lieflat less ai tone

Measuring AI writing patterns through linguistic experiments.

lieflat
less ai tone

An AI writing refinement skill built on 5.2 million Chinese characters and 500,000 judgments.

Less AI tone.
More
evidence.

As AI gets better at coding, its writing does not seem to be improving at the same pace. The generated text still reads strangely.

Many leading models are highly capable, yet I often have to read their longer pieces several times. Much of the difficulty comes from repeated “not this, but that” contrasts, dense em dashes and semicolons, and metaphors I cannot make sense of.

So I want to use linguistic experiments to turn the vague sense of “AI tone” into writing patterns we can examine one by one.

The experiment covers 5.2 million Chinese characters of human writing and AI-generated texts on the same topics, with 500,000 model judgments and extensive human blind evaluation. The project is still evolving. Stay tuned.

Selected significant features of AI writing

Writing pattern	Observed evidence	How to use the finding
“Not A, but B”	δ +0.88 / +0.72	Repeated contrasts; check whether the contrast carries meaning.
Renaming the same idea	MTLD δ +0.91 / +0.55	More lexical variation. Keep precise terms; avoid needless renaming.
More, shorter sections	Median 7 / 1	Median section counts: AI / human. Report structural patterns without automatically restructuring.
Uniform sentence lengths	δ −0.88 / −0.40	Lower sentence-length variance. Observe the whole draft; do not manufacture variation.
Long pre-noun modifiers	δ +0.48 / +0.54	Modifiers pile up before the noun. Keep every qualification when reordering.
Conclusion-leading colons	AI 2.1×	Check lead-ins such as “the key is:”, rather than removing every colon.
Over-explicit transitions	98.9% / 95.2%	Explicit links in AI / human writing. Judge redundancy in context.
Dense lists and enumeration	AI 1.45×	7.845 / 5.398 per 1,000 characters. Unpack crowded lists; keep useful ones.
Whether a metaphor fits	Human review	Model agreement was only 0–40%. Keep original metaphors by default and leave the judgment to people.

Paired δ values are GPT-5.6 / Claude Opus 4.6, in that order. Measures use different methods; they are not comparable “AI probability” scores.

Different models. Different habits.

Writing pattern	GPT-5.6	Claude Opus 4.6	What it means
Em dashes	−0.54	+0.95	Opposite directions; no blanket removal.
“的” per clause	−0.75	+0.62	Less in GPT, more in Opus.
Quotation marks	≈ 0	−0.67	Less frequent in Opus than in human writing.
Semicolons / list commas	+0.80 / +0.56	−0.11 / +0.14	The stronger differences appear in GPT.
Indirect sentence structures	3 significant	3 not significant	Observations of single-verb sentences, stative sentences and subject–verb distance.
Questions	−0.59	−0.12	Fewer questions in GPT; Opus is closer to human writing.
Long / medium sentences	−0.68 / +0.81	+0.05 / +0.04	GPT sentence lengths cluster in the middle.
Delayed focus	+0.05	+0.47	Opus more often places new information at the end.
Active metaphor density	2.099	2.851	Occurrences per 1,000 characters; human baseline 2.066. This row is not δ.

δ is Cliff’s effect size: positive means more in AI writing, negative means less. The final row uses metaphor counts per 1,000 characters. These observations concern specific Chinese corpora and model versions; they do not establish who wrote a text.