What Percentage of AI Is Acceptable? How to Read an AI Detection Score

Students ask what percentage of AI is acceptable as if there were a passing grade on a detector: under 20% and you are safe, over 50% and you are in trouble. There is no such number, and the reason is simple. A detector score estimates how much of your text looks machine-written. Whether any AI use is allowed, and for which steps, is a rule your course or institution sets, and most rules are written about what you did, not about what a detector shows. This page explains what a score on Turnitin and similar tools actually measures, why it is not a probability that you cheated, and where to find the rule that applies to you.

What an AI detection score measures

A detector measures text. A policy governs conduct.

Take Turnitin, the detector most students meet. Its AI writing detection FAQ says the percentage indicates the amount of qualifying text in the submission that its model determines was likely generated by AI, or generated and then modified by a paraphraser or bypasser tool. Qualifying text means prose sentences only; lists, bullet points and other non-sentence structures are left out. The FAQ adds that the percentage is not necessarily the percentage of the entire submission.

Under the hood, the same FAQ says, each sentence gets a score between 0 and 1, sentences over a threshold are labeled AI, and the document score is the share of sentences above the threshold. So a score of 40% is a count: about four in ten evaluated sentences crossed the line. It is not a statement that there is a 40% chance you used AI, and it is not a measure of how much of your thinking came from a tool.

What percentage of AI is acceptable, then?

The honest answer is the one your course gives. A course that bans AI for a take-home essay does not allow 10% of it; a course that allows AI brainstorming with disclosure may accept an essay whose detector score is high, as long as the student followed the rule. The number on the detector and the rule in the syllabus are answers to two different questions.

Detector makers say the same. Turnitin's guide to using the AI Writing Report says the model may misidentify human-written, AI-generated and AI-paraphrased text, so the score should not be used as the sole basis for adverse actions against a student, and that further scrutiny and human judgment, together with the organization's own policy, are needed to decide whether misconduct occurred. The FAQ adds that Turnitin does not make a determination of misconduct.

Two published university positions

Some institutions have gone further and switched the detector off. In an August 2023 post, Vanderbilt University explained why it disabled Turnitin's AI detector. It noted that Turnitin had claimed a 1% false positive rate at launch, and put that in context: Vanderbilt submitted 75,000 papers to Turnitin in 2022, so around 750 papers could have been wrongly labeled as partly AI-written. Its guidance to instructors was to state expectations clearly and, where AI use is allowed, to require disclosure and citation.

The University of Pittsburgh's Teaching Center takes a similar line on its academic integrity resource page. It recommends against AI detection tools, says they are not accurate enough to prove a violation, reports that it disabled the AI detection tool in Turnitin, and says it does not endorse or support any AI detection tools. Instead it advises instructors to state in the syllabus whether and how generative AI may be used, and to teach students how to cite it when it is allowed.

Neither page sets an acceptable percentage, and that is the point. Where a rule exists, it is about tasks: AI allowed for brainstorming, not for drafting; allowed with a disclosure note; not allowed at all. To find yours, read the syllabus, the assignment sheet and your institution's academic integrity policy, in that order, and ask if they disagree.

Why a single number misleads

Three features of how detectors work make any fixed cutoff unreliable, and Turnitin's own documents describe all three.

• Low scores are the least reliable. Turnitin hides scores from 1% to 19% because its testing found a higher incidence of false positives in that range.
• Scores undercount by design. To keep its stated false positive rate under 1% for documents over 20% AI writing, the FAQ says Turnitin accepts missing some AI text: a document scored at 50% could contain as much as 65%.
• Short pieces swing. In documents of a few hundred words, the FAQ says, predictions are mostly "all or nothing", so mixed text may be flagged as entirely AI.

Different detectors also use different models and thresholds, so the same paragraph can score very differently on two tools. For a fuller explanation of what any of them can and cannot establish, read what AI text detectors can tell you, and for Turnitin specifically, how Turnitin's AI indicator works.

A worked example: one essay, two courses

Here is an example built for this page, not a real case. A student writes a 1,200-word essay. They used a chatbot to suggest three possible thesis statements, picked one, rewrote it, then drafted the essay themselves. A detector reports 24%, with highlights on the thesis paragraph and two summary sentences in the conclusion.

Change one fact and the picture flips again: if the student had used no AI at all and still scored 24%, the useful evidence in either course would be the drafts and version history, not a second score.

How to check your own draft against the rule

1. Find the rule in writing

   Read the syllabus, the assignment sheet and your school's integrity policy. Highlight every sentence that mentions AI, tools or outside help.

2. List what you used, by task

   Write one line per tool: what it was, which step you used it for. Grammar checking, translation, brainstorming and drafting are different tasks, and policies treat them differently.

3. Compare the two lists

   Mark any use the rule does not clearly allow. When the rule is silent or vague, ask before you submit, not after.

4. Disclose what you used

   Add the note your course asks for. The method is in how to disclose AI help in your writing.

5. Keep your process

   Save drafts, notes and version history until the grade is final.

Some students also run a finished draft through a detector to see which passages read as generic before they revise. Walter Writes AI's AI detector is one option: Walter Writes AI is an AI humanizer with a built-in AI detector, supports over 80 languages, and includes voice matching and the detector on every plan. Its score is another estimate, not your school's rule, and not a forecast of what your instructor's tool will show. Detectors disagree, and no detector, Walter's included, can guarantee 100% accuracy.

Pitfalls when reading AI scores

• Treating a low score as permission. If the rule bans AI for a task, a low score does not make that use allowed.
• Treating a high score as a verdict. Turnitin says its score should not be the sole basis for action.
• Comparing scores across tools. Different models, thresholds and minimum lengths make the numbers incomparable.

Tips for students and teachers

• Students: ask early. One email before the deadline about which tools are allowed saves a hard conversation later.
• Teachers: write the rule by task. "AI allowed for X, not for Y, disclose in a note" is clearer than any threshold.
• Both: talk about the text. The highlighted sentences, the drafts and the student's own explanation say more than the percentage. More on fair review is in the AI writing guides.

How Turnitin's AI indicator works explains the most common score in schools, and the ranking of detectors for teachers lists what each tool documents about its limits.

So what percentage of AI is acceptable? The one your course allows, which is usually described as a set of tasks, not a number. Read a detector score for what it is, an estimate of how many sentences look machine-written, and read your syllabus for what matters, the rule. Then disclose what you used and keep the drafts that show how you worked.