What Percentage of AI Is Acceptable?

An AI percentage is a confidence estimate about a block of text, not a measurement of how much of that text a machine produced. A 46% reading does not mean nearly half your essay came out of a model. It means the classifier looked at the whole passage and landed at that level of confidence in a machine origin. Nothing in the number tells you which half, because there is no which.

That distinction matters most on the second confusion, which is more common still: an AI reading is not a similarity reading. The similarity score is Turnitin’s original plagiarism metric, built years before AI detection existed to measure text overlap against a database of sources, and the old student rule of thumb that grew around it, keep it under 15% and you are fine, made sense there because quoting and citing produce overlap by design. AI-writing detection arrived much later, measures something else entirely, and inherited none of that convention. Carrying the similarity habit across to the AI figure is how people talk themselves into calm and into panic on the same afternoon.

The same displayed number also means different things in different tools, and one vendor makes that unusually visible. Originality.ai’s own help documentation offers institutions a configurable “AI Allowance” set at 0%, 5%, 15%, 25% or 40%, with no level designated as the correct one (read August 2026). The threshold is a setting somebody chooses, not a property of your writing. Two tools can read one paragraph and disagree, and if you want the measured version of that disagreement, our detector-reliability statistics cover run-to-run variance and inter-tool agreement in detail. What the figure technically is inside one specific report is a separate question, answered in our breakdown of Turnitin’s AI checker, and as the next section shows, we checked all four major detection vendors’ published guidance directly; none of it names a cutoff.

A 20% AI reading on Turnitin is not a finding of misconduct and is not evidence of anything on its own. It sits at the exact edge of the range Turnitin is willing to print: readings from 1% to 19% are shown as an asterisk with no number attached at all. So a paper at 20% has crossed from “the vendor will not name this figure” into “the vendor will name it”, which is a statement about the instrument’s confidence, not about your integrity.

Turnitin’s own guidance is direct about why that band is blank. The suppression exists, in the vendor’s words, “to avoid potential incidence of false positives”: inside the 1% to 19% range its report attributes no score and no highlights whatsoever (Turnitin’s own guidance, read August 2026). Read that the way it is written, because it usually gets cited backwards. Turnitin is not certifying the band as acceptable. Turnitin is saying that inside that band it does not trust its own reading enough to put a figure on it.

A second sentence from the same guidance is the one worth carrying into any conversation about your paper: an AI Writing score “should not be used as the sole basis for adverse actions against a student.” That is the company that built the classifier, in its own documentation, telling institutions the output is not sufficient by itself. If you are ever asked to account for a percentage, that instruction is more useful to you than any threshold, and it does not depend on anyone believing your side of the story.

None of that explains how the number is produced, or how the report behaves when a paper is short or resubmitted. That is the instrument rather than the verdict, and it is covered properly in our Turnitin AI checker guide.

Outside Turnitin, there is no published band whatsoever, and the vendors say so themselves. GPTZero’s own site states that “there currently isn’t a broadly ‘acceptable’ AI percentage, as some institutions ban AI entirely, while others allow it in a variety of ways” (read August 2026). Copyleaks’ own policy guide takes the same position, that “one policy does not fit all, as each school’s situation is unique,” and describes its detection report as a data point rather than a compliance line. Originality.ai ships the adjustable allowance described above instead of a recommendation. Four vendors checked; not one publishes a number.

This is worth stating plainly, because the tidy band tables circulating on this question (0-15% safe, 16-30% caution, and so on) are not sourced to anything. The pages carrying them assert that no universal percentage exists and then publish a scale anyway, with no citation behind any row. We are not going to invent one here. What can be said honestly is qualitative: a small reading on a short passage is the least reliable output these tools produce; a large reading whose highlighted passages line up with the parts you actually drafted with help is telling you something real; and every reading is interpreted by a person whose rules you can go and read for yourself.

Two things make that qualitative read easier. The first is a breakdown instead of a bare figure: our free AI checker reports which specific patterns pushed a score up, so you can see whether a reading rests on two stray sentences or on the texture of the whole piece. The second is knowing how often these tools are wrong in the first place, which is measured rather than guessed in our analysis of detector accuracy.

In college, the acceptable percentage is set by your course, not by your university and not by the detector. We looked for a published numeric cutoff at institutional level and did not find one, and the institutions that have addressed the question directly describe discretion rather than thresholds. UC San Diego’s own student newspaper reported in November 2025 that the campus “does not seem to have a concrete answer; rather, it leaves it up to the instructor to decide.”

The University of Melbourne goes further and takes the number away from students entirely: its academic-integrity guidance states that the AI writing indicator and report are not visible to students at all, and that the result “is a prompt for further investigation” rather than a conclusion. No percentage appears anywhere on that page. A high score, in the only institutional language we could find on the point, starts a look, not a case.

Several universities have made a stronger version of the same judgment by switching the indicator off rather than publishing a line for it: Vanderbilt disabled Turnitin’s AI detector in August 2023, Yale’s Poorvu Center does not endorse AI-detection software or enable it in Canvas, citing a late-2025 University of Chicago Booth study that called such tools “unsuitable for high-stakes applications,” the University of Waterloo discontinued the function in September 2025 after internal testing flagged human-written text as 100% AI-generated more than once, and Curtin University disabled it across all campuses from 1 January 2026. UC San Diego’s Extended Studies division deactivated the indicator in April 2025. A fuller account of those decisions sits alongside the score mechanics on our Turnitin page.

