When the score lands high on writing you have reason to trust.
False positives happen. They happen most often on translated text, heavily edited writing, ESL register, very short input, and heavy genre conventions. This page is the working protocol for what to do next.
What to do when a passage you trust scores in the AI range.
Five steps. None of them is “auto-reject.”
- 01ReadRead the sentence highlights
A high aggregate score on a passage with one or two flagged sentences is a different story than a uniformly high passage.
- 02ReadLook at the confidence band
Wide confidence band = less commit. The detector is telling you it isn’t sure.
- 03ReadCheck input length
Very short input increases the false-positive rate. Slice the passage if needed and rescan a longer sample.
- 04TalkAsk the writer two specific questions
See the protocol below. The questions surface process; the answers are part of the evidence.
- 05DecideDocument the call
Whether you publish, hold, or send back, write down the call and the rationale. Save the report to the file.
Where the detector’s assumptions interact with real human writing.
Translation can flatten lexical diversity and shift sentence-length variance toward the AI baseline. Run on the original language when you can.
Multiple edit passes can converge on a registered, model-typical voice. The piece is human-written; the editing is doing the work.
Detection signals can interact with non-native English register. Pair with a conversation, especially in education contexts.
Confidence widens for short input. The aggregate score is less reliable below ~120 words for text.
Academic, legal, and regulatory writing compress lexical diversity by design. The genre is the work; the writer is doing it correctly.
A reviewer who saw mostly opinion writing this morning will read a regulatory submission as model-typical. Trust your tools, not your hot takes.
The two questions to ask the writer
When the score is high on a piece you have reason to trust, ask two specific questions before drawing a conclusion:
- Walk me through how you wrote it. Where did the structure come from? Did you draft from notes, or start fresh in the document? What did the second pass change?
- Where did the phrasing on these sentences come from? Cite two specific sentences from the sentence highlights. Listen for whether the writer can talk about their own writing.
These questions surface the writing process. A writer who wrote the piece can answer them. A writer who didn’t will avoid them. The score is the evidence; the conversation is the verdict.
What not to do
A detection score is probabilistic, not proof. Treat the report as evidence to discuss, not a verdict to publish.
- Don’t auto-reject on score alone, even a 95% score has a confidence band.
- Don’t share the raw score with the writer without context about what it means.
- Don’t escalate to formal proceedings on a single high score.
- Don’t apply the threshold inconsistently across a cohort.
- Don’t hide the methodology. Students and contributors deserve to know how the tool works.