Three steps, entirely on your device
- 1. Split
Your writing is divided into overlapping passages of about 240 words, aligned to sentence boundaries.
- 2. Score
A DeBERTa-v3-large classifier by Shantanu Thorat, from the PAN 2026 AI-detection competition and converted for the browser, scores each passage on your device. Each sentence takes the median signal of every passage that covers it.
- 3. Highlight
Sentences with a signal of 85 or more show a strong AI-like signal; 50–84 show a possible one. The possible tier starts at 50, the model's normal decision boundary, so borderline passages aren't hidden. Neither threshold has been validated on our own benchmark, and crossing one is not proof.
Your text stays local
The model (about 408 MB) is downloaded from Hugging Face the first time you run a check and cached by your browser. Your writing is processed locally, not uploaded to us. On devices with limited memory, or if the main model can't load, a smaller lightweight model is used instead; it is less accurate on modern AI writing.
What the two tiers mean
The strong tier (signal 85+) and possible tier (50–84) are reported as separate percentages: the share of analyzed words in sentences that crossed each threshold, plus the share below 50. You also see an overall AI-pattern signal from 0 to 100: the word-weighted average of every sentence's signal. It is a model signal, not the probability that someone used AI. For documents of 150–1,500 words, the text is also scored in shorter passages of about 100 words, and each sentence takes the median across both passage sizes. This helps surface a single edited or mixed paragraph without letting one odd passage dominate. Texts under 80 words are not checked, and texts of 80–149 words are labelled limited evidence.
Limitations
AI detection is probabilistic and no classifier is right every time. Formal or academic human writing can still produce false positives. Heavily edited or paraphrased AI writing may be missed, and the lightweight fallback model catches older, formulaic AI writing much more reliably than modern, natural-sounding AI writing. Short texts give the model less context and are less reliable.
Use the result as one signal, never as proof. Read our frequently asked questions for practical guidance.