# Lieflat Less AI Tone

[中文](README.md) | English

[![less-ai-tone.skill: quantifying what "AI tone" actually is, then removing it](assets/less-ai-tone-hero-en.png)](https://moxt.ai/hub?type=skill&id=lieflat-less-ai-tone)

**An AI-tone removal skill grounded in a 2.83-million-character corpus study**

This skill applies linguistic methods to quantify what "AI tone" actually is in Chinese writing, and to remove it. It was created at [moxt.ai](https://moxt.ai).

The study builds a contrastive corpus of 629 articles (2,826,972 Chinese characters, 95,551 sentences, 45,721 paragraphs). 300 were generated by five mainstream models under controlled conditions; 329 are human-authored texts. It tests 26 candidate features widely cited in public discussion as markers of "AI style", computing a frequency ratio between the generated and human sides for each.

Eleven features pass. The strongest discriminator is zero-anaphora paragraph-initial commentary (4.4×). Fifteen fail, and several run opposite to popular belief: humans use figurative language 2.4× as often as generated text, in-body rhetorical questions 17× as often, and intra-sentence parallelism more often as well.

The study also finds that "AI tone" is not one thing. Em dashes run 5.16 per thousand characters for DeepSeek and 4.25 for Claude, but only 0.11 for GPT — a fortyfold spread. Cue colons peak with DeepSeek and Claude; interrogative subheadings peak with Gemini.

The final sections document six measurement errors caused by operator design flaws, along with their corrections, and discuss three methodological limits: single-model sampling, denominator choice, and a corpus that cannot be independently verified.

**Keywords** generated-text detection; Chinese syntax; corpus linguistics; discourse cohesion; contrastive study

> [**Open Lieflat Less AI Tone on MoxtHub →**](https://moxt.ai/hub?type=skill&id=lieflat-less-ai-tone)

---

## 1 The problem

Text generated by large language models carries a perceptible stylistic signature that Chinese writing communities call "AI tone" (AI 味). The judgment enjoys broad consensus. The evidence behind it does not go much past anecdote.

The checklists in circulation run roughly as follows: em dashes used too densely; the antithetical construction "not A but B" recurring; sentence lengths too uniform; monosyllabic particles (就, 很, 了) underused; abstractions dressed in concrete imagery ("the warehouse of memory"); rhetorical questions immediately self-answered. These items propagate through social media and gradually acquire the standing of common knowledge.

None of them has been tested systematically. The items come from mixed sources — some from observing early model output, some transplanted directly from English-language findings, some from a single writer's impression. Which carry statistical discriminative power, which are artifacts of observer bias, and which point in the opposite direction, has gone unexamined.

This study does not propose a new checklist. It tests the existing one item by item and specifies a recomputable decision procedure.

## 2 Method

### 2.1 Design

A contrastive design. Text source (model-generated vs. human-authored) is the independent variable; the frequency of each candidate syntactic feature is the dependent variable.

Valid contrast requires controlling three classes of confound.

**Topic and genre.** The generated side's topic range and genre settings are aligned with the human side. If the two sides are not writing about comparable things, observed differences cannot be attributed to text source and may merely reflect genre.

**Model-individual variation.** The generated side draws 60 articles from each of five models rather than relying on one. An earlier round of this study used 30 articles from two models; most of its conclusions failed to replicate. Take em dashes: the early data supported "this feature is unique to Claude", but expanding to five models revealed a different model with a higher rate. One model's idiosyncrasy is easily misread as a property of generated text in general.

**Generation conditions.** The generated side ran without web retrieval and without style instructions, receiving only a topic. Added style constraints would mask each model's default tendencies.

### 2.2 Corpus

| Side | Articles | Characters | Sentences | Paragraphs |
|---|---|---|---|---|
| Generated | 300 | 1,179,105 | 37,642 | 14,549 |
| Human | 329 | 1,647,867 | 57,909 | 31,172 |
| Total | 629 | 2,826,972 | 95,551 | 45,721 |

Models: claude-opus-4-6, deepseek-v4-pro, gemini-3.1-pro, gpt-5.6-sol, kimi-k3, 60 articles each, spanning 38 topics.