What actually shifts between a discussion post, a term paper, a research paper and a thesis is not a percentage. It is how much of your own reasoning the work is meant to demonstrate, and what you are required to declare. A literature review assembled with heavy tool assistance and disclosed in a methods note can be entirely in order; the same assistance undeclared in a thesis chapter is a problem at any reading, including 0%. The governing document is your syllabus, your program handbook, or the academic integrity policy your institution publishes, one paragraph of which is worth more to you than every band table on the internet combined.

A high AI score does not by itself open a misconduct case. In every piece of institutional guidance we read for this page, a score functions as a prompt for a closer look, and the closer look is usually informal: an instructor reads the flagged passages, and often asks you about them. A formal integrity review is a separate step with its own process, and the guidance is explicit that a report is not enough to get there on its own. Melbourne states it in one sentence: “an AI writing detection report alone is not sufficient evidence for an allegation.” The percentage itself is not a grade deduction, either: no institutional guidance we read treats the score as a penalty on its own, separate from what a review actually concludes.

Two things are true at once here, and pages on this topic usually publish only one of them. The first is that instructors are paying real attention, and the vendor’s own data is why. Turnitin’s press release of 24 February 2026 reports that 14.8% of English-language submissions to its latest detector carried 80% or more AI writing between October 2025 and February 2026, up from 3.3% of submissions to its original detector in 2023. Those are Turnitin’s own figures with no published methodology behind them, so read them as a market signal rather than an audited statistic, but a rise of that size explains why a flagged paragraph gets read carefully now.

The second is that these instruments make errors with a documented shape. Liang and colleagues, publishing in Patterns in 2023 (DOI 10.1016/j.patter.2023.100779), ran seven detectors over 91 TOEFL essays written by non-native English speakers with no machine involvement at all. The detectors labelled 61.3% of them AI-generated. If English is not your first language, that is the single most relevant figure on this page for you, and the full picture is in our guide to ESL false positives.

Anyone quoting an error rate should also say what it was measured on, so here is the population behind ours. The calibration run behind our free detector, dated 30 July 2026, scored 15,542 human-written passages spanning TOEFL essays by non-native writers, native-English student essays, pre-ChatGPT student work, arXiv abstracts, literary prose and published long-form web writing. It flagged 34 of them, a 0.2% false-positive rate, and on the two sets most likely to sit in front of an integrity officer, the TOEFL essays and the native-English student control, it flagged none at all. That is the reading we hold the tool to, stated with its population and its date, and the same discipline is what to ask of any detector whose number is being held against you: a score is a signal about a passage, never proof about a person.

What to do about a high AI score depends entirely on how the words got onto the page, and the two paths are close to opposites. Taking the wrong one actively damages your position, so match the situation before you touch anything.

Start with disclosure, because it is the part that is cheap now and expensive later. If your course asks you to declare tool use, declare it in the form it asks for; a disclosed workflow with a middling reading is an ordinary conversation, while an undisclosed one is the conversation nobody wants. Then reduce the reading the only way that survives scrutiny: make the finished text genuinely yours. Open the report, find the passages it highlighted, and revise those with your own examples, your own transitions and your own argument: the machine-drafted sections are almost always the flat, evenly-paced ones, and they are the passages a human reader questions too. If the tool in your workflow was Grammarly, which of its features count as generative is a question of its own, answered in our breakdown of whether Grammarly gets flagged as AI.

Our humanizer is built for exactly that revision pass: switch the widget above to the Humanize it tab and it rewrites AI-drafted passages into ordinary prose, trained on a corpus of 2,590 real student essays rather than on synthetic text. A free account comes with four rewrites of up to 250 words each and no card, which is enough to bring a stubborn paragraph into range. If you are working through a whole essay in one pass rather than paragraph by paragraph, Basic at $18 a month raises the limit to 1,000 words per run. Either way, re-check the result before you submit rather than after.

One boundary on all of this: it applies to drafts that genuinely had AI in them. If the words are already yours, do not rewrite your own work to chase a lower number. That path is below, and it goes somewhere else entirely.

Do not edit the flagged document. Not a sentence, not through a rewriting tool, not “just to be safe”: the file exactly as it stands is the strongest thing you own, and every change made after you saw the score becomes something you have to account for later. Preserve the record of how the writing happened instead: revision history in Google Docs, version history in Word, timestamps, notes, sources, commits.

Then bring two borrowed sentences to the conversation. The first is the vendor’s own instruction that an AI Writing score should not be the sole basis for adverse action against a student. The second is Melbourne’s, that a detection report alone is not sufficient evidence for an allegation. Neither is your opinion, and neither asks anyone to take your word for anything. If English is your second language, the 61.3% finding above belongs in the same file. The complete version of this (what evidence to collect, in what order, and how an appeal actually proceeds) is set out in our guide to proving you didn’t use AI. If it is an employer rather than an instructor holding the number over you, start instead with our guide to being accused of using AI at work.

The cheapest version of this whole problem is the one that happens before submission rather than after. If you are unsure how a piece of writing reads, run it through the detector while you can still do something about it, four checks a day of up to 250 words with no account required, and treat whatever comes back as one reading of your prose, not a verdict on your honesty.