2.83 Million Characters Weighed: Most 'AI Tone' Writing Rules Are Backwards

Overview

A popular checklist circulating in writing circles claims to identify AI-generated text by signs like excessive dashes, too many metaphors, constant self-Q&A, and overly uniform sentence lengths. The open-source project lieflat-less-ai-tone put this checklist on the scale, measuring all 26 items across 629 articles: 300 AI-generated texts covering Claude, DeepSeek, Gemini, GPT, and Kimi, and 329 human-written pieces — 2,826,972 Chinese characters, 95,551 sentences, and 45,721 paragraphs in total. The data overturned most of the checklist.

Items Measured Backwards

Metaphors. Human writers use metaphor markers 2.4x more frequently per thousand characters than generated text, and start paragraphs with standalone metaphors 8x more often. The difference lies in the choice of vehicle: generated text favors idealized professional personas like "like a wise mentor" (R = 7.3); humans write "like an experienced mechanic" or "like a well-connected doctor friend."

Rhetorical questions. Human texts contain 17x more rhetorical questions in body text. Following the popular advice to delete them pushes text further from human writing.

Sentence-length uniformity. An early measurement showed AI uniformity at 51x human levels — caused by a segmentation bug that didn't filter Markdown tables and punctuation-less list items. After fixing the splitter, the ratio was 0.87: no difference.

| Popular checklist item | Popular claim | Measured ratio R (AI ÷ human) | Actual finding | | :--- | :--- | :---: | :--- | | Metaphors | "AI-only" | 0.42 | Humans use 2.4x more | | Body rhetorical questions | "AI-only" | 0.06 | Humans use 17x more | | Sentence-length uniformity | "AI too even" | 0.87 | No difference; early data was a bug | | Parallel structure | "AI loves it" | ~1.0 | No difference |

> Frequency ratio R: frequency of a feature in generated text divided by frequency in human text. R ≥ 2.0 means AI noticeably higher; R < 0.8 means humans higher.

The 11 Features That Actually Discriminated

Of 26 items, 11 passed, concentrated in discourse structure and syntax:

• Zero-anaphora paragraph-opening comments (R = 4.4, strongest discriminator). New paragraphs open with judgments like "worth noting" without specifying what's being commented on. Adding a single "this" fixes it.
• Colon-introduced empty lists (R = 9.4). An information-free sentence ending in a colon just to announce a list. Near-zero on the human side.
• Anthropomorphic hollow metaphors (R = 7.3). Comparing tools to "a wise mentor" or "a tireless digital secretary" with strings of flattering modifiers. Metaphors themselves aren't the problem — only this floating persona type.
• Antithetical constructions (R = 3.4). "Not A but B" forced where no contradiction needs correcting.
• Specific-number density (R = 0.35, reverse). Human texts contain nearly 3x more concrete numbers; AI prefers "significantly improved" over exact values.

Same Feature, 47x Different Across Models

Dash frequency: DeepSeek 5.16 per thousand characters, Claude 4.25, GPT only 0.11 — a range of ~47x:

\[\frac{\text{DeepSeek dash frequency: 5.16}}{\text{GPT dash frequency: 0.11}} \approx 46.9\]

DeepSeek and Claude lead on suggestive colons, Gemini on question headings, GPT on antithetical structures. Rules derived from a single model don't transfer — there is no unified "AI style."

From Statistics to Execution

The 11 validated features became executable rewrite rules (published as SKILL.md) under three hard constraints:

• Whitelist principle. Only the 11 items are processed; unmatched sentences preserved verbatim; heading hierarchy, paragraph order, and list structure untouched.
• Information conservation. Every content word after rewriting must trace to the original. No new names, numbers, dates, or causality; no deleting qualifiers like "possibly" or "usually." Changing "may improve" to "improves" is tampering with judgment strength.
• No subjective semantic judgments. Metaphor aptness, question necessity, and case-list piling depend on semantic understanding and can't be written as executable operators — excluded entirely.

Six Failure Records

The report documents six measurement errors, all from regex operators matching more broadly than the rule definitions. In one case, long pre-modifier frequency appeared to be 3.04/thousand characters; after checking hits, it dropped to 0.35. In another, a metaphor lookup table of 11 fixed terms missed all actual metaphors (initially 0.000; corrected to R = 0.42, humans higher). Four of the six would have entered the rule set uncorrected, two in the wrong direction — rewrites under the buggy rules would have made text more AI-like, not less.

References and Open-Source Archive

1. Open-source project:

• Name: *Lieflat Less AI Tone*
• Authors: Larashero & Moxt open-source community (2026)
• Repository: github.com/larashero3-dotcom/lieflat-less-ai-tone
• Corpus scale: 629 articles, 2,826,972 characters, 95,551 sentences, 45,721 paragraphs
• Models covered: Claude claude-opus-4-6, DeepSeek deepseek-v4-pro, Gemini gemini-3.1-pro, GPT gpt-5.6-sol, Kimi kimi-k3
• Core finding: 11 of 26 candidate features passed testing; 15 had no discriminating power (metaphors and rhetorical questions were reversed from popular belief)

2. Corpus linguistics foundation:

• Biber, D. (1988). *Variation across speech and writing*. Cambridge University Press — established the multi-dimensional multi-feature framework underpinning this analysis.