What does an AI detection score mean?
Percentages, probabilities, and confidence bands, decoded so you stop misreading detector results.
An AI detection score means the tool's estimated probability that a passage was machine-generated, not the share of your text that is AI. This distinction trips up most people in 2026, and misreading it leads to bad decisions and unfair accusations. This guide explains what the number actually represents, why it moves, and how to read a result the way the detector's makers intend.
What is an AI detection score?
An AI detection score is a single figure, usually a percentage, that expresses how confident a classifier is that the input text was written by a large language model. It is the output of a statistical model, not a direct count.
The score compresses many signals, mainly perplexity and burstiness, into one number. Because it is a probability, it carries uncertainty that the clean percentage on your screen tends to hide.
Does a 90% score mean 90% of my text is AI?
No. A 90% score means the detector is roughly 90% confident the passage is AI-generated overall, not that nine of every ten words came from a model. The percentage is confidence, not composition.
Some tools do offer a separate sentence-level view that highlights which parts look AI-generated. That view is closer to composition, but it is still built from the same probability estimates and can be wrong on individual lines.
| Score range | Common label | What it actually means |
|---|---|---|
| 0-20% | Likely human | Low model confidence in AI authorship |
| 20-50% | Uncertain | Mixed signals, treat as inconclusive |
| 50-80% | Possibly AI | Leaning AI but not decisive |
| 80-100% | Likely AI | High confidence, still not proof |
Why do AI detection scores change?
Scores change because they depend on the input and the tool's settings, not just the writing itself. The same paragraph can score differently across tools and even across versions of one tool.
- Text length: short passages give detectors less signal, producing swingy scores.
- Editing: varying sentence length and adding specifics lowers AI confidence.
- Thresholds: each tool sets its own cutoff for labeling text "AI."
- Model updates: vendors retrain classifiers, shifting scores on identical text.
- Genre: formal or technical writing pushes scores upward regardless of author.
That blind spot is why accusations built on a single score fall apart. We cover the evidence and defense in our guide to false positives in AI detection.
How should you read a detection score correctly?
Read a score as a confidence estimate with a margin of error, then look past the number to the details the tool provides. The headline percentage is the least informative part of most reports.
- Note whether the label is confidence or a share of AI text; almost always it is confidence.
- Check the text length; scores under a few hundred words are unstable.
- Open the sentence-level highlights to see where the model reacts.
- Cross-check the same text on a second [AI detector](/ai-detector) for agreement.
- Compare the result against the reported thresholds and known false-positive risk.
Different tools also report pass rates differently, which is why a document that passes one detector can fail another. Our breakdown of detector pass rates across Turnitin and GPTZero shows how far those numbers can diverge.
What should you do with a high AI score?
If your genuine writing scores high, gather evidence rather than panic: keep drafts, notes, and edit history that show your process. A single score should never override that record.
If the text is AI-assisted and you want a lower, fairer result, revise for specificity and rhythm or run it through UmanWrite's humanizer, then recheck. You can compare detection-plus-rewrite plans on the pricing page.
Frequently asked questions
+What does an AI detection score actually measure?
It measures the classifier's confidence that a passage is AI-generated overall. It is a probability estimate, not a count of how many words a model wrote.
+Does 100% mean the whole text is AI?
No. It means the detector is highly confident the passage is AI-generated. It is still a probability, and high-confidence scores can be wrong on human writing.
+Why did my score change between tools?
Each tool uses its own model, thresholds, and training data. Text length, editing, and genre also shift scores, so identical text can land in different bands.
+Is a low AI score proof of human writing?
No. Edited AI text often scores low. A low number reduces suspicion but does not prove authorship, so weigh it with other evidence.
+What score should worry me?
Any score above roughly 80% signals high AI confidence, but it is not proof. Confirm with a second tool and keep drafts if the writing is genuinely yours.
+Can I lower an unfair AI score on my own work?
You should not manipulate a score on genuine work; instead keep evidence of your process. For AI-assisted drafts, revise for specificity or humanize, then recheck.
+Why do formal writers get higher scores?
Formal and technical prose uses even structure and cautious phrasing, which resembles AI output statistically. That pushes scores up regardless of who wrote it.
