How our AI detection works
We split your text into overlapping windows of a few sentences and run each through a classifier trained on paired human and AI writing. Each sentence gets a probability from the windows that cover it. The document score is the share of words in sentences we flag.
Lists, tables, code, headings and bibliographies are excluded automatically, because they don't behave like normal prose and cause false positives.
What the colors mean
Amber highlights are sentences that read as AI-generated. Dashed amber marks text that looks AI-generated and then reworded by a paraphrasing or humanizing tool. Unhighlighted text reads as human-written.
Accuracy, honestly
No detector is perfect. False positives are more likely with short texts, heavily edited formal writing and writers whose first language isn't English. We calibrate our threshold on held-out human writing to keep false positives low, and we'll publish the measured rate with every model version.