Why popular beliefs about the "AI tone" are mostly wrong
An open-source linguistic study — lieflat-less-ai-tone — published on GitHub by Larashero and the Moxt open-source community (2026) built a controlled corpus of 629 articles, 2,826,972 Chinese characters, 95,000 sentences, and 45,000 paragraphs, placing five leading models (Claude, DeepSeek, Gemini, GPT, Kimi) and human writers under the microscope across 26 style features.
The measurement data show that most folk heuristics for detecting AI text are backwards:
| Linguistic feature | Popular belief | Frequency ratio R (AI ÷ human) | Diagnosis | | :--- | :--- | :---: | :--- | | Zero-anaphora comment at paragraph opening | Rarely noticed | 4.4x (highly discriminative) | New paragraphs open with judgments like "this sounds like a feature description" without an antecedent, breaking cohesion | | Anthropomorphic metaphor | Believed rare in AI | 7.3x | AI prefers hollow personas ("like a wise mentor"); humans write "like a mechanic who's fixed engines for 30 years" | | Colon-introduced empty list | Mistaken for good formatting | 9.4x | An information-free boilerplate sentence followed by a list | | Antithesis ("not A but B") | Strongly felt by readers | 3.4x | Forces false oppositions where no conflict exists | | Concrete number density | Believed AI-favored | 0.35x | Humans write "38.5% and 42 ms"; AI writes "significant improvement" | | Rhetorical questions in body text | Believed AI-exclusive | 0.06x (humans 17x higher) | Human writing has natural dialogic rhythm | | Total metaphor frequency | Believed AI-exclusive | 0.42x (humans 2.4x higher) | Real authors use more concrete imagery |
Where R ≥ 2.0 indicates the feature is significantly more frequent in generated text, and R < 0.8 means human writing favors it.
No single "AI style" exists
Different models diverge dramatically on the same syntactic feature. Dash usage ranged from 5.16 per 1,000 characters (DeepSeek) to 0.11 (GPT) — a spread of roughly 46.9x:
- DeepSeek & Claude: high dash frequency (4–5 per 1,000 chars), suggestive colons
- Gemini: frequent question-form headings
- GPT: more antithesis constructions; almost no dashes
- Kimi: denser ordinal headings ("First, Second…")
- Open-source project: *Lieflat Less AI Tone* — Larashero & Moxt Open-Source Community (2026) — https://github.com/larashero3-dotcom/lieflat-less-ai-tone
- Biber, D. (1988). *Variation across speech and writing*. Cambridge University Press.
Style rules derived from a single model's samples therefore lack generality.
Three surgical rewriting principles
1. White-list principle — only edit the 11 empirically confirmed features; everything else stays verbatim. 2. Information conservation — every content word after rewriting must trace back to the source; no fabricated data. 3. No subjective semantic judgment — only address localizable formal markers, not impressions.
Concrete fixes include restoring anaphora at paragraph openings ("This architecture brought the change"), concretizing metaphors ("like a mechanic with 30 years in the workshop"), and keeping specific data ("throughput up 2.59x, VRAM down 70%") instead of vague claims like "significant performance leap."
Scientific honesty: six documented measurement failures
The report transparently documents 6 measurement errors caused by regex defects — including an early, wrong conclusion that AI sentence-length uniformity was "51x that of humans" (caused by unfiltered Markdown tables). After correcting denominators and regex boundaries, the authors retracted that finding and published the full correction process.
Takeaway
Stiff, machine-sounding text comes not from rhetorical devices but from abstraction. Human writing anchors in concrete numbers, grounded references, and natural cohesion. De-AI-ifying text means returning to precise, specific, honest expression.