The human side consists of publicly available texts, tallied in groups. Grouped tallies test cross-group consistency: if a feature is elevated in only one group, it more likely reflects an individual author's style or a genre convention than a property of human writing. Monosyllabic particles and em dashes were both flagged as unreliable indicators this way.

### 2.3 Segmentation

**Sentences.** Split by line first, discarding heading lines, table rows, code blocks, block quotes, and list items; then split on `。！？；`, keeping fragments of at least 4 characters.

Discarding non-body lines is necessary. An early version split only on `。！？`, so Markdown tables and long lists without terminal punctuation counted as single sentences. One group's sentence-length standard deviation was consequently computed as 27 times its mean, yielding the conclusion that "generated text is 51 times more uniform in sentence length than human writing". After correcting the segmenter, the ratio is 0.87 — no difference.

**Paragraphs.** Split on blank lines, discarding headings, tables, code blocks, quotes, list items, and image lines; keep paragraphs of at least 8 characters.

### 2.4 Denominator selection

Chosen according to the linguistic unit the feature attaches to.

| Denominator | Applies to | Examples |
|---|---|---|
| Per thousand characters | Lexical and punctuation features | Em dashes, enumeration commas, antithetical constructions |
| Per hundred paragraphs | Paragraph-level structure | Adjacent-sentence isomorphism, figurative paragraph openings |
| Share of same-type elements | Features bound to a specific structure | Interrogative subheadings as a share of all subheadings |

A wrong denominator produces directional error. The generated side averages about 10 subheadings per article, the human side 1 to 2. Per thousand characters, interrogative subheadings show a ratio of 32; recomputed against total subheadings, the generated side is 2.7% and the highest human group 3.6%, and the difference disappears. The first denominator measures "generated text has more subheadings"; only the second measures "generated text tends to phrase subheadings as questions".

Paragraph-initial features use non-initial paragraphs as the denominator, since an opening paragraph has no prior text to refer back to.

### 2.5 Decision criteria

The primary criterion is the frequency ratio *R* = generated frequency ÷ human frequency.

| *R* | Decision |
|---|---|
| ≥ 2.0 | Include |
| 1.25 ≤ *R* < 2.0 | Conditional; requires cross-group consistency and sufficient absolute volume |
| 0.8 ≤ *R* < 1.25 | No discriminative power; exclude |
| < 0.8 | Higher on the human side; not usable as grounds for deletion either |

Three additional conditions apply.

**Cross-group consistency.** The range across human groups must not exceed a factor of 5. Em dashes span a hundredfold range across groups (0.01 to 1.29), so despite *R* = 3.0 the rule carries an added qualification.

**Operator locatability.** A feature must resolve to specific clauses and word forms. Judgments requiring semantic understanding — whether a metaphor is apt, whether parallel cases amount to padding — cannot be transcribed into executable operators and are excluded without exception, however much they subjectively resemble AI tone.

**Rewrite operability.** A feature holding up statistically does not make it usable for rewriting. Material density runs 2.8× higher on the human side, a clear direction, but "add data" would violate the information-conservation constraint; only the executable half is retained, namely that concrete data already present in the source may not be overwritten by a summary phrase. Monosyllabic particles are analogous: genuinely lower on the generated side, but "add particles" cannot be verified, so the item was dropped entirely.

### 2.6 Operator implementation

Each feature is implemented as a regular expression, defined at the top of the relevant script and open to modification and review.

This is the weakest link in the method chain. Regular expressions match surface form without parsing semantics; an over-broad operator admits large numbers of instances that do not belong to the feature. Six measurement errors arose this way (Section 5). The resulting working rule: before accepting any operator's frequency output, sample and inspect 20 matched instances.

