What Turnitin AI Score Counts as Bad?

There is no official "bad" Turnitin AI score

Turnitin has not published a specific percentage that counts as proof of AI-generated writing, and it explicitly does not want the score used that way. The percentage is the AI writing indicator, an estimate of how much of a submission Turnitin's model flags as resembling AI-generated text. It is an estimate, not a certainty score, and Turnitin's own guidance for educators states the number should prompt a conversation before any accusation.

What the percentage measures

The AI writing indicator estimates the share of the submitted text that resembles patterns Turnitin's model associates with AI-generated writing, shown as a percentage on the originality report alongside the separate similarity, or plagiarism, score. A 20% AI score does not mean "20% chance this is AI." It means the model flagged roughly a fifth of the sentences as resembling AI-typical patterns. Those are two different claims, and confusing them is the most common misreading of the number.

Why the score has no clean threshold

Turnitin reports that its detection model performs best on longer stretches of continuous text, and that shorter documents, quotations, or bulleted lists can produce unreliable results, which is part of why Turnitin has resisted naming a specific "bad" cutoff. A student's formulaic, template-following writing style, common among English learners and younger writers still developing their voice, can resemble the statistical patterns the model associates with AI-generated text even when no AI was used. Treating any single number as a hard cutoff builds a false-accusation risk directly into your policy.

A reasonable way to read the score

Rather than a fixed cutoff, treat the AI score as a prompt for a closer look, weighted by how it compares to work you already have from that student.

• Low, single digits: usually not worth a conversation on its own.
• Moderate, roughly 20 to 50%: worth a look at which sections are flagged and whether they match the student's usual writing style.
• High, above 50%: worth a direct conversation with the student that asks about process instead of assuming guilt.

At every level, the score is one input alongside what you already know about how that student writes. It never stands alone as a verdict. If your school runs Canvas, our guide to what Canvas records without Turnitin covers the quiz logs and page views it does keep.

What to do with a flagged submission

Compare the flagged sections against work you have seen the student produce before, in class, under supervision, or in an earlier draft. Ask the student to walk you through their process: what they researched, how they organized their argument, why they chose specific wording. A student who wrote the piece can usually explain these choices without hesitation. A student who did not usually struggles with specifics. Our guide on whether AI detectors are accurate covers the broader accuracy and false-positive research behind this caution in more depth.

Building this into your policy instead of reacting case by case

The strongest protection against a false accusation is a policy set before any score exists. A decision made under pressure once a number appears comes too late to be fair. A clear AI syllabus statement that spells out how you use detection tools, as one signal that prompts a conversation and never as sole evidence, sets the expectation early and protects both you and your students when a score does show up.

What can push a score up without AI being involved

Several ordinary writing habits can raise an AI score even when a student never touched a chatbot. Heavy use of a grammar or paraphrasing tool can smooth a sentence into the kind of predictable phrasing the model flags. A student writing in a second language often leans on simpler, more repetitive sentence structures, which reads the same way to the model. Direct quotations, bulleted lists, and short technical or formulaic sections, a lab procedure, a math proof, a works-cited page, can also score unusually high because they carry little of the sentence variety a longer personal essay would. None of these facts make the score meaningless. They explain why a single number, read without context, misleads far more often than teachers expect.

How this compares to other AI detectors

Turnitin is not the only tool producing a percentage like this. GPTZero, Grammarly's own AI-detection add-on, and several other checkers each run their own model and can disagree with each other on the same piece of writing. None of them has published the kind of independent, peer-reviewed accuracy study that would justify treating one score as more trustworthy than another simply because of which company built it. A student flagged by one tool and cleared by another is not evidence the second tool is right. It is evidence that every one of these percentages is an estimate. None of them is a verdict, regardless of the name on the report.

Pair the score with a conversation before any consequence

Read a flag as a prompt to look closer, never as proof on its own. Compare it against the student's writing history, then have a direct, non-accusatory conversation about process before any grade or discipline action follows. Save an actual consequence for cases where the writing history and the conversation both point the same way, and use the syllabus language above to tell students how you will use the score before the first one ever gets flagged.