Understanding Similarity and AI Detection Reports: What the Numbers Mean and What They Don't

Similarity checking and AI writing detection are often bundled in the same platform, most commonly Turnitin, but they work in completely different ways. A similarity report compares your text against a database of web pages, publications and previously submitted work, and highlights passages that match. It answers the question: does this text overlap with text that exists elsewhere?

An AI writing indicator does not compare your text with anything. It uses a statistical model to estimate whether passages have characteristics typical of text produced by large language models. It answers a much harder and less certain question: does this text look machine-generated? Turnitin's own documentation states that the AI score is independent of the similarity score, and the two should never be read as measuring the same thing.

Neither tool detects plagiarism or misconduct as such. Those are judgments that people make, following institutional procedures, after considering the evidence in context.

When you submit, the system breaks your document into strings of words and searches for matches in its repositories. Matching passages are highlighted and linked to their sources, and the overall similarity index is the proportion of your text that matches any source. The report usually lists sources in order of how much of your text they account for.

Several settings change the number substantially. Instructors or administrators can exclude quoted material, the bibliography, small matches below a set number of words, or specific sources such as your own earlier draft submitted to the same assignment. Whether these filters are on depends on the institution and assignment, so the same document can produce very different percentages in different settings.

Coverage also matters. A match can only be found if the source is in the database. Paywalled material, recent publications, some non-English sources and translated text may be missed, while commonplace phrases and template text may be matched.

There is no universal acceptable percentage. A literature-heavy review with properly quoted definitions may legitimately show more overlap than an original data analysis, and a short piece where a correctly cited quotation makes up a tenth of the text will show that tenth as a match. Markers are trained to look at what matched, not just how much.

Look through each highlighted passage and classify it. Most matches fall into a few categories, and only some require action.

AI detectors look for statistical patterns, such as how predictable each word is given the words before it, and how uniform sentence structures are. Text generated by language models tends to be highly predictable and even in rhythm. The detector outputs an estimate for segments of text and an overall percentage.

Vendors describe specific conditions. Turnitin, for example, states that its indicator needs at least 300 words of long-form prose, is not designed for poetry, code, bullet lists or tables, and shows scores from 1 to 19 percent only as an asterisk rather than a number because false positives are more likely at that level. Turnitin also states that the percentage should not be used as the sole basis for action against a student.

Used carefully, an indicator can prompt a conversation. It cannot establish who wrote a text, what tools were used, or whether any use broke the rules of a particular assignment.

Detectors also struggle with mixed authorship. A document in which a student wrote most paragraphs, used a permitted tool to rephrase one, and received tracked-change edits from a proofreader contains several layers of influence that no classifier can reliably separate. Segment-level highlights are estimates for stretches of text, not records of what happened while writing, which is why process evidence and conversation remain essential to any fair judgment.

Any statistical classifier makes errors. Even a low overall false positive rate means some genuine human writing will be flagged when many documents are checked. Researchers and educators have raised particular concern about writing by people using English as an additional language, whose prose may be simpler and more formulaic, and about highly conventional genres such as lab reports or legal summaries.

Legitimate tools can also shift text toward patterns the detector associates with AI. Grammar checkers, translation tools and heavy editing can make prose more uniform. Conversely, AI-generated text that has been rewritten may go undetected. For these reasons many universities have published guidance saying that AI indicators alone are not evidence of misconduct, and some have chosen not to enable AI detection at all.

• Short texts or short segments give the model little to work with.
• Formulaic genres use predictable phrasing by design.
• Writers using English as an additional language may write in simpler, more regular structures.
• Grammar, paraphrasing and translation tools can homogenize style.
• Heavily edited or co-edited text may lose individual voice.

The strongest protection against a mistaken concern is evidence of how you produced the work. Most of it is created automatically if you work in a cloud word processor or keep dated files, and it takes little effort to preserve.

If your course permits some use of AI tools, for example to brainstorm or to check grammar, keep a record of what you used, when and for what, and acknowledge it in the way your institution requires. Many universities now provide a template statement for acknowledging AI use.

