You've run a scan and you have a number. Now what? This guide walks through the common score bands, from under 10% to over 50%, answers how much AI detection is acceptable, and shows how to read the sentence highlights so a result becomes a sensible next step instead of a snap judgment.
Key takeaways
- A score is a likelihood estimate, not a count of AI-written words.
- Under 20% AI is common in fully human writing; 20% on its own is not bad.
- From 20% to 50%, look at where the flagged sentences cluster before deciding anything.
- Over 50% is a strong reason to follow up, but still evidence rather than proof.
- No detector sets an acceptable level: the policy that applies to the work does.
What the number is measuring
A detector compares text against patterns it learned from large amounts of human writing and AI writing. The overall score reflects how strongly the text as a whole matches the AI side. It answers 'how machine-like does this read?' and not 'how many words came from a chatbot?' Keep that distinction in mind and most results become easier to interpret.
That difference matters. A score that leans AI doesn't mean a fixed share of the document was generated. A piece that is half human and half AI can produce quite different overall scores depending on how distinct each part is. Two essays with similar scores can also have arrived there in completely different ways. That's why the sentence view exists, and why it deserves most of your attention.
What each AI detection score band means
Many detectors, Turnitin among them, report a percentage: roughly, how much of the text their model predicts was written by AI. The bands below are a reading guide for that kind of number, not a rule. The same percentage can mean different things in a five-paragraph essay and a one-page cover letter, so always read the highlighted sentences before you act.
- 0–10% AI: little or no signal. This is where most ordinary human writing lands. Stray flags usually sit on boilerplate such as definitions, set phrases or a formal opening, and rarely call for any follow-up.
- 10–20% AI: a weak signal. Fully human work lands here often, especially formulaic writing or text polished with grammar and rewriting tools. Note which sentences are flagged, but don't treat the number as a finding.
- 20–30% AI: worth a look, not a conclusion. Check whether the flagged sentences cluster in one place, such as an introduction or conclusion, or are scattered. A cluster is a question to ask the writer; scattered flags usually aren't.
- 30–50% AI: a meaningful share of the text reads as machine-written. Read the highlights closely, compare the piece with the writer's earlier work and look at drafts or version history. Mixed human and AI writing and heavy paraphrasing often land in this band.
- 50% AI and above: most of the text reads as machine-written. That is a strong reason to follow up, but it is still evidence, not proof. Ask how the piece was written before you decide anything.
Is 20% AI detection bad?
Not on its own. A 20% reading means a minority of the text looks machine-like, and human writing gets there for ordinary reasons: a formal or formulaic style, careful second-language prose, or edits from a grammar tool. Low readings are also where detectors are least reliable. Turnitin, for example, has said its detector produces more false positives when it finds AI writing in less than 20% of a document, and its report marks scores in that range with an asterisk rather than presenting them like higher scores.
So treat 20% as a prompt to look, not a verdict. If the flagged sentences are the ones that felt generic when you first read the piece, ask about them. If they are scattered across otherwise personal writing, the number probably reflects style rather than a chatbot.
How much AI detection is acceptable?
There is no universal acceptable percentage, and a detector can't set one. What counts is the policy that applies: a school's academic integrity rules, a course syllabus, an employer's or a client's guidelines. Many of those policies care about how AI was used, for brainstorming, editing or writing, rather than about a number, and a score can't tell those uses apart. If you're unsure where the line is, ask the person who set the work before you submit or decide anything.
How Winston AI's score compares
Winston AI shows a human score from 0 to 100, which runs the other way: higher means more human. Subtracting it from 100 gives a rough AI reading you can hold against the bands above, but it isn't a count of AI-written words. The score takes the higher of two estimates, the model's view of the whole document and the length-weighted average of each sentence's AI probability, so a few strongly flagged paragraphs can't be averaged away by a clean remainder. Long documents are scored in parts and combined by length.
- 60 to 100: likely human-written.
- 35 to 59: a mix of AI and human writing.
- Under 35: likely AI-generated.
Sentences are colored on their own probability: red at 65% or more, yellow from 40% up to that. Scans need at least 80 words, because shorter text gives any detector too little to work with.
Reading the sentence highlights
In Winston AI, each sentence gets its own assessment, shown as a color:
- Red: the sentence reads as likely AI.
- Yellow: mixed signals. This often shows up where a person wrote the text and a tool polished or paraphrased it, or where someone edited AI output by hand.
- No highlight: the sentence reads as human.
Hover or tab to any sentence to see its individual AI probability. The pattern of colors usually tells you more than the overall score. A single red paragraph in an otherwise clean essay points to one section to ask about. Red spread evenly across a whole piece is a different story from a few yellow sentences in an introduction. Picture an essay whose quotations and personal examples are unhighlighted while its topic sentences and conclusion are red. That might mean the student wrote the core and used a tool to frame it. It's a question to ask, not an answer.
When a score deserves less weight
Some texts give any detector less to work with. A result on these is still worth reading, but it should carry less weight in your decision. Be more cautious when:
- The text is very short, such as a single paragraph or a few sentences.
- The genre is formulaic, like a lab report, a templated cover letter or a legal clause.
- The writer is working in a second language and writes in a careful, textbook style.
- The text has been through grammar or rewriting tools, which can push human writing toward yellow.
- The writer was told to follow a strict structure or model answer, so many submissions look alike.
For a fuller look at where results go wrong, read can AI detectors be wrong.
A low score isn't a guarantee either
Text that reads as human may still have had help from AI. Someone might use a chatbot for an outline and then write every sentence themselves, or heavily rework generated text. Detection looks at the writing on the page. It can't see the process behind it. That's why a detector works best as one input in a review, not the whole review. Your knowledge of the writer, the assignment and how the work was produced fills the gaps a score can't.
Turning a score into a next step
- Run the scan in the AI detector and note both the overall score and where the highlights fall.
- Re-read the highlighted sentences yourself. Do they share the habits you associate with AI writing?
- Gather context: drafts, version history, earlier work, and the brief or prompt the writer was given.
- If questions remain, talk to the writer and ask about the highlighted sections specifically.
- Record what you found. Paid plans keep scans in your history and let you export a shareable PDF report. Teachers will find a full review routine on our page for teachers.