让 AI 文字，读起来像人写的。 全部在你的设备上完成重写。
免费的 AI 写作与重写工具。粘贴任意草稿——它会给机器感打分，在你的浏览器里完整重写，并逐词标出每一处改动。不上传，无需注册。
写得更好，依然是你的。
三套重写引擎都在你的设备上运行，草稿永远不会发往服务器。
100% 端侧 · 私密
重写通过 Chrome 内置 AI、本地端侧模型或离线规则引擎在浏览器中完成。任何内容都不会被上传，永远不会。
看得见的分数
句子节奏、可预测性、词汇、主动语态、开头多样性与 AI 措辞——六项信号合成一个分数，改写前后同时呈现。
逐词对照
每一处改动都会相对原文高亮显示。不会有文字在你背后被悄悄改写，保留什么由你决定。
四种模式，三种语气
人性化、润色、扩写或精简。可选学术、商务或口语，也可以完全保持原有语域。
粘贴、打分、重写。
- 01
粘贴你的文字
放入草稿、文章或一段模型输出。本地即时分析，无需下载任何东西。
- 02
读懂分数
面板会指出哪些信号显得机器化，并标出暴露它的具体措辞。
- 03
重写并对照
选择模式与强度，看着分数下降、对照高亮填充。复制、下载，或送回编辑器。
About the AI Humanizer
A free writing tool for anyone whose draft reads too much like a machine wrote it. Students polishing an essay, marketers editing generated copy, non-native speakers smoothing a translation, developers rewriting documentation — paste the text, see how machine-like it scores, and get a rewrite you can check word by word.
Everything happens on your own device. There is no server in the pipeline, no account, and no queue, so drafts that are confidential or unfinished stay on the machine you wrote them on.
How it works
Three engines sit behind the same interface. The offline rules engine is pure deterministic text processing — it rewrites AI-tell phrasing, breaks uniform sentence rhythm, varies openers, and shifts passive constructions to active, with no model and no download. Where the browser exposes Chrome's built-in AI (Gemini Nano), the Rewriter API produces a more natural rewrite, again with nothing to download. Finally, an optional local language model (Qwen2.5 1.5B or SmolLM2 1.7B, quantized to 4-bit) runs through Transformers.js on WebGPU for the highest-quality output — downloaded once, then cached for offline use.
Long documents are split into overlapping-by-sentence chunks so a ten-page draft is processed part by part with a live progress indicator, and the result streams in as it is produced.
Reading the AI score
The score panel combines six independent signals: sentence-length rhythm (burstiness), phrase predictability, vocabulary range, active-voice ratio, sentence-opener variety, and the presence of known AI-tell vocabulary. Each is shown as its own bar so you can see exactly which signal is dragging the score up, and the flagged-phrase list lets you swap a single term for a suggested alternative without rewriting the whole paragraph.
Treat this as a writing signal, not a verdict. Detectors disagree with each other and with themselves; a low score means the text carries fewer of the surface patterns they look for, not that any particular detector will clear it.
Modes, tones and intensity
Humanize strips the machine rhythm, Polish fixes flow while keeping your voice, Expand adds supporting detail without inventing facts, and Shorten cuts redundancy. Tone can be left as-is or pushed toward academic, professional, or casual. Intensity controls how far the rewrite is allowed to move from your original wording — light stays close, aggressive restructures freely.
Limits & requirements
Every engine is told to preserve facts, numbers, and names, but a language model can still drift; the word-level diff exists so you can verify each change before you accept it. The local model is roughly 1 GB and needs a reasonably modern device — if it fails to load, the tool falls back to the rules engine and tells you it did, rather than failing. The AI-score metrics are tuned for English and Chinese; other languages still get full text statistics, just without the phrasing analysis.
Privacy
Your text is analysed and rewritten inside your browser tab. Nothing is uploaded, stored, or logged, and there is no word limit or paywall. Once the optional model is cached, the tool keeps working with the network switched off.