Is it free?
Yes. The checker is free to use and does not require an account.
Is my text stored?
No. Your text is processed locally in your browser and is not sent to our servers.
How accurate is it?
No AI detector is consistently accurate across every kind of writing. The main model handles modern and casual AI writing better than older detectors, but results are still estimates and can be wrong in both directions. On phones, tablets and low-memory devices the smaller Compact model is suggested by default (you can still choose Full); it catches older, formulaic AI writing much more reliably than modern, natural-sounding AI writing.
Which model does it use?
A DeBERTa-v3-large classifier by Shantanu Thorat, developed for the PAN 2026 AI-detection competition and converted to run in the browser. It is about 408 MB and downloads once, then your browser caches it.
Why was my human writing flagged?
Formal, academic, technical, repetitive, or highly polished writing can resemble patterns learned by the classifier. A flag is not proof of AI use.
What models does it detect?
The classifier was trained to distinguish examples of human and machine-generated text. It cannot reliably identify which specific writing model produced a passage.
Does it work on short text?
Texts under 80 words are not checked. Texts of 80–149 words are labelled limited evidence, because the classifier has fewer language patterns to evaluate.
Can it be fooled?
Yes. Editing, paraphrasing, casual style, and other changes can reduce detection. We do not recommend treating any result as a guarantee.
Does it detect every AI model?
No. Writing systems and editing techniques change quickly, and a classifier may perform differently on text unlike its training data.
Which languages are supported?
The underlying model is intended primarily for English. Results in other languages should not be relied upon.
What do the percentages and highlight colours mean?
Each percentage is the share of analyzed words in sentences whose model signal crossed a threshold: 85 or higher for strong AI-like signals (coral) and 50–84 for possible ones (amber). 50 is the model's normal decision boundary, not a validated threshold. The overall AI-pattern signal (0–100) is the word-weighted average of every sentence's signal. None of these numbers are the probability that someone used AI, and they are not proof.
Is there a length limit?
Yes. You can check up to 10,000 characters at a time (roughly 1,500–1,700 words). Word documents can be up to 5 MB and text files up to 1 MB. For longer documents, check them in parts.
Why is the first check slower?
Your browser must download the open-source model files. It may cache them, so later checks are usually faster.
Are uploaded files sent anywhere?
No. Text and Word files are read locally in your browser. Only their extracted text enters the checker.
Should educators use this as evidence?
Never as sole evidence. Review drafts, version history, sources, prior work, and the student's explanation of their process.
Does this prove who wrote something?
No. A statistical classifier cannot establish authorship, intent, or misconduct.
Want the technical overview? Read How it works or view our research and sources.