AI detector highlights: how to review mixed writing

Imagine a writer drafts an introduction, uses an AI tool for a middle section, and then rewrites the conclusion. A detector may highlight part of the middle, miss another part, and also highlight an original sentence. The colour shows what the detector flagged—not a complete record of how the draft was written.

Mixed writing makes it especially important to distinguish a document verdict, an estimated AI-origin share and the words visibly highlighted. Our score guide explains those separate outputs. This guide focuses on what to do with the marked passages.

Does a detector identify words or sentences?

That depends on the product and display method. A model may produce token or word scores, combine them into a sentence score, then colour the entire sentence. Another system may present passage-level predictions. Consequently, the display can include words that did not individually cross a threshold.

Read the report's own explanation. GPTZero's sentence-scanning guidance is one example of a vendor documenting how to interpret its outputs. It does not establish that every highlighted word has a known origin, or that another detector uses the same method.

For Certaffy, treat highlighted passages as suspected AI writing for review. Do not call a highlighted word “proven AI,” or an unhighlighted passage “verified human.”

A worked example: coverage is not accuracy

Suppose an illustrative 1,000-word test document has a recorded construction history: 400 words came from AI output and 600 from a human source. A displayed mask highlights 300 of the AI words and 30 of the human words.

These are invented teaching numbers, not Certaffy performance results. The known AI share is 40%, yet the visible coverage is 33%. A separate percentage estimator could produce another number. None of these measures tells you the document-level false-positive rate across human documents.

Review the passage, not just its colour

1. Confirm the report refers to the current draft, not an earlier version.
2. Check the extracted text against the original, especially after a PDF upload.
3. Read the marked sentence with the surrounding paragraph.
4. Compare it with relevant drafts, notes, sources and any disclosed AI use.
5. Record what remains uncertain before making a decision.

Consider quotations and standard wording too. A routine phrase appearing in many documents is not, by itself, an authorship test. A polished human paragraph can be flagged, and an AI paragraph can be missed. Do not rewrite every highlighted sentence just to change the colour; revise for accuracy, evidence and clarity.

Why can a highlight change when text moves?

Many detectors use context, so the same sentence can receive a different prediction when surrounding text changes. Turnitin's discussion of sentence-level false positives describes boundary-related errors in its own evaluation. That is a reason to inspect neighbouring text, not permission to assume the neighbour is AI too.

When comparing tools or versions, keep the input identical. Adding headings or removing paragraphs changes the test. Use a word counter to check length, but save the actual text as well: identical totals do not guarantee identical documents.

What if the headline and highlights seem inconsistent?

A document-level model and a local highlighting model can disagree. Look for the score definitions and any uncertainty warning. “Likely human” should not be read as “no AI contribution,” and an estimated 25% AI share is not proof of exactly one AI-authored quarter.

Try Certaffy's AI detector to inspect your own permitted text, then use the academic review checklist or business review guide to decide the next step.

Prepared with AI assistance for Certaffy. Sources checked on September 29, 2026. Examples explain metrics and are not benchmark claims.