We've written before about what makes a similarity score "bad", and the AI score deserves the same honest treatment — because it behaves completely differently, and applying similarity-score intuition to it is how students misread their situation. The uncomfortable headline: there is no institutional consensus on an "acceptable" AI percentage, and unlike similarity, there's no innocent-by-default explanation for a high one. But the ranges do mean things, and knowing what they signal to a marker is worth a lot.
First: why low scores show as an asterisk
If a document's AI score comes back under roughly 20%, Turnitin doesn't display the number — it shows *%. That's a deliberate uncertainty flag: Turnitin's own data showed false-positive rates rise in the low range, so it declines to state a precise figure there. Practically, an asterisk is the best result you can get, and any argument about whether you're "really" 6% or 14% is an argument Turnitin itself refuses to have. (More on the report's mechanics in how to read the AI writing report.)
What the ranges tend to signal
These aren't official thresholds — institutions set their own policies, and many instruct markers that no score alone proves anything. But as a description of how reports get read:
- *% (under ~20). Effectively unflagged. Turnitin explicitly tells instructors to treat this range with caution because of false-positive risk.
- ~20–40%. A "look closer" signal, not a verdict. Markers check which passages flagged: summary and transition paragraphs flagging while the argumentative core reads clean is a very different picture from whole sections lighting up.
- ~40–70%. Uncomfortable territory. At this level most markers will look at the highlighted text seriously, compare it against your known writing, and may open a conversation.
- 70%+. The report is claiming most of the document reads as machine-written. If that's wrong, you'll want your receipts — drafts, version history — ready.
Note what's missing from that list: any percentage that's "fine" because your school published a threshold. Similarity thresholds exist at many institutions; AI-score thresholds mostly don't, because the number is a probability estimate, not a measurement.
Why "acceptable" is the wrong frame
A similarity score has innocent explanations at almost any level — quotes, bibliographies, templates. An AI score's only innocent explanation is the detector being wrong, which absolutely happens (here's the accuracy evidence, and who it happens to most) — but that's a defense you make, not a percentage you point to. So the practical question isn't "what number is acceptable?" It's "can I account for every highlighted passage?" If the flagged text is yours, you can defend it and revise it. If it isn't, no threshold will save you.
If you're checking before submission
This is the score's one great property: it's fully knowable in advance. Run the document through the real Turnitin and you get the same AI report a marker would see — percentage and highlighted passages — with time to rework anything flagged. An asterisk on Tuesday means a calm submission on Friday; a 45% on Tuesday means you found out while it was still an editing problem.