What Do AI Detection Scores Mean? How to Read a Percentage
Short answer
An AI detection score is an estimate, and the same percentage can mean different things in different tools. It might be the share of sentences flagged, the probability that the whole text is AI, or a confidence value. None of these proves who wrote the text. Read the verdict, the highlighted sentences and the text length together, and never treat the number alone as a finding.
Key points
- A percentage can mean the share of text that looks like AI or the probability that the text is AI. These are not the same thing.
- A score of 30% does not mean 30% of the writing was done by AI in a way anyone can verify.
- Different tools give different numbers for the same text because their training, scoring and thresholds differ.
- Verdict labels and sentence highlights are easier to interpret honestly than a single raw number.
- A score is a reason to look closer, not a conclusion. Use drafts and a conversation to settle real doubts.
Start with what the number is
Before you decide what a score means, find out what the tool is reporting. Many students see a figure like 40% AI and assume it means 40% of the essay was written by a machine. That is only one of several possible meanings, and often it is not the right one.
The same figure can describe quite different quantities depending on the tool, so always check its explanation of the score before comparing results.
Share of text flagged versus probability of AI
There are two common ways a tool turns its analysis into a percentage, and they answer different questions.
A share-of-text score says how much of the document looked AI-like: for example, 4 sentences out of 10 were flagged. A probability score says how likely the tool thinks it is that the document, taken as a whole, is AI-generated: for example, an 85% chance.
- A 40% share-flagged result on a human essay can mean four generic sentences, with the rest fine.
- A 40% probability is a fairly uncertain result about the whole text, not a measure of how much is AI.
- A fully AI text can get a high probability and a high share. A lightly edited AI text may get a high probability but a lower share.
Why different tools disagree
Run the same paragraph through three detectors and you may get three different answers. That does not mean two of them are broken. Each tool has its own training data, its own way of scoring windows and sentences, its own threshold and its own idea of what the number means.
Detectors are also tested on different material. A tool calibrated on academic writing will behave differently on advertising copy, and one tested on native speakers may behave differently on learners. Because their scores are not on a common scale, comparing raw percentages between tools is not meaningful. Compare the sentences highlighted and the reasoning instead.
How to read our three verdicts
AITurnitin gives one of three verdicts instead of a single score to interpret, and it highlights the sentences behind the result. Texts under 50 words get no verdict, because there is too little evidence.
- Likely human: the wording does not resemble AI text closely enough to flag. It is still an estimate, not a certificate.
- Uncertain: the text sits in a borderline band. Read the highlights, add specifics and consider checking again after revising.
- Likely AI: the wording resembles the AI text the detector was trained on. On human academic texts in our calibration this label appeared 1% of the time, so it can be wrong; unedited 2026 AI text received it 96.9% of the time.
What a score cannot tell you
A score cannot tell you who wrote the text, whether AI help was permitted, or how the work was produced. It cannot separate a student who wrote carefully and formally from a student who used a tool, because their text may look the same.
It also cannot tell you how much AI was involved. Someone who had AI fix grammar and someone who had AI write the argument can produce similar numbers. That is why decisions about integrity should rest on process evidence and a conversation, not on a threshold.
Using a score sensibly
If you are a writer, use the sentence highlights to find generic passages and revise them with specific details and your own reasoning, then check again. If you are an instructor, treat a flag as a prompt to ask about process, not as a verdict. If you are the subject of a flag, collect your drafts and version history and ask what evidence the concern rests on.
You can try this on your own draft: AITurnitin checks up to 1,000 words free without an account, up to 3 times a day. Even our well-calibrated detector labelled 1% and 0.7% of human academic text Likely AI on two held-out sets, so a score alone should never end a conversation.
Frequently asked questions
What does a 50% AI score mean?
It depends on the tool. It might mean half the sentences were flagged, or that the tool is unsure whether the whole text is AI. Either way it is not a measurement that half the text was written by a machine, and it is not proof.
Is a 0% AI score proof that I wrote it myself?
No. It shows that the wording did not resemble AI output to that tool. Lightly edited or rewritten AI text can score low, so authorship is best shown with drafts and an explanation of your work.
What AI score is considered too high?
There is no universal line. Tools and institutions use different thresholds, and a number from one tool cannot be compared with another. Look at which sentences are flagged and whether the writing is specific and yours.
Why did two detectors give me different scores?
They are trained differently, score differently and use different thresholds. Their numbers are not on the same scale, so a difference does not show that either one is right.
What does Uncertain mean in a detection result?
It means the text falls in a borderline range where the detector does not want to claim either way. Review the highlighted sentences, make generic passages more specific and check again if you wish.