2.83 Million Characters Tested: Most Beliefs About the 'AI Tone' Are Wrong
Like watermelon buyers trusting folk wisdom, Chinese writers have long relied on unverified rules to spot AI-written text. The open-source project lieflat-less-ai-tone put these beliefs to the test with a controlled corpus of 2.83 million Chinese characters — and debunked most of them.
Experiment Design
- AI-side corpus: 300 articles / 1.18M chars from 5 major models (Claude, DeepSeek, Gemini, GPT, Kimi), on 38 identical topics, with no retrieval or style instructions
- Human-side corpus: 329 publicly published high-quality articles / 1.65M chars, checked for cross-source consistency
- Verification method: frequency ratio R = generation frequency ÷ human frequency. A feature counts as an AI marker only if R ≥ 2.0 and is stable across models and human groups.
- Dash frequency: DeepSeek (5.16/1000 chars) > Claude (4.25) ≫ GPT (0.11) — a 47× spread
- Contrast framing: GPT (1.26) > DeepSeek (0.86) > Claude (0.61) > Gemini (0.29)
- Question subheadings: Gemini (0.173) ≫ Claude (0.043) > GPT (0.008) > Kimi (0.000)
- Project: Lieflat Less AI Tone
- Corpus: 629 articles, 2.827M chars, 95K sentences, 45K paragraphs
- Models tested: Claude Opus 4.6, DeepSeek V4 Pro, Gemini 3.1 Pro, GPT-5.6 Sol, Kimi K3
- GitHub: https://github.com/larashero3-dotcom/lieflat-less-ai-tone
- MoxtHub: https://moxt.ai/zh-CN/hub?type=skill&id=lieflat-less-ai-tone
Result: 11 core features passed; 15 popular beliefs were refuted.
Four Popular Myths Debunked
1. "AI overuses metaphors" — Humans use metaphor markers 2.4× more than AI; metaphor-only paragraphs are 8× more frequent in human writing. The real difference is tenor choice: AI prefers lofty abstract figures (R = 7.3, e.g., "like a wise mentor"), humans prefer concrete life-based ones ("like a mechanic", "like a doctor friend with many WeChat contacts"). 2. "Rhetorical questions betray AI" — Humans pose questions in body text 17× more often than AI (R = 0.05). Deleting all question marks makes text *more* machine-like. 3. "AI sentence lengths are too uniform" — An artifact of early detection scripts including Markdown tables/lists. After correction, coefficients of variation were 0.58 vs 0.67 (R = 0.87) — no meaningful difference. 4. "Add function words (就/很/了) to humanize" — Artificial insertion fails verification and damages style.
The 11 Real AI Markers
Mostly discourse- and structure-level flaws:
| # | Feature | R | |---|---------|---| | 1 | Paragraph-opening comments with zero referent ("Notably, …") | 4.4× | | 2 | Reversal/contrast framing ("not X, but Y") | 3.4× | | 3 | Anthropomorphic lofty metaphors | 7.3× | | 4 | Empty colon-intro sentences ("mainly as follows:") | 3.8–9.4× | | 5 | Dash overuse (DeepSeek/Claude habit) | 3.0× | | 6 | Filler openers ("simply put", "in short") | 3.2× | | 7 | Ordinal-number subheadings | 3.1× | | 8 | Structurally mirrored adjacent sentences | 2.0× | | 9 | Five typical translation-ese patterns | 2.6–5.3× | | 10 | Over-dense enumerations with pause marks | 1.8× | | 11 | Vague summaries replacing concrete data (reverse: 0.35×) | 0.35× |
Model Styles Differ Dramatically
There is no single "AI tone":
De-AI-flavor rules must therefore extract cross-model commonalities rather than generalize from one model.
Three Rewriting Rules
The project turns findings into an agent rewriting skill with three red lines:
1. Whitelist principle: only edit the 11 verified issues; preserve all other text verbatim 2. Information conservation: never add facts/data or alter the author's judgment strength (e.g., "may improve" must not become "improves") 3. Excluding subjective semantics: rules must be machine-locatable literal and syntactic operators
Conclusion
De-AI-ing is not about faking human quirks — it is about removing the logical laziness, empty boilerplate, and residual grammar glitches that emerge from probabilistic prediction. The result is writing with real structure, warmth, and force.