## 3 Results

### 3.1 Features that failed

Fifteen items with *R* < 1.25, i.e. no discriminative power or the opposite direction.

| Candidate feature | Measurement | *R* |
|---|---|---|
| Sentence-length uniformity | Coefficient of variation: gen 0.58 / human 0.67 | 0.87 |
| Adjacent sentence-length delta | gen 21.7 / human 21.7 | 1.00 |
| Paragraph-length uniformity | Robust dispersion: gen 0.94 / human 1.00 | 0.94 |
| Monosyllabic particle 就 | gen 2.93 / human 6.45 | 0.45 |
| Colloquial connectives | gen 1.01 / human 3.87 | 0.26 |
| Noun repetition (pronoun avoidance) | Same noun opening adjacent sentences: gen 0.02 / human 0.04 | 0.50 |
| Abstract passives | gen 0.09 | Below threshold |
| Self-answered questions | gen 0.13 / human 0.13 | 1.03 |
| Interrogative subheadings | Share of subheadings: gen 2.7% / highest human group 3.6% | 0.75 |
| In-body rhetorical questions | gen 0.10 / human 1.83 | **0.05** |
| Intra-sentence parallelism | Two items with same opening character: gen 2.35 / human 3.87 | 0.61 |
| Figurative markers | gen 0.16 / human 0.38 | 0.42 |
| Abstract subject with concrete predicate | gen 0.001 / human 0.001 | 0.70 |
| Nominalization | gen 0.003 / human 0.005 | 0.52 |
| In-body ordinals | Sentence-initial 首先: gen 0.06 / human 0.03 | 2.00, volume too low |

Three results run opposite to popular belief and warrant separate comment.

**Figurative language.** The claim that generated text dresses abstractions in metaphor does not hold. The human side uses figurative markers 2.4× as often, and opens paragraphs with a figure 8× as often. The difference lies not in whether metaphor appears but in the choice of vehicle: the generated side favors idealized professional personae ("like a wise mentor", "a tireless junior reviewer", *R* = 7.3), while the human side reaches for specific individuals ("like an old master craftsman", "like a doctor friend with a wide network"). The former performs sentiment; the latter performs explanation.

**Rhetorical questions.** In-body questions occur 17× as often on the human side. Following the popular advice to delete them moves text further from human writing.

**Sentence-length uniformity.** This study accepted the claim in its early phase and reported a ratio of 51. The cause proved to be a segmenter defect (see 2.3). After correction, *R* = 0.87.

### 3.2 Features that passed

| No. | Feature | Measurement | *R* |
|---|---|---|---|
| 1 | Antithetical construction (not A but B) | gen 0.73 / human 0.22 | 3.4 |
| 2 | Dense enumeration commas | gen 3.21 / human 1.78 | 1.8 |
| 3 | Adjacent-sentence isomorphism | Per hundred paragraphs: gen 9.41 / human 4.81 | 2.0 |
| 4 | Em dashes | gen 2.38 / human 0.80 | 3.0 |
| 5 | Colon overuse | cue phrase gen 0.29 / human 0.08; empty line into list 0.29 / 0.03 | 3.8 / 9.4 |
| 6 | Ordinals as subheadings | gen 0.19 / human 0.06 | 3.1 |
| 7 | Personifying vehicles | gen 0.018 / human 0.002 | 7.3 |
| 8 | Summary overwriting existing data | Numeral density: gen 6.34 / human 17.92 | 0.35 (inverse) |
| 9 | Opening formulae (说白了) | gen 0.025 / human 0.008 | 3.2 |
| 10 | Translationese constructions (5) | See 3.4 | 2.6–5.3 |
| 11 | Zero-anaphora paragraph-initial commentary | Share of non-initial paragraphs: gen 0.61% / human 0.14% | **4.4** |

The features that pass cluster at the discourse and structural level rather than the lexical or punctuation level. Item 11, the strongest discriminator, is a cohesion problem: generated text opens a new paragraph by delivering an evaluation ("sounds like a feature description", "it is worth noting that") without marking what is being evaluated, so the reader must look back to establish the referent. Notably, overall paragraph-initial connective density does not differ between sides (gen 14.4%, human 15.1%); the difference lies solely in evaluative clauses omitting the anaphoric element. Inserting a demonstrative pronoun restores cohesion.

### 3.3 Between-model variation

| Feature | Claude | DeepSeek | Gemini | GPT | Kimi |
|---|---|---|---|---|---|
| Em dashes | 4.25 | **5.16** | 0.51 | 0.11 | 2.32 |
| Cue colons | 0.38 | **0.43** | 0.22 | 0.25 | 0.32 |
| Antithetical construction | 0.61 | 0.86 | 0.29 | **1.26** | 0.51 |
| Ordinal subheadings | **1.00** | 0.69 | 0.22 | 0.73 | 0.82 |
| Overlong pre-modifiers | 0.64 | **0.65** | 0.48 | 0.06 | 0.22 |
| Interrogative subheadings | 0.043 | 0.086 | **0.173** | 0.008 | 0.000 |

The range for a single feature across models reaches a factor of forty (em dashes: DeepSeek 5.16, GPT 0.11). Two implications follow. First, there is no unified "AI style"; model-individual variation is not negligible. Second, feature descriptions derived from single-model samples do not generalize, which is the source of several early misjudgments in this study.

