Plagiarism

Plagiarism spans a spectrum of distinct behaviors, and the distinctions matter because they are handled differently in practice.

Types

• Verbatim copying — reproducing another author’s exact wording without quotation marks or citation.
• Mosaic (patchwork) plagiarism — stitching together lightly reworded fragments from one or more sources without adequate attribution; individually each fragment may look paraphrased, but the underlying structure and content are not the author’s own.
• Self-plagiarism (text recycling) — reusing substantial portions of one’s own previously published text without disclosure. This is a distinct case from plagiarizing another author: under the US federal research-misconduct definition (42 CFR Part 93), plagiarism is scoped to appropriating another person’s work, so self-plagiarism does not meet the federal misconduct bar — but it is still treated as a publication-ethics violation by COPE and most journals, because it misrepresents the novelty of the submission. See CASRAI’s self-plagiarism entry for the full distinction.
• Plagiarism of ideas — using another person’s uncredited intellectual contribution (a hypothesis, an analytical approach, an unpublished insight encountered in peer or grant review) even when every word is the plagiarizing author’s own. ORI treats this as a separate category from plagiarism of text specifically because it cannot be caught by text matching.

What similarity software does and does not detect

Tools like Turnitin and iThenticate (the engine behind Crossref’s Similarity Check service) generate a similarity report by matching a submission’s text against a reference database — archived web content, prior submissions, and licensed publisher content. That process can flag verbatim copying and much mosaic plagiarism, where the wording itself overlaps a known source. It structurally cannot detect plagiarism of ideas, because there is no string match to find when the wording has been fully reworded. It also cannot, by itself, distinguish properly quoted-and-cited material, common boilerplate phrasing (methods-section stock language), or an author’s own previously published text from actual misconduct — all of these can inflate a similarity score without being plagiarism at all.

Why a similarity score is not a plagiarism finding

No vendor, COPE, or the US federal misconduct regulation (42 CFR Part 93) sets a field-wide numeric similarity threshold above which a submission is deemed plagiarized. Any specific percentage circulating informally (commonly figures around 15-20%) is a local journal or institutional policy choice, not a standard. A similarity report is an investigative aid that a human reviewer, editor, or integrity officer must interpret in context — it identifies passages worth checking, not a verdict. See CASRAI’s acceptable similarity percentage and originality report entries for more on how these numbers are actually used.

The institutional process, described neutrally

When a similarity report or a reviewer flags a concern, the typical path — described here as process, not as an accusation against anyone — moves through: an initial editorial or institutional screening to decide whether the concern is credible enough to pursue; if so, a preliminary inquiry to establish whether a fuller investigation is warranted; a formal investigation, conducted under an institution’s or journal’s own misconduct policy (in US federally funded research, this tracks the ORI process under 42 CFR Part 93), during which the respondent has a defined opportunity to respond; and a determination by the responsible body, which may range from no action, through correction or a mandated citation fix, to a finding of misconduct with consequences up to retraction. At every stage before a determination, the matter is an allegation under review, not an established finding — see CASRAI’s research misconduct entry for the fuller process.

Example: A submission returns a 35% Turnitin similarity score driven mostly by a correctly quoted-and-cited literature review and boilerplate methods language — an editor reviewing the highlighted matches determines no plagiarism occurred.

Counter-example: A submission returns a 12% similarity score, but the flagged 12% consists of unattributed verbatim paragraphs lifted from a single uncited source — a low overall score does not clear it.

References

• ORI, Plagiarism and Plagiarism of Text vs. Plagiarism of Ideas guidance (ori.hhs.gov)
• 42 CFR 93.227 (Plagiarism) and 93.234 (Research misconduct) — the US federal definitions; 93.103 sets the requirements for a finding. Older institutional SOPs may still cite the plagiarism definition as 93.103(c)
• COPE Core Practices; COPE flowcharts for suspected plagiarism
• Turnitin, “Understanding the similarity score”