Why Most Beliefs About "AI Tone" Are Wrong: A Microscopic Linguistic Study of 2.83 Million Characters

Imagine a physics experiment: the whole town believes that whenever a ghost (AI-written text) is present, the air smells of lavender (too many dashes, metaphor-heavy prose, self-answered rhetorical questions). Vigilante patrols flag any text containing a metaphor or rhetorical question as machine-generated. Yet no one has ever brought out a precision mass spectrometer to weigh the molecules one by one.

The open-source community has now done exactly that. The GitHub project lieflat-less-ai-tone presents a controlled linguistic study built on a corpus of 629 articles, 2,826,972 Chinese characters, 95,000 sentences, and 45,000 paragraphs, placing five top models (Claude, DeepSeek, Gemini, GPT, Kimi) and human writers under the microscope across 26 stylistic features. The measurements show that much of the folk wisdom about "AI tone" is backwards: humans use metaphors 2.4x as often as AI, and rhetorical questions in body text 17x as often. The real giveaways cluster in 11 features at the paragraph-cohesion and structural level.

Key Points

1. Popular Bias vs. Statistical Truth

• "AI loves metaphors" — False: humans use metaphors 2.4x more often; metaphor-opened paragraphs reach 8x.
• "AI always self-answers rhetorical questions" — False: human body-text rhetorical questions are 17x more frequent.
• "AI sentence lengths are too uniform" — False: after fixing the segmentation code, the sentence-length standard-deviation ratio is 0.87 — no statistical difference.

Measured features with strong discriminative power (frequency ratio \(R = \text{AI} \div \text{human}\)):

| Feature | \(R\) (AI ÷ human) | Diagnosis | | :--- | :---: | :--- | | Zero-anaphora comment at paragraph opening | 4.4x | New paragraphs open with bare evaluations ("sounds like a feature description") with no subject, forcing readers to backtrack | | Anthropomorphized metaphor vehicles | 7.3x | Prefers hollow personas like "like a wise mentor"; humans write "like a mechanic who's repaired engines for 30 years" | | Colon-introduced filler lists | 9.4x | An information-free boilerplate colon sentence followed by a list | | Forced contrastive structures ("not A but B") | 3.4x | Manufactures false oppositions where no conflict exists | | Concrete-number density | 0.35x | Humans write "38.5% and 42 ms"; AI writes "significantly improved, greatly optimized" (humans ~3x higher) | | Rhetorical questions in body text | 0.06x | Humans have natural conversational rhythm; removing questions makes prose stiff | | Total metaphor frequency | 0.42x | Real authors draw on concrete objects; AI's metaphor use is comparatively low |

> Frequency Ratio (\(R\)): frequency of a feature in generated text divided by its frequency in human writing. \(R \ge 2.0\) means significantly more common in generated text; \(R < 0.8\) means humans favor it.

2. Per-Model Differences: There Is No Single "AI Style"

Mixing all LLMs together is inaccurate. The range across models on the same feature can be enormous:

\[\text{Dash-usage range} = \frac{\text{DeepSeek: 5.16 per 1k chars}}{\text{GPT: 0.11 per 1k chars}} \approx \mathbf{46.9\,\text{times}}\]
• DeepSeek & Claude: high dash frequency (4–5 per 1k chars), more suggestive colons
• Gemini: high use of question-form section headings
• GPT: more contrastive structures; almost no dashes (0.11)
• Kimi: denser ordinal-number headings ("first, second...")

Rules derived from a single model's samples therefore lack generality.

3. Surgical Rewriting: Three Principles for Removing AI Tone

1. Whitelist principle — only edit the 11 explicitly matched features; every other sentence is preserved verbatim. 2. Information conservation — every content word after rewriting must trace back to the source; no invented data. 3. No subjective semantic judgment — no edits based on subjective feel; only localizable formal markers are handled.

Concrete operations:

• Restore anaphora: "Notably, this brought change" → "This architecture design brought change"
• Concretize metaphors: "like a guide versed in all ages" → "like a master mechanic who's rebuilt engines for thirty years"
• Keep concrete data: delete "achieved a significant performance leap" and write "throughput up 2.59x, VRAM down 70%"

4. Scientific Honesty: Six Measurement Errors, Publicly Documented

The report transparently documents six measurement errors caused by regex design flaws. Early code failed to filter Markdown tables when segmenting sentences, briefly producing the erroneous conclusion that "AI sentence-length uniformity is 51x that of humans." After correcting the denominator and regex boundaries, the team retracted the claim and published the full correction process.

5. Conclusion: Authentic Writing Rests on Concrete Detail

Whether prose reads stiffly depends not on rhetorical flourish but on whether content is grounded in concrete physical reality. Machine text lingers in abstraction, macro-level generalization, and hollow personas; genuine human writers favor concrete data, down-to-earth reference points, and natural tonal flow. Removing machine traces means returning to precise, specific, honest expression.

References

1. Open-source project: *Lieflat Less AI Tone*, Larashero & Moxt Open-Source Community (2026) — https://github.com/larashero3-dotcom/lieflat-less-ai-tone 2. Corpus linguistics foundation: Biber, D. (1988). *Variation across speech and writing*. Cambridge University Press.