The Turnitin AI score is the percentage of your document's prose that Turnitin's model believes was AI-generated. It is a separate number from the similarity score, it uses different colors, and instructors see it while you do not. Because students constantly read one score using the other score's rules, this guide lays out both systems side by side, then answers the questions that actually get asked: what is acceptable, what do the colors mean, and is 36% bad.
The two Turnitin scores, untangled
Similarity score: how much of your text matches existing sources, word for word. Catches copying and missing citations. You usually see this one.
AI writing score: how much of your prose reads as machine-generated, statistically. Catches ChatGPT-style text. Instructors only.
A paper can be 3% similarity and 85% AI, or 40% similarity and 0% AI. The first looks original but machine-written. The second is human-written but quotes heavily. Different problems, different conversations.
What do the Turnitin similarity colors mean?
The color flag on a similarity report is just the percentage in bands:
Blue: no matching text at all.
Green: 1-24% matching. The most common band for honest work with normal quoting and citations.
Yellow: 25-49% matching. Worth reviewing: heavy quoting, a long properly-cited passage, or sloppy paraphrase.
Orange: 50-74% matching. Usually structural over-reliance on sources.
Red: 75-100% matching. Largely copied, or you submitted a draft the system already stored.
Two things the colors do not tell you. First, context: a law essay quoting statutes runs yellow legitimately. Second, direction: a green 20% made of one uncited paragraph is worse than a yellow 30% made of properly cited quotes. Instructors read the match list, not just the color.
What does green mean on Turnitin?
Green means 1-24% of your text matched sources, which for most assignments is normal. Quotes, common phrases, your reference list, and boilerplate assignment language all contribute. A green flag with clean citations is not a problem, and chasing 0% by rewording your bibliography is wasted effort.
Is 36% similarity on Turnitin bad?
It depends entirely on what the 36% is. Open the report and look at the matches. If it is quoted sources with citations plus your reference list, you are fine and can say so confidently. If it is paragraphs of paraphrase tracking one source's structure, tighten it before the instructor asks. The percentage is a prompt to look, never a verdict by itself.
What is an acceptable Turnitin AI score?
For the AI score, there is no official acceptable number, and that is not evasion, it is how Turnitin built it. Scores below 20% are hidden or asterisked because Turnitin itself considers them unreliable noise. Above that, most institutions treat under roughly 25% as background risk, and high scores as grounds for a conversation. Policies vary wildly: some departments ignore the AI score entirely, some treat 50%+ as presumptive misconduct. Read your syllabus, because the number that matters is the one in your institution's policy.
From the field: a student sent us a panic message about a '19% AI' rumor from a classmate who had seen the instructor's screen. Turnitin suppresses sub-20% scores precisely because they are statistically shaky. The instructor never raised it. Knowing how the reporting works would have saved a weekend of dread.
Why your AI score can be wrong in both directions
False positives: non-native English writing, rigid templates like lab reports, and heavily grammar-polished prose all read as statistically smooth, which is what the model flags. Turnitin claims under 1% document-level false positives; universities like Vanderbilt found enough real-world cases to disable the feature.
False negatives: properly rewritten AI text with rebuilt sentence rhythm often scores low. Turnitin acknowledges it cannot catch everything.
The accuracy evidence, including the studies, is collected in how accurate is Turnitin AI detection.
Protecting yourself before the score exists
Write with version history on. A revision trail beats any percentage in an integrity meeting.
Pre-check important work with a free AI detector so you see the risk before your instructor does.
If you drafted with AI assistance, rewrite it for real: our AI humanizer rebuilds the statistical rhythm, then your edits add the specifics that make it yours.
Quote and cite cleanly. Most scary similarity scores are citation hygiene problems, not misconduct.
How the AI score is built behind the scenes
The model splits your prose into overlapping chunks of several sentences, classifies each chunk as human or AI, and rolls those results into the document percentage. Three consequences follow. First, the document needs roughly 300 words of prose before any score is produced, so short submissions often have no AI report at all. Second, the score describes coverage, not confidence: 30% means the model believed about a third of the qualifying text was AI-like, not that it is 30% sure of anything. Third, low scores are deliberately soft-pedaled, which is why anything under 20% carries an asterisk.
The model reads running prose and nothing else. It was not designed to judge reference lists, and it cannot see your intentions, your process, or which tool touched the text. It sees sentence statistics.
What to do if your instructor questions your score
Move the conversation from a number to a process, because the process is where your evidence lives.
-
Ask which score is at issue. AI and similarity get conflated in accusations too, and the right response to each is different.
-
Ask to see the flagged passages, not just the percentage. You need to know what you are answering for.
-
Bring your drafts, notes, sources, and version history. A document that visibly grew over days is the strongest counter-evidence there is.
-
Mention Turnitin's own guidance that the score should not be the sole basis for an accusation. Instructors are told this; a calm reminder is fair.
-
Offer to talk through the paper's content. Someone who wrote the work can discuss it, and that conversation resolves most cases faster than any report.
Mistakes students make reading their reports
Chasing 0% similarity. Quotes, references, and standard phrasing always match something, so a modest green score with clean citations is the healthy state.
Applying similarity color logic to the AI score. The color bands belong to the similarity side; the AI percentage has no colors.
Comparing raw percentages between classmates or classes. Assignment settings change what counts as a match, so the numbers are not comparable.
Panicking over rumors of a low AI score. Sub-20% results carry an asterisk precisely because Turnitin considers that range unreliable.
Resubmitting repeatedly to nudge a number down. Each submission can be stored and re-analyzed, and on the similarity side a resubmission can even match your own earlier draft.
Frequently asked questions
Can students see their own Turnitin AI score?
No. The AI report is instructor-only. If you want a self-check before submitting, public detectors are your only option.
Does a high AI score automatically mean an academic integrity case?
No. Turnitin explicitly tells institutions the score is not proof and should open a dialogue. Most integrity processes require more evidence than a percentage.
What does a yellow flag mean on Turnitin?
Yellow is a similarity band, 25-49% matched text, not an AI verdict. Review your matches: heavy but cited quoting is usually fine, uncited paraphrase is not.
Why is my Turnitin AI score zero when I used ChatGPT for ideas?
Using AI to brainstorm or outline leaves no statistical trace in prose you wrote yourself. The detector reads your sentences, not your process. Whether idea-level AI use is allowed is a policy question for your syllabus.
Do quotes and references count toward the AI score?
The AI model is aimed at your running prose. Reference lists are not what it judges, and quoted material mostly plays out on the similarity side. If your AI score is high, the cause is in your paragraphs, not your bibliography.
Can the AI score change if the same paper is resubmitted?
Yes. Each submission is analyzed fresh, and Turnitin updates its model over time, so a resubmission months later can score differently. That drift is normal, and it is one more reason no single percentage should decide anything on its own.