GPT's low em-dash rate has been cited as evidence that the feature is obsolete. But Claude's rate exceeds the human side more than fivefold, and that model is widely used for writing, so the rule is retained.

### 3.4 Filtering translationese constructions

The initial list specified 18 items on the criterion of direct transfer from English syntax. Five passed.

| Included | Generated freq. | Excluded | Generated freq. |
|---|---|---|---|
| Overlong pre-modifiers | 0.42 | Abstract passive (被认为) | 0.09 |
| 当…时 clauses | 0.26 | 在…的过程中 | 0.07 |
| Fronted topic shells (对于…来说) | 0.22 | 不仅仅是 | 0.05 |
| Sentence-initial connectives (然而，) | 0.18 | 存在着 / 有着 | 0.04 |
| 这意味着 restatements | 0.15 | 进行了…的分析 | 0.03 |
| | | 使得…能够 | 0.03 |
| | | 扮演…角色 | 0.02 |
| | | 在某种程度上 | 0.02 |
| | | 值得注意的是 | 0.02 |
| | | 以一种…的方式 | 0.01 |
| | | 一系列的 | 0.00 |

Three of the excluded items were part of this rule set's own initial version (以一种…的方式, 使得…能够, 扮演…角色), all measuring below 0.03. These are indeed translationese constructions, but they do not distinguish generated text; modern written Chinese has absorbed such syntax thoroughly.

## 4 From statistics to rewrite rules

Statistical significance does not imply executability. Transcribing the eleven surviving features into rewrite rules required three constraints.

**Whitelist principle.** Only the eleven listed items are processed. Sentences matching no rule are preserved verbatim; heading levels, paragraph order, lists, tables, and code-block positions are left unchanged.

**Information conservation.** No names, numbers, dates, quotations, sources, causal relations, or any detail absent from the source may be introduced; equally, no viewpoint, conclusion, or hedge may be removed. Changing "may improve" to "improves" falsifies the strength of a claim and is not style cleanup. The test: every content word after rewriting must be traceable to the source.

**Exclusion of semantic judgment.** Whether a metaphor is apt, whether a question addresses a real gap in the reader's understanding, whether parallel cases constitute padding — all three depend on semantic understanding and cannot be transcribed into executable operators. Writing them into rules would only license rewriting by subjective feel, so all are excluded.

Full rule set in [SKILL.md](./SKILL.md).

### 4.1 Application

The rule set ships as an Agent Skill. We recommend opening [`lieflat-less-ai-tone` directly on MoxtHub](https://moxt.ai/hub?type=skill&id=lieflat-less-ai-tone), where Moxt's large context window and AI-native file workflow make it easier for an agent to process complete long-form text.

It can also be installed locally:

```bash
npx skills add larashero3-dotcom/lieflat-less-ai-tone
```

Once installed, submitting text applies the rule set. `SKILL.md` also works directly as a system prompt in any tool accepting custom instructions.

When used alongside writing-style distillation there is no need to install this rule set separately; [writing-dna-skill](https://github.com/larashero3-dotcom/writing-dna-skill) bundles a copy, so installing that repository brings it along. The two form successive stages: the former approximates a target style, this rule set removes generation artifacts. Where distilled artifacts are present in the same directory, read `language-dna.md` first; where the two conflict, the artifacts take precedence, since they record how the target author actually writes and are not generation artifacts.

