What “AI Detected” Means on a Checker Score
By AIDetector365 Editorial Team · 11 min read · Published · Updated
A banner that says AI detected is easy to panic over and easy to over-read. On every major checker — Turnitin, Scribbr, GPTZero, Grammarly, QuillBot, a free webpage — that phrase means the model found writing that looks like typical AI output. It does not mean the tool watched you open ChatGPT, and it does not mean a percentage of sentences were pasted.
This glossary is the page to open when the number is already on screen. For how detectors work, see What is AI detection?. For why no score is 100% trustworthy, see How accurate are AI detectors?. Then try the free AI checker if you need highlights with reasons.
Key takeaways
- “AI detected” is a resemblance score, not a recording of the writing process.
- A percentage is not “this share of the file was generated.”
- Mixed or AI-refined usually means some sentences look generic and others look specific.
- Turnitin has long treated very low percentages as less reliable (historically flagged with an asterisk around 0–20%).
- Read highlighted sentences. If two checkers disagree, the evidence is weak.
Try it now: Free AI Checker
Free AI checker to see if text was written by ChatGPT or other AI. Instant sentence highlights, no login, runs in your browser.
The sentence under the percentage
Translate the UI into plain English before you act on it:
| What you see | What it means | What it does not mean |
|---|---|---|
| AI detected / likely AI / 80% AI | The prose matches patterns the model associates with chatbots | ChatGPT wrote 80% of the words, or there is an 80% chance of cheating |
| Human / likely human | Few of those patterns showed up | The work is original, cited, or allowed by policy |
| Mixed / mixed signals | Some sentences look generic, others look specific | Exactly half the paper is AI |
| AI-refined / AI-assisted | The model thinks a human draft was polished, or AI was edited | Proof someone used a forbidden tool |
| 0% with an asterisk (Turnitin-style) | Little qualifying prose looked AI-like, and low scores are noisy | A clean bill of health you can screenshot as a defense |
Why 100% is still not a verdict
A 100% AI result means the model is as confident as it gets that the sample sits on the AI side of its line. Formal lab reports, templated cover letters and cautious ESL essays can sit on that line without a chatbot. Heavily edited AI can sit on the human side. OpenAI retired its own classifier in 2023 for low accuracy; the category problem did not disappear when other vendors kept shipping.
- Short samples (under ~80–300 words) do not give the model enough rhythm to judge.
- Quotes, tables, code and bullet-only pages are not the prose most detectors were built on.
- A second tool that disagrees is information. A second tool that agrees on raw ChatGPT and disagrees on your edited draft is the usual real-world case.
If the number is being used against a student, read AI detector false positives before anyone treats the screenshot as evidence.
Turnitin, Scribbr and a webpage will not match
Turnitin’s AI indicator only looks at qualifying prose in a submitted file and is separate from the similarity score. Scribbr’s free checker often uses a four-way generated/refined split on a capped paste. A public AI checker highlights sentences in the browser. Three different inputs, three different cutoffs, three different percentages.
A five-step read of any AI score
For an essay-shaped walkthrough, use How to check if an essay was written by AI. For ChatGPT-specific copy, use the ChatGPT detector.
- Ignore the headline percentage for ten seconds. Open the sentence highlights.
- Ask whether each flagged line is generic (stock transition, no names, no data) or specific (a source you read, a method you ran, a memory).
- Check length, language and genre. A 90-word discussion post is a weak sample.
- If you can, run the same text through a second checker and compare highlights, not just scores.
- Decide the next human step: rewrite a flat intro, ask the writer to explain a claim, or leave it alone.
What to do with mixed and AI-refined
If you wrote it
Replace flagged generic sentences with evidence from class or from your notes. Do not run a humanizer to “beat” the number on a graded assignment if AI was not allowed — that is still a policy problem. See AI detector and humanizer.
If you are marking it
Use the highlights as a conversation script: “Walk me through this paragraph.” A mixed result is the worst moment to set an automatic fail line. Teachers’ process is in the AI detection guide for teachers.
Frequently asked questions
What does “AI detected” mean?
It means the checker’s model found writing that resembles typical AI output. It is not proof of which tool was used or that someone cheated.
What does a 100% AI score mean?
The model is fully on the AI side of its threshold for that sample. Formal or cautious human writing can still land there. Treat it as a strong flag to read the sentences, not as a conviction.
What does AI-refined mean?
Usually that the detector thinks AI was used to polish or paraphrase, or that a human draft was run through an assistant. It is a weaker claim than “fully generated.”
Why do two checkers give different percentages?
Different training data, different cutoffs, different handling of quotes and short text. Disagreement means the case is messy.
Does 0% AI mean the paper is original?
No. It means few AI-like patterns showed up. Plagiarism, poor sourcing and empty argument are separate issues.
Put this into practice: Free AI Checker
Free AI checker to see if text was written by ChatGPT or other AI. Instant sentence highlights, no login, runs in your browser.