An AI score is a detector's estimate of how likely it is that a piece of text was generated by AI, usually shown as a percentage or a label like "likely human" or "likely AI." It is a probability based on writing patterns, not a measurement of how many words came from a chatbot, and not proof of anything. Read it as a starting point for a closer look, alongside drafts and context.
What the number actually represents
AI detectors are classifiers. They have been trained on large collections of human-written and machine-generated text, and they learn statistical differences between the two. When you paste in a document, the detector compares its patterns with what it learned and outputs a confidence level.
Depending on the tool, the headline number can mean different things:
| Display | What it usually means |
|---|---|
| "82% AI" | Often a confidence estimate that the text resembles AI output; some tools mean the share of text flagged |
| "18% human" | The same estimate, expressed the other way around |
| "34% of sentences flagged" | Roughly a third of sentences individually looked AI-like |
| "Likely AI / Mixed / Likely human" | The score converted into a band with set thresholds |
| Highlighted sentences | Sentence-level estimates that feed into the overall result |
The most common misunderstanding is reading "82% AI" as "82% of this was written by AI." For many tools, that is not what it means. Check the tool's documentation or help text to be sure.
What goes into the score
Detectors rely on signals such as:
- Predictability. Language models tend to choose highly likely next words. Text where every word is the "expected" one can score higher.
- Variation. Human writing often varies sentence length, structure and rhythm more, though plenty of humans write very evenly.
- Phrase patterns. Some transitions and stock phrases appear more often in AI output.
- Consistency across the text. A sudden shift in style partway through can affect sentence-level results.
None of these signals is unique to AI, which is why scores can be wrong in both directions. We explain how specific content types affect results in AI detection by type.
What an AI score cannot tell you
A score does not tell you:
- Who wrote the text. It only estimates resemblance to patterns.
- Whether a rule was broken. Many courses allow some AI help, such as brainstorming or grammar suggestions.
- Whether the content is accurate. Human and AI writing can both be wrong.
- Whether the text is plagiarized. Copying from sources is a separate question that plagiarism checkers address.
- How the text was produced. A human-written draft polished by an editing tool and an AI draft heavily revised by a person can land at similar scores.
Why scores differ between tools
Paste the same paragraph into three detectors and you may see three different numbers. That is expected. Each tool uses different training data, different model designs and different thresholds for labels like "likely AI."
Scores can also shift within a single tool when you add or remove a paragraph, because longer text gives the model more to work with. If a result matters, look at the sentence-level highlights rather than fixating on the headline number.
How to read a result step by step
- Check the length. If the text is only a few sentences, treat the score as unreliable.
- Read the label and the explanation. Find out whether the number is a likelihood or a share of sentences.
- Look at the highlights. Which sentences drove the result? Are they definitions, lists, boilerplate or quotes?
- Consider the genre. Technical reports, templated assignments and formal emails naturally read as predictable.
- Compare with what you know. Does the writer have drafts? Does the style match their other work?
- Decide what action, if any, fits. For your own draft, that might be revising generic passages. For someone else's work, it might be a conversation.
If you are checking your own writing
Running your own draft through a detector can be useful, as long as you use it the right way.
Good uses:
- Spotting paragraphs that read as generic or padded, so you can add specifics and your own reasoning
- Checking that AI-assisted brainstorming has not crept into your final wording, if your course requires your own words
- Preparing for a conversation by understanding how your text might look to a reviewer
Uses to avoid:
- Rewriting over and over to push the number down. This rarely improves the writing and can undermine your ability to show your real process.
- Treating a low score as permission to submit work that breaks your course policy. The policy matters, not the number.
If your genuine work gets a high score, read our guide on AI detection false positives for what to do next.
If you are reviewing someone else's work
For teachers, editors and managers, a few principles keep score-based decisions fair.
- Never act on a score alone. Combine it with drafts, version history, prior work and a conversation.
- Know the tool's limits. Read what the provider says about accuracy, false positives and minimum length.
- Watch for groups at higher risk of false positives, including second-language writers and students following strict templates.
- Be transparent. If detectors are used in your course or organization, say so, and explain how results are handled.
- Focus on learning. In education, the goal is understanding whether a student did the thinking, not catching them out.
Two examples of the same score meaning different things
Scores only make sense in context. Consider two documents that both come back at roughly the same high AI likelihood.
Document A is a 1,200-word argumentative essay. The highlights are spread across most paragraphs, the writing has no personal examples, every paragraph uses the same structure, and the student has no earlier drafts. The score is a reasonable prompt for a conversation about how the essay was written.
Document B is a 150-word lab procedure. It follows the department template line by line, uses standard terminology and passive voice, and the student's lab notebook shows the work was done in class. Here the score mostly reflects the genre. It tells the reviewer almost nothing.
Same number, very different meaning. That is why experienced reviewers spend more time on the highlights, the genre and the writer's process than on the percentage at the top of the page.
Scores change when text changes
It also helps to know that scores are not fixed properties of a piece of writing. Adding a paragraph of personal reflection, removing a block quote or splitting a document into sections can move the result. That is not a flaw you should exploit; it is a reminder that the score describes a particular input at a particular moment, measured by a particular model. Treat it with the same caution you would give any single measurement.
Common score ranges and sensible responses
Detectors differ, so treat this as a general guide rather than a rule.
| Result | Sensible response |
|---|---|
| Low AI likelihood | No action needed based on the score; still review content quality as usual |
| Mixed or moderate | Look at highlighted sentences; often formulaic passages or quotations |
| High AI likelihood on long text | Look at process evidence and talk with the writer before drawing conclusions |
| High AI likelihood on short text | Treat as inconclusive; there is not enough text |
Key takeaways
- An AI score estimates how much text resembles AI writing; it is not a word count of AI content and not proof.
- Different tools and text lengths produce different scores, so look at sentence highlights and context.
- Short, formulaic and second-language writing are more prone to misleading scores.
- Use scores to improve your own writing or to start a fair conversation, never as the only basis for a decision.