What a percentage can and cannot tell you

Reviewed by Benjamin Lohrer, editor at corrected.eu · Updated: 11 August 2026

Meaning, limits and interpretation of the values

A plagiarism check and an AI detector give you a number, but not a verdict. This guide explains what both methods can tell you, where their limits are and how to interpret the result: why a plagiarism value above 0 per cent often does not mean real plagiarism, why correctly cited passages and the bibliography are flagged by mistake, and how to read the result properly.

Why a percentage does not yet prove cheating

Both the check and the AI detector are probability tools. They measure anomalies, not facts. A hit is a hint that a person has to follow up, not proof. That is where the central limits lie: both methods narrow down suspicious passages, and the actual assessment remains the job of the examiner.

If you know what the values can tell you, you read the report more calmly. A high value is not a guilty verdict, and a low value is not a free pass. The following sections show the concrete limitations: where the source check systematically flags too much and where the AI detector reaches its limits.

The first page of your report summarises everything: the plagiarism share, the distribution of the hits across the document and the split by levels of similarity. The number alone says nothing yet: what matters is which passages it comes from.

Correctly cited passages often count as plagiarism

A plagiarism check compares your text with sources and reports every match. The problem: even correctly cited and properly referenced passages are verbatim matches with the original source. The scanner cannot reliably tell whether a passage is cited sufficiently or insufficiently, so it flags it as a precaution.

You have to ignore such hits deliberately when you read the report. That is exactly why you often do not reach a value of 0 per cent even though every source is correctly given. A direct quotation with a source reference is unproblematic, but it appears in the report as a match.

A frequent technical special case reinforces this effect. If typographic quotation marks come out as question marks or other special characters when your file is exported to PDF or converted, the software no longer recognises the quotation style reliably. Even neatly set verbatim quotations are then flagged as a match, although the source is given correctly. In that case, check whether the quotation marks display correctly in the file you uploaded.

A typical example from an unchanged report: the quotation in quotation marks is backed by a source and is still flagged as a match, just like the unreferenced borrowing directly after it. To the software, both look the same at first.

Look at the similarity of each flagged passage

How strongly a flagged passage matches its source can be read in two places: from the colour of the marking in the text and from the percentage of the source in the source list. Yellow stands for a similarity of 35 per cent or more, orange for 75 per cent or more and red for 90 per cent or more. If a source stays at 50 per cent or below, a false positive is very likely: the scanner has then found only a few commonplace phrases again, not a passage that was taken over.

Conversely, red passages and sources with more than 80 per cent similarity deserve a close look. You check such hits one by one, fastest through the source comparison in the online report, which places your text and the source side by side. Often it is a correctly set verbatim quotation, but sometimes it is a passage that lacks its reference.

When you select a source, you see its degree of similarity, the number of matching words and the share of the whole text.

The bibliography and the references are counted too

The plagiarism check reacts particularly often to the bibliography. References, titles, publishers and author names are almost identical in many papers because they come from the same works. The technology behind our scan, PlagAware, compares these lines with huge source collections and reports them as matches, although a bibliography by nature consists of other people's details.

In an unchanged report, almost every line of a list of journal articles with DOIs can be marked. These hits can be explained technically and are no sign of cheating. When you evaluate the report, ignore the bibliography and pure source references mentally and concentrate on the running text.

Two evaluations: the filtered report and the PlagAware original

So that you do not have to work out these false hits yourself each time, we supply two PDFs with every plagiarism check. The filtered report carries the label "False-positives filter (BETA)" and leaves out exactly the hits described here: referenced quotations, bibliographies and very short chance matches. You also receive the PlagAware original unchanged, with everything the software found.

In the online report you switch between the two evaluations with a single click. The switch sits directly under the plagiarism share, and the value at the top changes with it:

• With the filter active, the same paper shows 9.4 per cent, of which 5.6 per cent is on the highest level of similarity.
• One click on "Original" and it is 18.9 per cent: the difference is the referenced quotations and the bibliography.

Both numbers are right, they just answer different questions. The original tells you how much of the text already stands somewhere else. The filtered report tells you how much of it is left after the explainable hits are deducted, and so comes closer to what an examiner finally assesses. In this example the value almost halves, and exactly these 9.5 percentage points are what you would have to subtract mentally when reading the original.

The filter deliberately carries the addition BETA: it works by rules and can miss a special case, such as a quotation with unusual punctuation. That is why the original is always included, not replaced. If the filtered value differs strongly from the original, it is worth looking at the original to see which passages the filter removed.

Where in the document most false alarms arise

It helps to read the flagged passages by their position in the document. At the beginning, the title page, the table of contents and other lists produce many hits, and at the end the list of sources and the bibliography. These areas naturally consist of formulaic or third-party details and deliver particularly many false positives.

The actually relevant part is the running text in between. Concentrate your check on this middle area, because that is where it is decided whether unreferenced borrowing is present.

Why universities look more closely only from 10 to 15 per cent

Because quoted passages and the bibliography raise the value artificially, a low percentage is completely normal. That is why many universities look more closely only at values above 10 or 15 per cent to see whether real plagiarism is present. The threshold is not a fixed limit but a rule of experience, above which a close look pays off.

Below this range, the reported share mostly sits in correct quotations and bibliographies. Only above it does the probability grow that unreferenced borrowing hides underneath. The meaning of the value therefore does not rise in a straight line, but only from a relevant threshold.

A value like this is harmless: 2.8 per cent in total, of which only 0.3 per cent on the highest level of similarity. The display stays in the green range.

AI detector: false positives and false negatives

With the AI detector, the meaning of the result is limited even more than with the plagiarism check. Two figures are decisive: the false-positive rate (a text written by a human is classified as AI) and the false-negative rate (an AI text is not recognised). Both are noticeably higher on German, soberly academic texts than the advertised accuracy values suggest.

The result of an AI detector therefore serves only as an assessment for examiners, who decide on that basis whether a follow-up is worthwhile. It replaces neither a conversation nor a documented writing process. How reliable the tools really are is compared in the AI detector comparison.

Unlike the similarity scan, the AI check does not give a percentage but a score from 0 to 100 in four levels: up to 25 a text counts as human, up to 40 as inconspicuous, up to 50 as acceptable, and above that as conspicuous. The strip above the report also shows whether the conspicuous passages are spread across the whole text or sit in individual sections.

A score of 61 is in the conspicuous range. The strip below it shows where in the text the conspicuous passages sit. What remains important: the value is a statement of probability about writing style, not proof that a text was produced by a machine.

In the report itself, passages typical of AI are marked red, while white and green passages count as written by a human. This is exactly where the limit of the method shows: sober, formulaic sentences stand out even if you wrote them yourself. A single red section is therefore no reason to worry, but a hint to sharpen the passage once more in your own words.

How to interpret your result properly

If you want to interpret both values properly, read the report passage by passage instead of staring at the total. Mark quoted passages and the bibliography as uncritical and check the remaining matches specifically for missing references. With the AI detector, treat the result as an indication and secure your writing process through interim versions.

In this way you use both methods sensibly without overrating their limitations, and you can interpret any value with confidence. A full plagiarism check and the corrected.eu AI detector show you exactly these sentence-level markings, so that you know before submission which passages really need attention.

Frequently asked questions about the meaning of the results

Where to go next

If you also want to see how the tools rank against each other, read AI detectors compared. To avoid problems in the first place, see how to avoid plagiarism.