Popular belief | Measured frequency ratio R (AI ÷ human) | Reality | | :--- | :--- | :---: | :--- | | Zero-anaphora comment at paragraph opening | Rarely noticed | 4.4x | AI opens paragraphs with evaluative remarks lacking referents (e.g., "Notably, ..." with no antecedent) | | Fake anthropomorphic metaphors | Thought humans do this more | 7.3x | AI favors hollow personas ("like a wise mentor") vs. humans' concrete ones ("like a mechanic who's fixed engines for 30 years") | | Colon-led filler lists | Seen as normal formatting | 9.4x | Information-free lead-in colon followed by bullet-point dumps | | Contrastive structure ("not A but B") | Widely perceived | 3.4x | Confirmed: AI manufactures false oppositions | | Concrete-number density | Assumed AI is good with numbers | 0.35x | Humans use 3x more concrete figures; AI prefers vague terms like "significantly improved" | | Rhetorical questions in body text | Believed AI-specific | 0.06x | Humans ask 17x more; deleting them makes text stiffer | | Overall metaphor frequency | Believed AI-specific | 0.42x | Humans use 2.4x more metaphors | | Uniform sentence lengths | Believed AI-specific | No significant difference (R=0.87) | Disappears after proper segmentation correction |
Here, R ≥ 2.0 indicates an AI-heavy feature; R < 0.8 indicates a human-preferred feature.
No Single "AI Tone"
Style quirks vary dramatically across models — e.g., em-dash frequency differs by 46.9x between DeepSeek (5.16 per thousand characters) and GPT (0.11). Reported per-model patterns:
Consequence: detection/editing rules built on a single model's quirks fail immediately on other models.
The De-AI-Toning Rules (SKILL.md)
Three surgical principles:
1. Whitelist: touch only the 11 empirically validated features; never edit unmatched sentences. 2. Information conservation: every content word after rewriting must trace back to the source; no invented facts or numbers. 3. No subjective semantics: only morphological/structural restructuring, no gut-feeling deletion.
Example fixes:
Scientific Integrity
The authors openly report 6 systematic regex-related errors corrected during the study — e.g., an early failure to strip Markdown tables led to an absurd "AI sentence-length uniformity 51x human" claim, which was retracted after fixing the denominator and regex boundaries.
Conclusion
AI text feels "fake" not because of any specific rhetorical device but because it lacks grounded, concrete physical detail. Removing the "AI tone" means restoring specificity, precision, and referential clarity — not following a checklist of surface tics.