2.83-Million-Character Corpus Study Debunks Myths About Chinese AI Writing Style

Forum topic · 小凯 · 2026-08-30

Summary

An open-source project called lieflat-less-ai-tone conducted a controlled corpus-linguistics study of 2.83 million Chinese characters (629 articles: 300 AI-generated texts from Claude, DeepSeek, Gemini, GPT, and Kimi across 38 identical topics, versus 329 human-written deep-dive articles) to empirically test popular assumptions about so-called "AI flavor" in writing. Using a frequency-ratio metric (R = generation frequency / human frequency, with R ≥ 2.0 as the threshold), the study overturned 15 widespread beliefs and validated 11 genuine AI-style markers. Debunked myths include: metaphors are actually 2.4x more common in human text; humans ask rhetorical questions 17x more often than AI; sentence-length uniformity shows no significant difference (R=0.87). Real AI tells include zero-anaphora paragraph openings (4.4x), lofty anthropomorphic metaphors (7.3x), performative reversal constructions (3.4x), filler colons (3.8-9.4x), and dash overuse (3.0x, notably DeepSeek at 5.16/thousand characters vs GPT's 0.11). The project, released as an agent rewriting skill, enforces a whitelist of only verified issues and strict information conservation—no invented facts or altered judgment strength.

2.83-Million-Character Study: What Actually Makes Text Sound "AI-Generated"?

Imagine buying watermelons in summer: everyone repeats folk wisdom like "a dull thud means it's sweet." Then a rigorous scientist measures 629 watermelons and finds that 15 of 26 popular tricks are pure fiction. According to this post, that's exactly how the writing community treats the notion of "AI flavor" — until now.

The open-source project lieflat-less-ai-tone (a de-AI-tone agent skill) ran a controlled corpus study on 2.83 million Chinese characters to empirically quantify what really distinguishes AI writing from human writing.

Study Design

Generation side: 300 articles / 1.18M characters from 5 mainstream models (Claude, DeepSeek, Gemini, GPT, Kimi), across 38 identical topics, with no retrieval or style instructions
Human side: 329 articles / 1.65M characters of published high-quality human deep-dive articles, cross-checked for inter-group consistency
Metric: frequency ratio R = generation frequency ÷ human frequency; a feature passes only if R ≥ 2.0 and is stable across models and human groups

Result: 11 core features validated; 15 popular beliefs refuted.

4 Popular Myths Destroyed by the Data

| Myth | Reality | |---|---| | ❌ AI overuses metaphors | Humans use metaphors 2.4x more than AI; standalone metaphor paragraphs 8x more | | ❌ Rhetorical questions are an AI tell | Humans pose-and-analyze questions 17x more often (R = 0.05) | | ❌ AI sentence length is too uniform | After fixing tokenization (excluding Markdown tables/lists), coefficients of variation are 0.58 vs 0.67 (R = 0.87) — no real difference | | ❌ Add filler particles (就/很/了) to sound human | Forced insertion fails verification and ruins style |

The real metaphor difference is in tenor choice: AI prefers lofty abstract figures ("like a wise mentor," "a tireless guardian," R = 7.3), while humans prefer concrete everyday craftsmen ("like a mechanic," "like a doctor friend with lots of WeChat contacts").

The 11 Real AI Tells (Discourse & Structure Level)

| # | Feature | R | |---|---|---| | 1 | Zero-anaphora paragraph-opening commentary ("Notably, ..." with no referent) | 4.4x | | 2 | Reversal/contrastive framing ("The real barrier isn't technology, it's cognition") | 3.4x | | 3 | Lofty anthropomorphic metaphors | 7.3x | | 4 | Empty signposting colons ("Mainly as follows:") | 3.8–9.4x | | 5 | Dash overuse | 3.0x | | 6 | Verbal filler openers ("Plainly speaking, ...") | 3.2x | | 7 | Ordinal-number subheadings ("1. Background 2. Status 3. Summary") | 3.1x | | 8 | Mirror-symmetric adjacent sentences | 2.0x | | 9 | 5 typical translationese patterns (overlong premodifiers, "when..." clauses, fronted topic shells, sentence-initial "However,") | 2.6–5.3x | | 10 | Dense enumeration commas | 1.8x | | 11 | Generalizing away specific data ("grew 35%" → "grew significantly") | 0.35x (reversed) |

Different Models, Different Habits

There is no single unified "AI style" — model-specific habits differ dramatically:

Dash frequency: DeepSeek (5.16/thousand chars) > Claude (4.25) >>> GPT (0.11) — a 47x gap
Reversal 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)

So conclusions drawn from a single model (e.g., GPT) are unreliable; de-AI-tone rules must extract cross-model commonalities.

Three Iron Rules of the Rewriting Skill

1. Whitelist principle: only edit the 11 verified issues; everything else is preserved verbatim 2. Information conservation: no new facts, data, or examples; never alter the author's judgment strength (e.g., "may improve" must not become "improves") 3. No subjective semantics: rules must be machine-locatable lexical/syntactic operators

Takeaway

De-AI-toning is not about faking human quirks — it's about removing the probabilistic laziness, hollow filler phrases, and structural tics that LLMs produce. Kill the empty signposts, fix paragraph-opening referents, and drop the guru metaphors, and what remains is writing with genuine backbone.

Project Metadata

Project: Lieflat Less AI Tone (de-AI-tone.skill)
Corpus: 629 articles, 2.827M characters, 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