Fully local processing No account required
↗ Source

Understand your writing.
Keep it private.

Run a trained AI-text classifier entirely in your browser. It shows a model score, not proof of AI authorship. First use downloads ~125 MB of model weights and a JavaScript inference runtime; your text is never uploaded.

Your text
Paste, type, or bring your own document.
100% ON DEVICE
0 words · 0 charactersNothing saved

Drop documents here, or
TXT · Markdown · DOCX · DOC · RTF · multiple files

Analysis report
Explainable signals, not a definitive verdict.
NO CLOUD AI

Ready when you are

Your local model score and writing insights appear here after analysis. English passages of at least 120 words are required.

Batch documents

Select or drop multiple supported files to score each independently. Switch documents above to inspect their independent on-device model scores and style highlights.

Writing insights
—Words / sentence
—Length variation
—Speaking time

Most common content words

Analyze text to see keywords.

How the on-device model works

NoGPT v4 uses TMR, a RoBERTa-base text classifier trained on the RAID benchmark. Its ONNX quantized weights (~125 MB) and Transformers.js runtime are downloaded from Hugging Face and jsDelivr when you first run an eligible analysis. These hosts can see standard asset-request metadata (such as your IP address), but no pasted text, imported document contents, or inference requests are transmitted to them. Analysis happens with WebAssembly in your browser, without our server or login. The score is an uncalibrated model softmax average, not a proven authorship probability. Long documents are split into passages; sampling coverage and agreement are reported. Inputs under 120 words and apparently non-English input receive no numerical score rather than a fake 50%. Passages with different styles, new AI models, unusual topics, editing and translation can fool the detector. Highlighted style signals are separate from the classifier and must not be interpreted as evidence of which passages were AI-written. Its publicly reported benchmark metrics were produced under specified RAID conditions and do not prove accuracy on your documents. NoGPT has not been independently validated on a fresh out-of-distribution benchmark.

Do not use these results as sole evidence of AI use in school, employment, publishing, or any consequential setting.