Similarity score

The mechanics are string matching at scale: the checker compares your document against web pages, published articles, and previously submitted student work, and highlights every overlapping passage. The percentage is the share of your words that appear in those matches. Crucially, the software has no concept of legitimacy — a properly quoted, cited passage matches exactly the same way a stolen one does.

That is why the number means less than students fear. Bibliographies, quoted material, methods boilerplate, and assignment-prompt language all inflate it; a high score on a quote-heavy literature review can be entirely honest, and a near-zero score can hide a stolen idea reworded top to bottom. This is also why no universal "safe" threshold exists: whether the review line sits at 15% or 25% is each institution's policy, not science, and the report is designed to be read by a human, not enforced as a verdict.

A similarity score is also a different instrument from an AI-detection score, though the two now often sit side by side in the same dashboard. Similarity asks does this text match existing sources? — a checkable string-matching question. AI detection asks does this text look machine-generated? — a statistical inference with a documented false-positive problem, especially for non-native writers (see flagxiety). Conflating the two is a category error. And the classic way honest work inflates similarity is patchwriting: paraphrase that stayed too close to a cited source.