03 · What You Need to Know
How Similarity-Detection Software Responds to AI-Generated Academic Writing
What does similarity-detection software actually detect?
Similarity-detection software compares submitted text with material available in its reference databases. These databases may include academic publications, web pages, previously submitted manuscripts, and other accessible documents.
Turnitin explains that its Similarity Report highlights matching or similar passages and calculates an overall similarity percentage based on detected textual overlap.
The system does not determine plagiarism simply by identifying matching language. A legitimate quotation, properly cited passage, standard methodological expression, or previously submitted draft may generate matches.
Conversely, a manuscript may contain unattributed intellectual borrowing that produces few detectable matches because the wording has changed substantially or the source is unavailable for comparison.
The similarity report is therefore a source-comparison tool rather than a direct measure of research integrity.
Why can AI-generated text match existing publications?
Generative AI systems produce language using patterns learned during training. Their outputs may contain conventional academic expressions, familiar methodological descriptions, or passages resembling previously published material.
Research on language-model memorization has demonstrated that some models can reproduce portions of their training material. However, memorization is not the only explanation for textual overlap.
Two independently produced passages may share wording because they describe the same established concept or use standardized disciplinary terminology.
For example, an AI-generated Methods section might contain familiar expressions about participant recruitment, statistical significance, or ethical approval. Some of those expressions may match existing manuscripts without necessarily indicating plagiarism.
More distinctive or extensive overlap warrants closer examination, particularly when it involves substantive source-dependent explanations.
The possibility of AI-generated text reproducing previously published phrases is one reason researchers should not assume that newly generated wording is necessarily original.
Is Turnitin's similarity score the same as its AI-writing score?
No. Turnitin provides separate functions for identifying textual similarity and estimating AI-generated writing.
| Feature |
Similarity Report |
AI Writing Report |
| Primary purpose |
Identify matching or similar text in available sources |
Estimate whether qualifying text may have been generated or modified using AI |
| Basis of assessment |
Comparison with reference databases |
Model-based classification of writing characteristics |
| Reported percentage |
Proportion of text matched to sources under the report's settings |
Proportion of qualifying text classified as potentially AI-generated or AI-modified |
| Identifies original sources? |
Provides information about matched sources |
Does not establish the original scholarly source of ideas |
| Proves plagiarism? |
No |
No |
| Proves AI use? |
No |
No, the classification may be inaccurate |
Turnitin explicitly states that its AI-writing percentage is independent of the similarity score. The two reports answer different questions and should not be interpreted interchangeably.
A manuscript can therefore have substantial similarity and a low AI-writing indicator, or little similarity and a higher AI-writing indicator. Neither combination automatically establishes misconduct.
Can AI-generated text receive a zero similarity score?
Yes. A similarity report may identify no matching text within the available comparison sources and report a score of zero.
However, zero similarity does not establish that the writing is entirely original, ethically acceptable, or free from unattributed intellectual borrowing.
A generated passage might communicate an existing scholar's distinctive argument using different language. It may also reproduce material from a source not included in the comparison database.
Turnitin's own guidance emphasizes that low similarity does not guarantee the absence of plagiarism.
The implications of low similarity scores in AI-assisted academic writing therefore require consideration of source attribution and intellectual provenance, not merely matching percentages.
Can AI-generated text produce a high similarity score?
Yes. AI-generated writing may match existing publications, web content, or previously submitted documents.
A high score can also result from factors unrelated to AI generation, including extensive quotations, references, standardized descriptions, or the reuse of material from an earlier manuscript.
For example, a manuscript containing a lengthy instrument description or conventional procedural wording may produce substantial matches even when the research itself is original.
The appropriate response is to inspect the matched passages and determine whether the overlap is legitimate, properly attributed, or potentially problematic.
No universal similarity percentage can determine whether a manuscript should be accepted, rejected, or classified as plagiarized.
Does AI paraphrasing affect similarity scores?
It can. Substantial rewriting may reduce direct textual overlap even when the resulting passage remains dependent on the original source.
This limitation is not unique to AI. Conventional paraphrasing can also produce new wording while preserving borrowed ideas or arguments.
Some detection systems incorporate methods intended to recognize particular forms of AI-modified writing. However, those methods do not guarantee identification of the original intellectual source or establish that plagiarism occurred.
Researchers should therefore distinguish the detection of textual transformation from the ethical assessment of attribution.
The relationship between AI rewriting and reduced visibility of plagiarized material illustrates why similarity scores cannot replace direct engagement with sources.
What are the limitations of AI-writing detection?
AI-writing detection systems estimate whether text exhibits characteristics associated with AI-generated or AI-modified language. These estimates can be inaccurate.
Turnitin acknowledges that its AI-writing model may misclassify human-written, AI-generated, or AI-paraphrased material. Its documentation advises against using AI-writing results as the sole basis for adverse academic decisions.
Turnitin also treats low AI-writing percentages cautiously. Its current reporting approach does not display exact percentages or highlights for results between 1% and 19%, instead showing an asterisk because such low-range classifications are less reliable.
This reporting threshold is not a declaration that AI use below 20% is permissible. It is a measure intended to reduce misinterpretation of uncertain detection results.
Detection capabilities also depend on supported languages, document characteristics, and the version of the system being used.
Watch Out
Do not confuse an AI-writing indicator with a similarity percentage. A 20% AI-writing result does not mean 20% of the manuscript is plagiarized, and a 0% similarity result does not mean the manuscript was written without AI. Both metrics describe different forms of automated assessment.
Can AI-generated references or quotations affect similarity results?
Yes, although their effects depend on the software's settings and the material being compared.
Reference lists may contain bibliographic entries that match other documents. Direct quotations may also produce matches even when they are correctly attributed.
Turnitin allows certain matches, quotations, and bibliographic material to be excluded from similarity calculations, depending on the report and available settings.
These exclusions can help reviewers focus on potentially meaningful overlap. However, they do not change whether the underlying citations are accurate or whether the quotations have been used appropriately.
AI-generated references require independent verification because a plausible citation may refer to a nonexistent publication or inaccurately describe a real source.
Should researchers use similarity software to verify AI-generated writing?
Similarity software can be useful as one component of a broader manuscript review. It may identify passages requiring attribution checks, quotation adjustments, or closer comparison with existing publications.
However, it cannot verify every scientific claim, identify all unattributed ideas, or determine whether AI-generated interpretations accurately represent the cited literature.
Researchers should combine source comparison with direct reading, citation verification, and critical evaluation of the manuscript's intellectual content.
When using AI to assist with checking whether academic writing is too close to a source, the same limitations apply. Automated assessments should support, not replace, scholarly judgment.
Does a similarity match prove that AI copied the source?
No. A similarity match establishes a textual relationship between the submitted material and a comparison source. It does not independently explain how that relationship arose.
The overlap might result from conventional language, direct quotation, legitimate reuse, independent formulation, or inappropriate copying.
Even when the passage is known to have been generated by AI, the match does not prove that the model reproduced the source through memorization. Similar wording can arise through different mechanisms.
Determining the academic significance of a match requires examination of the source, the content, the attribution, and the relevant circumstances.
