Reading the report
How to read an AI detection score without overinterpreting it
The percentage is the beginning of a review. Understanding what was analyzed, what the score means and what evidence is missing makes the result more useful.
Start with the definition, not the number
Different detectors use percentages differently. A figure may describe a classification estimate or a proportion under a particular method. It is not automatically the percentage of words written by AI. Read the explanation supplied with the result before making that interpretation.
Check AI presents the estimate and any available passage highlights returned by its processing service. These are model outputs. They are not a record of keystrokes, a document revision history or proof that a named model wrote a sentence.
Also confirm which text was submitted. A result for an introduction cannot stand for an entire report. A selected passage may exclude the sources, examples or discussion that give a document its context. Keep the original version available so you know exactly what you reviewed.
An illustrative review of a flagged paragraph
Imagine a draft with a high detection estimate and a highlighted opening paragraph. The passage explains a familiar topic in broad, repetitive language. That observation gives the editor something to examine; it does not establish how the text was produced.
The editor checks whether the opening accurately introduces the article, whether its claims have sources and whether it contains anything useful for the intended reader. If the paragraph is vague, the author can add a relevant, verified example. That change improves the article regardless of the model’s next score.
If the context requires disclosure of AI assistance, discuss the drafting process and the applicable rule with the author. Do not treat a fluent rewrite or a lower second score as evidence that the disclosure requirement has been satisfied.
Five checks before acting on a result
Confirm the input
Record the version, passage boundaries, language and approximate length. Avoid comparing reports for different extracts as if they described the same document.
Read the surrounding passages
A highlight is a reason to look more closely. Read the argument around it and inspect the source material where relevant.
Check facts independently
A detector does not verify dates, quotations or references. Open the original sources and correct unsupported claims.
Allow for error
Human writing may be flagged and generated writing may be missed. A single result cannot settle an accusation or certify authorship.
Choose an evidence-based next step
Ask for clarification, improve a vague explanation or review attribution. State the specific issue being resolved instead of making a target score the goal.
Why a changed score is not proof of improvement
Editing changes the text the model sees. Shortening sentences, translating a passage or changing the selected extract can change the output. That difference alone does not tell you whether the revised version is more accurate or better written.
Evaluate a revision against a concrete objective: preserving meaning, making an instruction clearer or supplying a missing reference. Compare versions side by side and keep the one that best serves the reader. If you do compare detection results, record the tool and inputs so the comparison has context.
For academic or workplace decisions with serious consequences, follow the relevant policy and review process. A score should not be the only basis for penalizing a person. The reliability guide explains the two types of classification error and a broader review method.
Your next draft
Try it with your own words.
One eligible 100-word humanization trial after email verification. No card required.