What Turnitin’s *% Score Actually Says

Ai Detection In Education

Share free access to this member-only story with a friend: Read it free here.

A Turnitin report can replace a low AI-writing percentage with *%. The symbol does not mean zero, and it does not mean the system caught a student cheating. It means the score falls somewhere from 1% through 19%, while the exact number stays hidden.

That distinction matters when an instructor must decide whether to examine the paper, speak with the student, or begin a misconduct case. Turnitin's Help Center guidance says false positives are more common in this range. It also warns that the model can mistake human writing for AI writing and should not be the sole basis for action against a student.

The asterisk is a reason to look closer, not a verdict about who wrote the paper.

What the percentage measures

Turnitin's percentage concerns the share of qualifying prose that its model classifies as likely AI-written. It is not the probability that the student used AI.

A hypothetical score of 12% would not mean there is a 12% chance of AI use. It would not establish that the remaining 88% came directly from the student. It would describe how the model classified parts of the text it could evaluate.

The model analyzes language in the submitted paper. It does not observe the student researching, drafting, revising, or receiving help. It does not know whether the assignment allowed a grammar assistant, prohibited generated prose, required disclosure, or said nothing about AI. It also cannot determine a student's intent.

Those limits become especially important when the report displays *%. Turnitin withholds the exact low-range figure because showing a number such as 7% or 14% could suggest more precision than the company believes instructors should rely on there.

An instructor should not silently convert the asterisk into 19%, the highest possible value. Treating it as zero would also be wrong. The honest reading is narrower: the model returned a low result that Turnitin considers particularly vulnerable to false positives.

The report cannot decide whether a rule was broken

Even a correct detection would not settle a misconduct case. The next question is what the assignment permitted.

One course might allow AI for brainstorming but ban generated paragraphs. Another might permit editing tools if students disclose them. A third might prohibit all outside assistance during an assessment. The same use could be allowed in one setting and violate the rules in another.

The instructor therefore needs the exact assignment instructions, syllabus language, and institutional policy that applied when the student submitted the paper. Standards added later should not be used to punish earlier work. If the instructions were vague, that uncertainty belongs in the review rather than being placed entirely on the student.

Educators already make a similar distinction with text-matching reports. A high similarity percentage can come from quoted passages, references, common phrases, or copied material. The number points toward text that needs interpretation; it does not decide whether plagiarism occurred. AI classification uses different technology, but the same restraint applies: software can identify a concern without resolving the academic question.

Review the paper and its history

A fair review looks for information that exists independently of the masked result. No single item automatically proves innocence or guilt, and missing evidence should not be treated as an admission.

Useful checks include:

Read the paper closely and verify how its claims, quotations, and citations connect to the listed sources. A false or irrelevant citation is a real problem, but it does not by itself prove AI use.
Examine outlines, notes, drafts, or version history when those materials normally exist and institutional rules permit the review. Some students draft offline, replace files, or paste completed sections into a final document, so an incomplete history is not conclusive.
Compare the submission with relevant earlier work, while allowing for tutoring, revision, changing subject knowledge, disability accommodations, and ordinary improvement.
Ask the student to explain the thesis, source choices, and development of a disputed passage. Give the student a clear account of the concern and a genuine chance to answer.

The conversation should begin with what is known: the report returned *%, the exact result is hidden, and the software can make mistakes. Neutral questions are more useful than an accusation.

An instructor might ask how a particular source shaped the argument, what changed between drafts, or which writing tools the student used. The goal is not to stage a surprise oral examination. Anxiety, language differences, or difficulty speaking under pressure can affect an answer without revealing who wrote the paper.

The student's explanation also needs context. A confident account is not automatic proof that no prohibited tool was used, just as a hesitant answer is not proof of misconduct. Its value comes from how well it fits the paper, drafts, sources, and other available records.

Consistency matters here. If one student must produce extensive document history because of a low detector result while other students are never expected to retain drafts, the software has created a new burden that the course never disclosed. Clear retention expectations should be set before submission and applied evenly.

When formal action makes sense

A misconduct process may be justified when independent evidence supports a specific violation. That could include a student's acknowledged use of a prohibited tool, reliable records showing prohibited assistance, or contradictions between the submitted work and its claimed sources that remain unexplained after review.

The evidence still has to match the rule. Poor citations can support a citation-related concern. Fabricated information can support a finding that the work is unreliable. Neither fact should automatically be relabeled as AI misconduct without evidence connecting it to the prohibited conduct.

If a case moves forward, the instructor should record the assignment rule, the report's displayed result, the software version or report date when available, the passages or records examined, the student's response, and the remaining uncertainty. The institution's own standard of evidence, notice requirements, privacy rules, and appeal procedures should control what happens next.

When *% is the only evidence, an accusation is difficult to justify. The report supplies no exact low-range figure, no knowledge of the student's actions, and no finding that a course rule was violated. An instructor can clarify expectations for future assignments without pretending the asterisk established what happened in this one.

Fair review keeps the rule credible

Taking a masked result seriously does not require treating it as conclusive. Ignoring every concern would weaken legitimate academic standards, but punishing students from a signal known to produce false positives creates a different problem. Honest students bear the cost, disputed cases multiply, and confidence in the misconduct process falls.

The more durable approach is to require evidence that can answer two separate questions: what assistance was probably used, and whether that assistance violated a rule the student could reasonably understand. A detector may contribute to the first inquiry. It cannot answer the second.

Before acting on *%, the instructor should be able to name the applicable rule and the independent evidence supporting the concern. If either is missing, the review is not ready to become an accusation.

The asterisk does not clear the paper, and it does not condemn the student. It tells the instructor that certainty is low and judgment still belongs to the human beings responsible for the course.