Screen text for AI-like writing signals in your browser. The text stays on your device; only public model files are downloaded.
This is the pinned model's precomputed result for the exact example text, so the model is not downloaded on first visit. Click Check text to run a fresh local analysis.
This is a screening signal, not proof of authorship. Editing, translation, paraphrasing, domain differences, and new models can change the result.
Use this tool as a first-pass signal when reviewing essays, articles, support replies, or other text. It does not claim to identify authorship with certainty.
The current page language is selected by default, and you can change it when the text uses another language. A language must be validated before the tool can show a confident result.
Use at least 120 characters. Longer passages give the model more context and allow multiple windows to be compared.
The browser downloads and caches the model, then analyzes the text locally.
Use the output as one signal alongside source history, drafts, citations, and human review.
These fixed examples were run with the pinned browser models. They illustrate how wording and language can change the signal; they are not universal accuracy benchmarks.
今天早上出门时才发现昨晚下过雨，楼下那块地砖还湿着。我本来想直接去地铁站，走到巷口又临时买了一杯豆浆。店里的阿姨正在数零钱，后面排队的人也不着急，大家就站在屋檐下聊了几句天气。到公司以后，鞋边还是沾了一点泥。下午回家时雨又下了一阵，我只好把伞往旁边偏，结果左边袖子还是湿了一小块。
Specific details and casual sequencing produced a low signal in this run.
有创意的在线小众图片工具集 ROCKIMG提供小众的，有趣或实用的与图片相关的创意或工具，无需注册，免费使用。 分类模型会将文本中的语言模式与训练时看到的人工文本和 AI 文本模式进行比较。结果是模型信号，不是作者身份测量。 这只是初步检测信号，不是作者身份证明。编辑、翻译、改写、领域差异和新模型都可能改变结果。
This is a known false-positive risk pattern. A high score is not proof of AI authorship.
I left the house early and walked beside the river before the streets became busy. The air was cool, and the small change in routine made the morning feel calmer.
The English baseline gave a lower signal, but it is still not proof of authorship.
English uses the original English-only RoBERTa baseline. Simplified Chinese uses the multilingual q4 browser model in the current supported prototype path. Traditional Chinese and Japanese use the same model as experimental signals because they have not been separately validated.
The model estimates an AI-like signal from distributed language patterns such as token choices, sentence transitions, phrasing, repetition, and rhythm. It does not read hidden authorship metadata.
The score comes from one model and one data distribution. Translation, editing, paraphrasing, prompts, domain jargon, and newer generators can create false positives or false negatives.
Language selects the tokenizer, calibration, and validation profile. The page defaults to its current language, while experimental languages are clearly marked until a detector model is validated.
A classifier compares language patterns in the passage with patterns learned from human and AI-generated examples. The result is a model signal, not a measurement of who wrote the text.
English uses a validated ONNX baseline. Simplified Chinese uses the current multilingual browser model, while Traditional Chinese and Japanese remain experimental until language-specific evaluation is available.
No. Paraphrasing, translation, editing, prompts, model updates, and domain differences can create both false positives and false negatives. Do not use one detector score as disciplinary, employment, academic, or legal proof.
Yes. The browser-based prototype is free to use. Your browser downloads the public model files and performs the analysis locally.
No. The entered text is not sent to the application server. Only model and tokenizer files are fetched from the configured public model host.
The browser needs to download the quantized ONNX model and tokenizer. The model can be reused from the browser cache after the first successful load.
Yes. Simplified Chinese uses the current multilingual browser detector path. Traditional Chinese and Japanese use the same model as experimental signals because they have not been separately validated, so none of these results should be treated as authorship probabilities.
No. Treat it as an initial screening signal only. For important decisions, inspect drafts, sources, edit history, and other provenance evidence.
Review the public source code, model cards, conversion scripts, and pinned browser model files behind this tool.