Overview
An open-source linguistic study, published on GitHub as lieflat-less-ai-tone, set out to test popular folk beliefs about "AI tone" — the assumption that AI-generated writing is easy to spot because it overuses metaphors, rhetorical questions, or dashes.
The authors built a controlled corpus of 629 articles, 2,826,972 Chinese characters, 95,000 sentences, and 45,000 paragraphs, covering five top models (Claude, DeepSeek, Gemini, GPT, Kimi) plus human writers, and measured 26 stylistic features.
Key points
- Popular beliefs are mostly backwards:
- Humans use metaphors 2.4x more often than AI (metaphor-opening paragraphs: 8x).
- Human self-answering rhetorical questions in body text are 17x more frequent than AI's.
- Sentence-length variability shows no statistical difference (std-dev ratio 0.87) after fixing a segmentation bug.
- The 11 genuinely discriminative features are structural:
- Zero-anaphora comments at paragraph openings (4.4x)
- Personified, abstract metaphors like "like a wise mentor" (7.3x), versus humans' grounded comparisons
- Colon-introduced empty list preambles (9.4x)
- Forced "not-A-but-B" contrast structures (3.4x)
- Lower concrete-number density — humans write "38.5% and 42 ms", AI writes "significant improvement"
- No single "AI style" exists: dash frequency ranges from DeepSeek's 5.16/1,000 characters to GPT's 0.11 — a 46.9x spread. Gemini favors question-form headings; Kimi overuses ordinal headings.
- Transparent methodology: the authors publicly documented 6 measurement errors (e.g., a regex bug that briefly suggested AI sentence length was "51x more uniform" than humans') and their corrections.
- Larashero & Moxt Open-Source Community (2026). *Lieflat Less AI Tone*. GitHub:
https://github.com/larashero3-dotcom/lieflat-less-ai-tone - Biber, D. (1988). *Variation across speech and writing*. Cambridge University Press.
Three rewriting principles for de-AI-ifying text
1. Whitelist principle — only edit the 11 empirically identified patterns; everything else is preserved verbatim. 2. Information conservation — every rewritten content word must trace back to the source; no invented facts. 3. No subjective semantics — only locally identifiable surface markers are touched, not subjective judgments of rhetoric.
Typical fixes: restoring anaphora ("this architecture change brought..."), concretizing metaphors ("like a mechanic who has rebuilt engines for thirty years"), and replacing vague claims with the actual numbers already in the text.
Conclusion
Stiff, machine-sounding prose comes not from rhetorical devices but from abstraction. Human writing is grounded in concrete numbers, relatable references, and natural discourse flow — effective de-AI rewriting means returning to precise, specific, honest expression.