Plagiarism Detector Flagged You: What the Score Means

A plagiarism detector compares your text against a database of published sources. An AI detector asks a different question: does this read as machine-generated? If you got flagged for plagiarism, the fix depends on which one you’re actually looking at. Matching a source and sounding like a model are separate failures, and they need separate repairs.

What a plagiarism detector is actually measuring

Most tools sold as plagiarism detectors run a text-matching check. Your sentences get fingerprinted, compared against indexed web pages, journals, student papers and books, and anything that overlaps gets highlighted with a source link and a similarity percentage. That is it. There is no authorship judgement in the comparison itself.

Some vendors now bundle an AI-writing score into the same report, which is where most of the confusion starts. You see one red number, assume it means “you copied”, and start rewriting the wrong things. The two signals come from different engines and fail in different ways.

So read the report properly before you touch a word:

1. Find the matched text. Every serious tool links each highlighted span to a source. If the source is a real paper or page you used, that is a citation problem, not a detection problem.
2. Check the match type. A block of quoted material in quotation marks with a reference is a formatting issue. A paraphrase that tracks the original sentence structure word-for-word is a genuine attribution failure.
3. Look for an AI score on a separate tab or line. If it’s there, treat it as a probability estimate from a different model, not as proof that you copied anything.

A high similarity percentage with no source links usually means the tool matched your own previously submitted work, a template, or a widely mirrored phrase. That happens constantly with methods sections, standard definitions and legal boilerplate. It is not misconduct, and it is fixable by quoting and citing rather than by scrambling the words.

Why your plagiarism detector score and your AI score disagree

This is the part that makes people panic. Turnitin’s own guidance says a document it identifies as roughly half likely AI-written “could contain as much as 65% AI writing”, and the same page states that scores in the 1% to 19% range get no percentage attributed at all (source: turnitin.com). GPTZero’s FAQ explains its class probability as the chance the detector is correct on similar documents — a 90% reading means 90% of the time on comparable texts the prediction holds, not that 90% of your essay is machine-made (source: gptzero.me).

Both vendors are describing a probabilistic classification. Neither is describing a measurement of your intent.

Our own measurement makes the overlap problem concrete. On 2026-09-28 we scored 709 hand-written DeAIze guides with our detector (local ensemble, lite tier) at a 30% flag threshold. The mean AI score was 26%, the median 25%, and the range ran from 0% to 69%. 239 of those 709 human-written documents crossed the flag line. That is our instrument, on our own writing, and it still flagged a third of it. Any claim that a single score settles authorship should be read against numbers like those.

Tool / signal	What it compares	What a high reading means	Reported confidence language
Text-matching plagiarism check	Your text vs indexed sources	Overlap with a specific source exists	Source-by-source match, usually with a percentage per source
Turnitin AI detection	Text patterns vs model output	Likely AI-written segment; 1–19% gets no score attributed (source: turnitin.com)	50% identified as likely AI “could contain as much as 65% AI writing” (source: turnitin.com)
GPTZero	Text patterns vs model output	Classified as AI, with a stated confidence	Class probability read as the chance the detector is correct on similar documents (source: gptzero.me)
DeAIze detection	Sentence-level AI probability	Which lines carry the signal, plus a document score	Our own predicted score, not a third-party verdict

The practical takeaway: a plagiarism match points at a source you can open. An AI score points at a writing pattern you have to edit. Fix the first with citation, the second with sentence-level rewriting.

If you want to see which specific sentences are carrying the signal before you rewrite anything, run the text through DeAIze’s AI detector — sentence-level highlighting shows you where the flag lives, so you edit five lines instead of the whole document.

The edits that actually lower a rate

Rewriting to dodge a detector is the wrong frame. Rewriting to make your own argument clearer is the right one, and it happens to move the score. Work in this order.

Fix attribution first. If the report links a source, add the citation, put borrowed phrasing in quotation marks, or restate the point in your own words while keeping the reference. This is the only part of the process where the plagiarism tool is giving you a factual answer rather than a guess.