## 5 Record of measurement errors

Six systematic errors occurred at the operator-design stage. In every case the operator's coverage was broader than the rule's definition, and in every case the frequency output was accepted before matched instances were inspected.

| Initial result | Defect | After correction |
|---|---|---|
| Overlong pre-modifiers 3.04 | Operator admitted ordinary constructions | 0.35 |
| Coordinating conjunctions 0.26 | Rule specifies "twice or more within one sentence"; operator measured single occurrences | 0.00 |
| 的…的…的 nesting 0.25 | Operator matched across enumeration commas, counting ordinary coordination as nesting | 0.06 |
| Cue colons, no difference | Of 6,431 matches: 1,266 headings, 1,225 list items, 207 dialogue, all explicitly exempted by the rule | *R* = 3.8 |
| Interrogative subheadings 32× | Denominator was character count, ignoring that the generated side has 5–10× as many subheadings | No difference |
| Figurative packaging 0.000 | Operator was a fixed list of 11 lexical items; actual figurative expressions matched none of them | *R* = 0.42 (human higher) |

Publishing this record serves two purposes. First, four of the six would have entered the rule set uncorrected, two of them pointing the wrong way, such that rewriting on their basis would have moved text further from human writing. Second, these errors share a structure: a coverage gap between the operator and the rule it is named for, which the frequency figure itself does not reveal. The corresponding working rule appears in 2.6.

## 6 Limitations

**Limited human-side corpus.** Cross-group variation is already substantial (a hundredfold range for em dashes, threefold for particle density). A larger corpus may change some conclusions.

**Some features lack a human-side comparison.** The translationese filter (3.4) rests on generated-side absolute frequency against a threshold, without computing ratios. A construction frequent on the generated side is not thereby infrequent on the human side.

**Semantic-level features cannot be measured.** Judgments at the discourse and rhetorical level (whether a question is warranted, whether a metaphor is apt, whether cases are padding) fall outside what regular-expression operators can reach, leaving only locatable formal features. If the principal component of "AI tone" resides at the semantic level, this method cannot reach it in principle.

**Topics are not strictly paired.** The generated side was produced in a single batch over one topic set; the human texts span a longer period and a wider range. Part of the observed difference may derive from topic rather than text source.

**Findings are time-limited.** Model iterations continue converging toward human style. GPT's em-dash frequency fell from near one per sentence to 0.11 in under a year. All frequency values here should be read as observations at a particular point in time.

**The corpus is not verifiable.** The corpus is withheld for copyright and privacy reasons, so the values above cannot be independently checked. What can be checked is the segmentation procedure, denominator selection, decision thresholds, and operator definitions, all published with the scripts. This is a substantive defect, not a disclaimer.

## 7 Recomputation

```bash
# Sentence level: per-feature frequency and ratio across both sides
python3 scripts/compare-human-ai.py --human <dirs...> --ai <dirs...>

# Paragraph level: adjacent-sentence isomorphism, zero-anaphora openings
python3 scripts/check-structure.py --human <dirs...> --ai <dirs...>

# Generated side: translationese construction frequencies
python3 scripts/check-translationese.py <dirs...>
```

All three scripts read `.md` files from directory arguments only. They contain no corpus data and do not use directory names as output labels. Operator definitions sit at the top of each script.

Substituting your own corpus recomputes every metric. If results diverge from those reported here, check first whether topic and genre are aligned across sides, and whether the denominator matches the unit being measured.

## About Moxt

[Moxt](https://moxt.ai) is a platform for building an AI-native team that takes on long-horizon work. It has four main parts:

- **Workflow** — split complex work across specialist agents and let it run for days rather than minutes, unattended
- **Agent Board** — see what's done, what's running, and what needs you; step in to steer at any point
- **MiniApp** — build the internal tools a team actually needs without writing code, backed by real databases and used by humans and agents alike
- **Agent Native Workspace** — a file system of Markdown, HTML, CSV, and structured files that keeps agents in context across sessions

This skill depends on rereading long documents and rewriting them sentence by sentence, which is what the workspace's context handling is built for.

> [**Open Lieflat Less AI Tone on MoxtHub →**](https://moxt.ai/hub?type=skill&id=lieflat-less-ai-tone)

---

**Data and code** [scripts/](./scripts/) · **Full result tables** [RESEARCH.md](./RESEARCH.md) · **Rule set** [SKILL.md](./SKILL.md) · **MoxtHub** [less-ai-tone.skill](https://moxt.ai/hub?type=skill&id=lieflat-less-ai-tone) · **中文** [README.md](./README.md)