English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Why Popular Beliefs About the 'AI Tone' Are Wrong: A 2.83-Million-Character Corpus Study

Forum topic · 小凯 · 2026-08-26

Summary

An open-source linguistic study from the GitHub project lieflat-less-ai-tone analyzed a controlled corpus of 629 articles, 2,826,972 Chinese characters, ~95,000 sentences, and ~45,000 paragraphs written by five major LLMs (Claude, DeepSeek, Gemini, GPT, Kimi) and human writers, testing 26 stylistic features. Results contradict common folk wisdom: humans use metaphors 2.4 times more often than AI and rhetorical self-questions 17 times more often, while sentence-length uniformity shows no statistical difference. The 11 features that genuinely distinguish AI text are structural and discourse-level traits, including zero-anaphora comments at paragraph openings (4.4x), anthropomorphic personified metaphors (7.3x), colon-introduced filler lists (9.4x), forced contrastive constructions (3.4x), and low concrete-number density (humans nearly 3x higher). Individual models differ sharply—for example, DeepSeek uses dashes 46.9 times more per 1,000 characters than GPT—so no single 'AI style' exists. The project also transparently documents six measurement errors and corrections, and proposes three rewriting principles: white-list edits, information conservation, and no subjective semantic judgment.

Overview

An open-source linguistic study published on GitHub as lieflat-less-ai-tone examines which stylistic features actually distinguish LLM-generated Chinese text from human writing. The authors built a controlled corpus of 629 articles, 2,826,972 Chinese characters, ~95,000 sentences, and ~45,000 paragraphs, comparing five frontier models (Claude, DeepSeek, Gemini, GPT, Kimi) with human writers across 26 stylistic features.

Key Findings: Popular Beliefs vs. Measured Data

Most commonly cited markers of the "AI tone" turned out to be wrong:

| Feature | Folk belief | Measured frequency ratio \(R\) (AI ÷ human) | | :--- | :--- | :---: | | Zero-anaphora comment at paragraph opening | Rarely noticed | 4.4x (highly discriminative) | | Anthropomorphic metaphor personae | Believed rare in AI | 7.3x | | Colon-introduced empty filler lists | Considered normal formatting | 9.4x | | Contrastive "not A but B" structures | Strongly sensed by readers | 3.4x | | Concrete number density | Believed AI excels at data | 0.35x (humans ~3x higher) | | Rhetorical questions in body text | Believed AI-specific | 0.06x (humans 17x higher) | | Total metaphor frequency | Believed AI-specific | 0.42x (humans 2.4x higher) |

Here \(R \ge 2.0\) indicates the feature is significantly over-represented in generated text, while \(R < 0.8\) means human writers use it more.

  • "AI loves metaphors" — wrong: humans use metaphors 2.4x more, and open paragraphs with them 8x more.
  • "AI always asks and answers its own questions" — wrong: human rhetorical questions are 17x more frequent.
  • "AI sentences are too uniform" — wrong: after fixing the segmentation program, the sentence-length standard-deviation ratio was 0.87, not statistically significant.
  • The 11 genuinely discriminative features concentrate at paragraph-cohesion and structural levels: paragraph openings that drop referential subjects (forcing readers to backtrack), floating anthropomorphic metaphors ("like a wise mentor" vs. a human's "like a veteran mechanic"), and vague abstractions replacing concrete data.

    Model-Level Differences: No Single "AI Style"

    Dash usage varies enormously across models:

    \[\text{Dash range} = \frac{\text{DeepSeek: } 5.16 / 1000\text{ chars}}{\text{GPT: } 0.11 / 1000\text{ chars}} \approx \mathbf{46.9\,\text{times}}\]
  • DeepSeek & Claude: high dash frequency (4–5 per 1,000 characters), more suggestive colons.
  • Gemini: frequent question-style section headings.
  • GPT series: more "not A but B" constructions, very few dashes (0.11).
  • Kimi: denser ordinal-number headings.
  • Style rules derived from a single model therefore generalize poorly.

    Three Surgical Rewriting Principles

    1. White-List principle — only edit the 11 explicitly matched features; everything else is preserved verbatim. 2. Information conservation — every content word after rewriting must trace back to the source; no invented facts. 3. No subjective semantic judgment — only address localizable formal markers, not personally disliked rhetoric.

    Concrete techniques: restore anaphoric pronouns at paragraph starts ("this architectural change..."), concretize metaphors ("like a mechanic who has rebuilt engines for thirty years"), and keep specific numbers ("throughput improved 2.59x, VRAM reduced 70%") instead of vague claims like "significant performance gains."

    Transparency: Six Documented Measurement Errors

    The report openly documents six measurement failures caused by regex defects — including an early bug (failing to filter Markdown tables in sentence segmentation) that produced an erroneous "AI sentence-length uniformity is 51x human" conclusion, later retracted with the full correction procedure published.

    Conclusion

    Whether text reads as stilted depends not on rhetorical flourishes but on whether it is grounded in concrete physical reality. Human writing favors specific data, grounded references, and natural conversational flow; removing machine traces means returning to precise, specific, honest expression.

    References

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

Tags

#ai-generated-text-detection#stylometry#corpus-linguistics#large-language-models#nlp#text-analysis#writing-style#lieflat-less-ai-tone

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178634048