Break the uniform rhythm. Machine-generated and heavily templated text tends to hold a steady sentence length and a steady clause structure. Read a flagged paragraph aloud. Where every sentence lands at the same length, split one and merge two. Where every sentence opens with a transition word, delete most of them.

Replace abstract nouns with the specific thing you mean. “The implementation of effective strategies” becomes “what the team did in week three”. Detectors and human readers both respond to this, because specificity is hard to fake at scale.

Add one detail only you could supply. A date, a room, a number from your own dataset, a disagreement you had with a supervisor. This is the single highest-value edit and it cannot be automated, which is exactly why it works.

Re-score, then stop. If the number moved down and the meaning survived, you are done. Chasing a zero produces stiff, unnatural prose that reads worse to a human marker than the original flag did.

Here is an illustration of the pattern, not a real report. Before: “It is essential to consider the various factors that contribute to effective communication within organisational contexts.” After: “Managers who send one clear ask on Monday get faster replies than managers who send a five-point memo on Friday.” The second sentence carries a claim, a timeframe and a comparison. It also has a rhythm a person would actually use.

What to do when the flag is wrong

False positives are real and documented by the vendors themselves. Turnitin’s own page acknowledges the risk, noting that human-written documents can be flagged, and its policy of withholding scores below the 20% threshold exists precisely to avoid over-flagging (source: turnitin.com). GPTZero’s published tuning claims that when its high-confidence category applies, 99.1% of human articles are classified as human and 98.4% of AI articles are classified as AI (source: gptzero.me). Those are the vendor’s own figures about its own tuned categories — not a general accuracy rate you can apply to your specific document.

If you are accused, the useful response is documentation, not argument. Keep your drafts, your notes, your version history, your source list, and any research artefacts. If you write in a second language, say so plainly — formal, heavily edited academic English from a non-native writer is a well-known trigger, and why an AI detector flags non native English writers walks through the specific patterns involved.

It also helps to know what the two accusations actually mean legally and institutionally. Copying someone else’s work and producing machine-generated text are different failures with different consequences, which AI detection vs plagiarism sets out clearly. If your case is already at a panel stage, what an academic integrity hearing assesses is the more useful read than any detector explainer.

When this is the wrong choice

Be honest about the boundary. Editing to lower a score is the wrong move in three situations.

• You did submit work you did not write. Rewriting the surface does not fix that, and no tool should be used to obscure it. The ideas, data and conclusions have to be yours.
• Your institution bans AI assistance outright. Then the fix is a conversation with your tutor about what you used, not a humanizing pass. Check the policy before you edit anything.
• You are editing someone else’s text to hide its origin. That is misrepresentation regardless of which detector is involved.

If none of those apply — you wrote the draft, you used a model for parts of it within the rules, and now a tool is reading it as machine-made — then sentence-level editing is a legitimate, ordinary part of the writing process. The humanize AI text pillar guide covers the full method, and AI checkers flagged you handles the case where several tools disagree with each other.

A short routine you can run today

1. Open the report. Write down which signal flagged you: source match, AI score, or both.
2. For every source match, add the citation or rewrite the sentence while keeping the reference.
3. Read the flagged paragraphs aloud. Mark every sentence that starts with a transition word and every sentence longer than about thirty words.
4. Rewrite those marked sentences only. Leave the rest alone.
5. Insert at least one concrete detail per flagged section that only you could know.
6. Re-score. If the number dropped and the argument still holds, publish or submit.

One more thing worth internalising: a detector score is a probability estimate, not proof of authorship. DeAIze reports its own predicted scores and does not guarantee any result on any third-party tool. Neither does anyone else, whatever their marketing says. Treat the number as a signal about your prose, fix the prose, and keep the receipts.

---

Measurement note: figures in this article come from our own detector run on 2026-09-28 — n = 709 hand-written guides versus n = 12 unedited model outputs, median AI score 25% and 34% respectively. Method: /methodology/.