Read the words beside the score before you act on the number. “Human Score,” “AI probability” and an AID review rating mean different things. A high score is a reason to look more closely, not proof of who wrote the text.
Three numbers that should stay separate
- The document label: The tool’s answer, such as Human, AI, Mixed or uncertain.
- Confidence or probability: The tool’s estimate. Check how that tool defines the number.
- The AID review score: Our rating for accuracy, protection against false alarms, ease of use and value. A 4.8/5 review score does not mean 96% detection accuracy.
How our three reviewed products present results
GPTZero shows separate Human, AI and Mixed probabilities. Keep those labels with the numbers. Do not turn them into a new score of your own. See GPTZero’s FAQ on classifications and probabilities.
Winston AI uses a Human Score alongside its explanation and highlighted passages. Keep the “Human Score” label with the number and read the explanation. For example, a 39% Human Score does not give us a verified 61% AI probability. See Winston’s results guidance.
Originality.ai describes its AI score as confidence in the result. A score of 78% does not mean that 78% of the words were AI-written. It also does not identify the writer. See Originality.ai’s guidance on flagged human writing.
False positives, false negatives and uncertainty
A false positive is a false alarm: the tool flags human writing as AI. A false negative is missed AI writing: the tool calls it human. An unclear result leaves the question open. If a scan fails or only checks part of your document, it has not given you a full-document answer.
In our eight-document test, GPTZero missed the AI03 script, Winston AI gave an unclear answer, and Originality.ai flagged it as AI with 78% confidence. All three tools recognized our four human documents as human. Those human samples came from the editor-in-chief’s books, doctoral thesis and script writing. Four samples from one writer cannot show how the tools will treat everyone. Read the test comparison for settings and limitations.
Why highlighted sentences need context
A highlighted sentence is a place to look more closely. It is not proof that AI wrote that sentence. Save the full text you submitted, including the sentences around the highlight. Avoid rewriting just to chase a lower score. That changes the work you are trying to understand.
A five-step review process
- Save the exact text, full report, date, tool, plan and settings.
- Read the result label, any warning and the tool’s explanation of its numbers.
- Check drafts, notes, sources and revision history where available.
- Ask the writer about the flagged passages and how they wrote them. Give them a chance to explain.
- Write down what you can support and what you still do not know.
Does agreement between tools prove authorship?
No. Another tool may add a useful clue, but counting votes does not give a reliable probability. We have not shown that the tools make errors independently. When we know how a test sample was written, a detector’s answer does not change that fact.
How to use this when choosing a detector
Choose a tool that fits your full documents, gives clear reports and meets your privacy needs. Start with the selection guide, then read the GPTZero, Winston AI and Originality.ai reviews. See how we test to understand our ratings.
Common questions
Does 100% AI mean the tool is certain?
It means the tool returned its highest displayed score under its own rules. It still does not prove authorship. Read the report and check the writing process.
Does a low score prove the work is human?
No. A detector can miss AI writing. Our script test shows one example.
What should I send if my work is flagged?
Ask for the full report. Share relevant drafts, notes, sources or revision history and explain how you developed the work. For a simple starting point, see our student guide.