• Keep version history in Google Docs or Microsoft 365, or save dated drafts.
• Keep notes, outlines, reading notes and annotated sources.
• Keep your reference manager library, which shows when sources were added.
• Save any permitted AI prompts and outputs, and note how you used them.
• Keep data files, analysis scripts and outputs for quantitative work.
• Save feedback from supervisors or writing centers and show how you responded.

Similarity reports also reveal issues that students do not always recognize as problems. Reusing substantial parts of your own previously submitted work without permission, often called self-plagiarism or duplicate submission, is prohibited at many institutions even though the words are yours. If you want to build on earlier work, ask your instructor and cite your previous submission as required.

Collusion means working with other students on assessed work that is meant to be individual, for example sharing drafts or answers. Similarity systems that check against other submissions in the same class can detect close overlap between students. Contract cheating, where someone else produces work you submit as your own, is treated as one of the most serious integrity breaches; it is often identified not by similarity software but by changes in style, metadata, or a student's difficulty discussing the work.

The line between legitimate help and misconduct depends on the assignment's rules. Feedback from tutors, writing centers and editors on your own draft is usually acceptable within institutional limits; having someone else write or rewrite content usually is not.

For instructors and integrity officers, good practice mirrors the vendors' own advice. Review the matched sources rather than the headline percentage, check the filter settings applied, and consider the assignment type. For AI indicators, treat the result as one signal to be weighed alongside other evidence, such as the student's previous work, drafts and version history, and a conversation about the content.

Many universities recommend designing assessment to reduce ambiguity, for instance by requiring process evidence such as proposals, annotated bibliographies or drafts, and by stating clearly which AI uses are permitted. Clear expectations protect students from uncertainty and make any necessary investigation fairer. Staff should also be aware of potential bias against writers using English as an additional language and avoid acting on AI scores alone.

Stay calm and read your institution's academic integrity procedure, which will explain the stages, your rights, and whether you can bring a representative such as a student union adviser. Ask to see the report and the specific concerns. A flag is the beginning of a process, not a finding.

Prepare a clear, factual explanation of how you wrote the work, supported by your version history, notes and sources. Be ready to discuss your argument, sources and choices in your own words, since a conversation about the content is often the most persuasive evidence of authorship. If you used a tool in a permitted way, say so plainly. If you made a genuine citation mistake, acknowledge it and explain how you will avoid it in future.

Before submission, a formative similarity check with a clear explanation of each match can help you catch citation problems while there is still time to fix them properly.

Imagine an invented 3,000-word master's essay that returns a similarity index of 24 percent. At first glance this might seem alarming. Opening the report shows that the bibliography accounts for 9 percent because the filter was not applied, three properly quoted and cited definitions account for 5 percent, and a scattering of discipline-specific phrases makes up another 4 percent.

The remaining 6 percent is the part worth attention. Two paragraphs in the literature review closely follow the sentence structure of a review article; the student had cited the article but paraphrased too closely, having written from notes that copied phrasing without quotation marks. The fix is to rewrite these paragraphs from an understanding of the source, possibly synthesizing it with other studies, and to keep the citation.

The example is illustrative, but the pattern is common: the headline number overstates the problem, while one small section reveals a genuine skill gap in paraphrasing. Reading the report source by source, rather than reacting to the percentage, is what turns it into useful feedback.

Most problems in similarity reports come from note-taking habits rather than intent. Copying phrases into notes without quotation marks, then forgetting their origin, is a common route to accidental close paraphrase. Keep quotations clearly marked in your notes, record page numbers, and write summaries from memory rather than from the open source.

Paraphrase properly by changing the structure as well as the words, and always cite the source. Quote sparingly and only where the exact wording matters. Check that every in-text citation matches a reference entry, and never try to lower a score with character substitution, hidden text or paraphrasing tools designed to evade detection; these tactics are themselves treated as misconduct at most institutions and are often visible to markers.