03 · What You Need to Know
Why Similarity Scores Cannot Determine Whether AI-Assisted Writing Is Ethical
What does a similarity percentage actually measure?
A similarity report compares submitted text against material accessible to a particular system. It identifies matching or sufficiently similar passages and calculates a percentage according to its matching and reporting rules.
Turnitin explains that its Similarity Report highlights matching text from its comparison databases. The resulting score is not a plagiarism determination.
For example, a reported similarity of 5% generally means that the system identified matching text corresponding to approximately 5% of the submitted material under the applicable report settings. It does not mean that the remaining 95% has been independently verified as original.
Database coverage, quotation exclusions, bibliographic material, and report configuration may affect the percentage. Different systems can produce different results for the same manuscript.
Why can AI-generated plagiarism produce a low similarity score?
Plagiarism can involve the appropriation of intellectual material without retaining its original wording.
Suppose a researcher uses generative AI to rewrite a published theoretical explanation. The resulting sentences may differ substantially from the original, but the argument could still reproduce another author's distinctive intellectual contribution.
A text-matching system may identify little overlap because its primary evidence is linguistic similarity rather than the complete intellectual relationship between two works.
Research on language-model plagiarism and memorization also demonstrates that generated outputs may reproduce or reformulate existing material in different ways. The presence or absence of exact matches cannot establish the provenance of every idea.
This is why AI-assisted rewriting can make some plagiarism harder to detect without resolving the underlying attribution problem.
Can a manuscript have zero similarity and still contain plagiarism?
Yes. A zero score indicates that the software did not identify matching text under the conditions of that report. It does not establish that every intellectual contribution originated with the author.
A researcher could reproduce a distinctive argument in entirely new language while omitting the original source. Alternatively, a manuscript might contain material from a publication unavailable to the comparison database.
Neither situation necessarily produces a detectable textual match.
The appropriate question is whether borrowed ideas, methods, findings, and interpretations are attributed, not simply whether the wording resembles existing documents.
What ethical problems can a similarity checker miss?
Similarity-detection software is not designed to evaluate every dimension of academic integrity.
| Potential concern |
Can a low similarity score rule it out? |
What requires verification? |
| Unattributed paraphrasing |
No |
Whether distinctive ideas or arguments originated from another source |
| Fabricated references |
No |
Whether each cited publication exists and supports the claim |
| Invented research findings |
No |
Whether reported results correspond to genuine research evidence |
| Unsupported interpretations |
No |
Whether conclusions follow from the data and methods |
| Undisclosed AI assistance |
No |
Whether AI use was reported as required |
| Unauthorized AI use |
No |
Whether the assistance complies with institutional or journal rules |
A manuscript could contain none of these problems, several of them, or concerns unrelated to AI. The similarity score does not distinguish among those possibilities.
Is low similarity better than high similarity?
Not necessarily. Lower similarity may be desirable when it reflects independent writing and appropriate source use. However, the percentage alone cannot establish manuscript quality or integrity.
A paper containing accurately quoted passages, standardized instrument descriptions, or conventional methodological language may produce substantial legitimate overlap.
Conversely, a manuscript containing heavily paraphrased but unattributed intellectual material may produce very little overlap.
Consider two hypothetical manuscripts. One has 22% similarity because of correctly attributed quotations and methodological terminology. Another has 2% similarity after a distinctive published argument has been rewritten without citation.
The second manuscript is not necessarily more ethical simply because its percentage is lower.
The meaningful assessment concerns the nature of the matches and the provenance of the ideas.
Does a low similarity score mean the writing was not generated by AI?
No. Similarity detection and AI-writing detection are separate processes.
Similarity score
Reports textual overlap with available comparison sources under specified matching conditions.
AI-writing indicator
Estimates whether qualifying text exhibits characteristics associated with AI-generated or AI-modified writing.
Turnitin provides separate reporting functions for these purposes. Its AI-writing indicator is not calculated from the similarity percentage.
Neither metric independently proves misconduct. AI-writing classifiers can misidentify human-written text, while similarity reports can miss unattributed intellectual borrowing.
The distinction is essential when interpreting similarity-detection results for AI-generated academic text.
Does a low score establish that the researcher contributed intellectually?
No. Authorship involves intellectual contribution and accountability, neither of which can be measured through textual overlap.
A researcher might submit an entirely AI-generated discussion containing no identifiable copied passages. Even with a zero similarity score, the researcher must still establish that the interpretations are scientifically defensible, the sources are genuine, and the use of AI complies with the relevant publication requirements.
The ICMJE requires human authors to remain accountable for AI-assisted material, including its accuracy, attribution, and integrity.
The distinction between submitting AI-generated academic text under one's own name and plagiarism is therefore important. A submission may raise authorship or policy concerns without containing identifiable copied text.
Can an acceptable similarity threshold be determined universally?
No universal percentage establishes that a manuscript is free from plagiarism or ethically acceptable.
Some institutions, journals, or academic programs may use numerical thresholds as screening or administrative criteria. Such thresholds are local requirements rather than universal scientific definitions of originality.
A manuscript below an institutional threshold may still contain inappropriate borrowing. A manuscript above it may contain legitimate matches requiring explanation rather than rejection.
Researchers should distinguish compliance with a local screening threshold from substantive compliance with academic-integrity standards.
Why can pursuing the lowest possible score become counterproductive?
When researchers treat similarity as a target to minimize, they may focus on replacing words rather than improving scholarship.
Repeated paraphrasing can remove technical precision, distort source meaning, or obscure the intellectual origin of ideas. It may also encourage researchers to avoid necessary quotations or standard terminology simply because these produce matches.
Generative AI makes extensive rewriting convenient, but linguistic novelty is not equivalent to scholarly originality.
Using AI to humanize copied academic text does not resolve the ethical problem if the underlying source remains unattributed.
The appropriate objective is accurate, independent, and properly attributed scholarship rather than a particular percentage.
What should researchers verify beyond the similarity report?
A responsible review examines the manuscript's intellectual provenance and evidential basis.
Researchers should check that cited publications exist, claims accurately represent the sources, quotations are properly attributed, and the manuscript's scientific interpretations are supported by the study's findings.
They should also examine whether AI introduced new facts or altered the meaning of previously drafted material. Finally, they must follow the applicable journal's AI-use and disclosure requirements.
These checks address different risks. No single automated percentage can replace them.
