humanize-chinese · open source · pure Python
We measured six Chinese genres against five current models. The detector is still useful on short explainers, already weak on academic abstracts, and on long-form blogs the direction inverts — human writing scores higher than the machine's.
So the goal here is text that reads like a person wrote it,
not a smaller number.
Say it plainly
August 2026. Claude Opus 5, GPT-5.6, DeepSeek V4, GLM-5.3 and Kimi K3 each produced 50–100 samples per genre, set against 50 public human texts, recomputed after matching length on both sides.
| Genre | Separation | Verdict |
|---|---|---|
| Short explainer / news | 0.837 | still useful |
| Social posts | 0.771 | usable |
| Academic abstracts | 0.621 | already weak |
| Fiction | 0.479 | a coin flip |
| Long-form blogs | 0.155 | inverted |
Read it as: pick one machine text and one human text at random — this is the probability the machine one scores higher. The tick is 0.5, a coin flip. Below that, the direction is reversed: asked to pick "the one that looks AI-written", it picks the human more often. The long-form row has a thinner sample (15 machine vs 22 human) than the others, but the direction is unambiguous.
60 real CNKI abstracts published before large language models existed average 60/100, and 35% land in the "almost certainly AI" band. Chinese academic prose is full of set connectives; the detector cannot tell convention from generation.
Short expository writing is a genuine exception — don't throw that out. But using it to judge whether a long blog post or a novel chapter was machine-written gives you answers that are wrong in a systematic direction.
Still readable afterwards
Every rewrite passes two gates, and failing them means the text comes back untouched.
| Measured (1–5, three judges; a second round when the first is not clean-cut, unioned) | Fluency | Broken sentences found |
|---|---|---|
| This tool · five exaggerated in-repo samples | 4.0 – 5.0 | none found |
| Previous rule-only rewriter · same samples | 2.0 | 5 – 17 |
| This tool · real model output, 3 texts | 4.0 – 5.0 | 1 total |
Two of the three real-model texts came back with nothing flagged; one had a single unbalanced coordination. The table says none found rather than zero on purpose — review misses real defects some of the time, so “none found” reports what this round caught, not what exists. Three samples is not a pass rate — but the direction is clear: real text is harder to rewrite than caricatured text, and the tool is not defect-free.
Install
git clone https://github.com/swaylq/humanize-chinese.git
| Tier | Command | API key |
|---|---|---|
| Offline | ./humanize rewrite file.txt | none |
| Boilerplate removal | ./humanize rewrite file.txt --llm | required |
| Draft from scratch | ./humanize write "…" | required |
Pure Python standard library, zero dependencies. Also installs as a Claude Code skill. Licence is MIT Non-Commercial — free to use, modify and redistribute; commercial use excluded.