Detecting AI in writing is much harder than in images. Only style can be judged, and human editing or translation mixes the traces easily. That is why Origma keeps text detection in beta and states first where it goes wrong.
The text detector reads statistics generation models leave, such as how predictable word choice and sentence flow are. The box above shows the text gate: it catches AI text at a high rate but also flags a fair number of human texts.
Human texts that went through machine translation often read like AI text and make up most false alarms. Short texts and quote-heavy texts lack the material to judge style and are returned as inconclusive. Pasted Korean text uses a stricter threshold against flagging people.
A human text with one or two AI-continued sentences is counted separately. Such texts read as human overall and are easily missed, so the accurate reading is that they may be mixed.
Never punish or reject someone on a text result alone. Use it to start a conversation, together with other evidence such as the writing process, draft history or a short oral check.
This is general guidance, not legal advice. A scan result is a probabilistic assessment, not proof.