What This Study Did
An open-source research project, lieflat-less-ai-tone, published on GitHub by Larashero and the Moxt open-source community (2026), set out to test whether the widely believed markers of "AI tone" actually hold up statistically.
The controlled corpus:
- 629 articles
- 2,826,972 Chinese characters
- 95,000 sentences
- 45,000 paragraphs
- Five frontier models compared against human writers: Claude, DeepSeek, Gemini, GPT, Kimi
- 26 stylistic features quantitatively screened
- DeepSeek & Claude: dash overuse (4–5 per 1,000 chars) plus suggestive colons
- Gemini: prefers question-form section headings
- GPT: heavy "not A but B" constructions, but almost no dashes
- Kimi: dense ordinal headings (First, Second, Third...)
Popular Myths vs. Measured Facts
| Feature | Popular belief | Frequency ratio R (AI ÷ human) | Verdict | | :--- | :--- | :---: | :--- | | Paragraph-opening zero-anaphora comments | Rarely noticed | 4.4x | Strongest single discriminator — AI opens paragraphs with evaluative lines ("Notably...") without an antecedent subject | | Vacuous anthropomorphic metaphors | Assumed AI dislikes personification | 7.3x | AI favors hollow personas ("like a wise mentor"); humans write grounded ones ("like a veteran mechanic") | | Colon-introduced empty lists | Seen as normal formatting | 9.4x | AI pairs an information-free lead-in colon with a bulleted list | | "Not A but B" contrastive structure | Widely sensed | 3.4x | Confirmed — AI manufactures false oppositions | | Concrete number density | Assumed AI is good with numbers | 0.35x | Humans use ~3x more specific figures (e.g., "38.5%, 42 ms"); AI prefers vague intensifiers ("significantly improved") | | Rhetorical questions in body text | Believed to be AI-specific | 0.06x | Humans ask rhetorical questions 17x more often — deleting them makes text stiffer | | Metaphor frequency overall | Believed to be AI-specific | 0.42x | Humans use metaphors 2.4x more | | Sentence-length uniformity | Believed uniform in AI | No significant difference | An earlier "51x" result was a segmentation artifact (Markdown tables not excluded) and was retracted |
> Frequency ratio (R): feature frequency in AI text divided by frequency in human text. R ≥ 2.0 indicates an AI-leaning marker; R < 0.8 indicates a human-leaning one.
No Single "AI Tone" Exists
Model-specific quirks diverge dramatically — dash usage ranges from DeepSeek's 5.16 to GPT's 0.11 per 1,000 characters, a ~46.9x spread:
Three Surgical Editing Principles
The project's rule set (SKILL.md) enforces:
1. Whitelist principle — only the 11 empirically validated features may be edited; untouched sentences stay byte-identical. 2. Information conservation — every content word in the rewrite must trace to the source; no invented numbers or facts. 3. No subjective semantic rewriting — no deleting metaphors because they "feel bad"; only form-level restructuring.
Concrete fixes: restore anaphora ("This architecture change brought the shift" instead of a subjectless "Notably, this brought..."), ground the tenor of metaphors, and replace vague claims with hard numbers ("throughput up 2.59x, latency down 42%").
Scientific Transparency
The report openly documents six systematic corrections caused by regex bugs — e.g., the retracted claim that AI sentence-length uniformity was 51x human — with results overturned and deleted once denominators and regex boundaries were fixed.
Conclusion
AI text feels "off" not because of any rhetoric it uses, but because it hides behind abstract nouns, hollow personas, and vague intensifiers. Human writing leans on concrete numbers, grounded comparisons, and conversational self-questioning. Removing "AI tone" is not about satisfying a checklist — it is about restoring precision and candor.
References
1. Source project: *Lieflat Less AI Tone*, Larashero & Moxt Open-Source Community (2026) — https://github.com/larashero3-dotcom/lieflat-less-ai-tone 2. Theoretical grounding: Biber, D. (1988). *Variation across speech and writing*. Cambridge University Press — the multi-dimensional, multi-feature framework for corpus-based register analysis.