When you open the AI Writing Detection result in Turnitin, what you see is often not a clean percentage but an asterisk: *%. The first reaction is usually "is it broken?"
No. This is designed behaviour.
How AI detection works
Unlike similarity matching, AI detection does not search a database for identical text. It analyses statistical properties of the writing sentence by sentence — how predictable the word choices are, how much the sentence structure varies — and estimates the probability that each sentence was generated by a large language model, then aggregates that into an estimated proportion.
The key point: this is a probabilistic estimate, not evidence. The system has no way of knowing what tools you actually used. It only says "these sentences have statistical properties close to generated text".
Why you get *%
1. The document is too short (under ~300 words)
Turnitin requires roughly 300 words of continuous prose for AI detection. Below that the sample is too small for the estimate to mean anything, so the system declines to give a number.
Note that the word count is of prose — bullet lists, tables, code, formulae and reference lists generally do not count. So an 800-word report that is mostly bullet points can still trip this limit.
2. The estimate is below 20%
According to Turnitin's own documentation, the false positive rate rises noticeably when the estimated AI proportion is below 20%. Rather than show a 5% or 12% that a marker might easily over-read, they display an asterisk to indicate "a small amount detected, but not with enough confidence to report a figure".
So *% is generally good news for you, not bad — it means no substantial AI-generated content was detected.
3. Unsupported language or format
AI detection covers a much narrower set of languages than similarity matching. If your document is not in a supported language, or the format prevents clean text extraction, no figure is produced.
Who gets false positives
Because the underlying question is "how predictable is this writing" rather than "was this written by AI", human writing in these situations shares statistical properties with generated text:
- Non-native English writers. Conservative vocabulary, repeated sentence patterns and close adherence to a template are exactly the low-variance signals the model looks for.
- Highly formulaic sections. Methodology, experimental procedure, standard definitions.
- Text polished — not written — by AI. Even if the argument and structure are entirely yours, sentences rewritten by a model carry its fingerprints.
Put differently: a high score is not proof of cheating, and a low score is not proof of innocence. This is a shared limitation of every AI detection tool today, and it is why many institutions specify that an AI score cannot on its own be the basis of a misconduct finding.
What to do if you are flagged
- Keep your writing history. Drafting in Google Docs or Word online records version history automatically, showing how the piece was built. This is currently the strongest evidence you can offer.
- Keep drafts, outlines and notes. Photographs of handwritten notes count.
- Be ready to talk through your argument. If you are asked to a meeting, being able to explain on the spot why you chose your framework and how you reached your conclusion is more persuasive than any tool report.
- Read your course rules before using AI. Many courses permit AI for research or grammar checking but prohibit generated content. The line is set by the course, not by your guess.
Want the sentence-level marks?
The headline percentage tells you how much, but not which sentences. TurnCheckHK supports requesting the AI highlight report, so you can see exactly what the system marked and revise those passages specifically rather than rewriting the whole thing blind.