Turnitin AI Writing Report vs Similarity Report: do not read the scores as the same thing

A Turnitin Similarity Report finds wording in a submission that matches material in Turnitin's databases. An AI Writing Report uses a classification model to estimate what share of the submission's qualifying prose may be AI-generated, including text the model considers to have been further modified by an AI paraphraser or bypasser. The two calculations do not feed into one another.

The confusion is understandable. Both reports belong to the same submission workflow, and some Turnitin interfaces put an AI indicator beside the Similarity Report. In the newer interface the full result opens from an AI Writing tab; in the classic interface the indicator can sit in the right toolbar. That proximity is navigation, not shared mathematics. Turnitin's AI report guide explicitly describes the AI percentage as different from and independent of the Similarity Score.

A matching passage and an AI-classified passage are also different kinds of evidence. A similarity match points to another item - a web page, publication or earlier student submission - so a reader can compare the words and inspect attribution. An AI highlight points to a model category. It does not name a source, model, prompt, user or writing session.

Similarity asks, "Where else does this wording appear?" AI detection asks, "Does this qualifying prose resemble text in the model's AI categories?" Neither asks, or can answer by itself, "Did this student commit misconduct?"

That distinction determines the next action. A source match should lead you to the source, quotation and citation. An AI highlight should lead you to the surrounding prose, the writer's drafts and the applicable AI-use policy. Treating both as a generic "problem percentage" sends the reader to the wrong evidence.

The fastest way to avoid a category error is to compare the reports by question, denominator and next step. The table below uses Turnitin's current terminology; product access and layout can still vary by license and interface version.

The denominator is the most important technical difference. Turnitin defines the Similarity Score as the percentage of submission text that matches other sources. The AI percentage is calculated over qualifying prose only. A paper containing tables, code, equations, short-form responses or an annotated bibliography may therefore have an AI denominator much smaller than its visible page count. The detailed AI report guide explains that denominator and the asterisk state without repeating it all here.

The highlight semantics matter just as much. In the Similarity Report, selecting a match should take you toward a source and the overlapping words. In the AI Writing Report, a highlight represents a classification category, not a source link. It is invalid to look at an AI-highlighted sentence and say it was copied from ChatGPT; the report contains no comparison capable of establishing that provenance. There used to be two of those categories. Turnitin dropped the split on 4 August 2026, folding AI-paraphrased text into one blue AI-generated category and stating that purple no longer appears. Older reports still carry both colours until the file goes through again.

Report dates and settings belong in any comparison. A Similarity Report with quotations and bibliography included is not directly comparable with one that excludes them. An AI report generated or refreshed under a newer model may also differ from an earlier version. Record the file, report date, interface or model notice, and active exclusions before trying to explain why two screenshots disagree.

Finally, do not infer access rights from somebody else's screenshot. What a student can see depends on the institution, Turnitin product, license and assignment settings. If a decision uses a report you cannot open, ask for the complete report and the settings used, not merely a percentage copied into an email.

Because the reports are independent, none of the four high/low combinations is contradictory. Here, "high" and "low" are descriptive only; Turnitin does not provide one universal misconduct threshold, and institutions can apply different policies.

• Higher similarity, lower AI. A human-written literature review may quote extensively, use standard definitions, reproduce an assignment template, or have been submitted previously as a draft. Those features create database matches without making the prose look AI-generated. Review the largest matches and their attribution; do not rewrite properly quoted material just to move a percentage.
• Lower similarity, higher AI. Newly generated wording need not resemble any indexed source. It can therefore produce few source matches while the AI classifier labels part of the qualifying prose. A low Similarity Score does not rebut an AI result, just as the AI result does not establish authorship. Review drafts, permitted tools and the highlighted context.
• Higher similarity, higher AI. The same document can contain substantial quoted or reused material and separate prose that the classifier flags. It can also contain a template, long references or prior-submission matches alongside those classifications. Investigate the two findings separately; one percentage does not explain the other.
• Lower similarity, lower AI. This means neither system surfaced much under its own rules. It does not certify originality or human authorship. A source may be outside Turnitin's databases, a consequential passage may be small relative to the paper, content may be filtered or non-qualifying, and any classifier can miss cases.

Do not add, subtract or average the two scores. Twenty percent similarity plus twenty percent AI is not forty percent of one problem; the percentages use different denominators and describe different events.

This is why a traffic-light reaction to the cover page is weak practice. The useful unit is not "the score" but the passage and the question attached to it: What source matched? Is the overlap quoted? What text qualified for AI review? Is there drafting evidence? Only then does a number become a route into evidence rather than a substitute for it.

Turnitin's own Similarity Score guide is unambiguous: the system checks similarity against its databases; it does not decide whether plagiarism occurred. Its comparison set includes current and archived web pages, past student submissions, and periodicals, journals and publications. A match is therefore a location to inspect, not a misconduct label.

A high score can be innocent. Correctly quoted and referenced material still matches. Bibliographies contain titles that exist elsewhere. Methods sections may repeat required procedures. A draft deposited in a student-paper repository can match a later submission. Quote-heavy qualitative research can legitimately contain more shared text than an original mathematical proof.

A low score can still contain a serious problem. One uncited paragraph copied from a source may form only a small percentage of a long thesis. Patchwriting can distribute close paraphrases across multiple sources. An obscure, private or newly published source may not be in the comparison set. The aggregate percentage cannot tell you the seriousness of the most important match.

Read a Similarity Report in this order:

• Check the report settings first. Note whether quotations, bibliography, citations, assignment templates, small matches or repositories are excluded. Compare percentages only when those choices are materially the same.
• Start with passages, not the colour band. Open the largest sources, then scan for smaller but consequential matches. A 2% match may matter more than a 15% bibliography.
• Classify the reason for each meaningful overlap. Is it a direct quotation, a cited close paraphrase, common terminology, required instructions, the author's earlier submission, or unattributed borrowing? Each reason calls for a different response.
• Fix the scholarly problem, not the meter. Add a missing citation, quote wording that is too close, paraphrase from understanding, or document legitimate reuse. Deleting citations or laundering copied wording through a rewriter only hides the evidence while leaving the attribution problem intact.

The filters and exclusions guide explains that reports can exclude quotations, bibliography, in-text citations, small matches, templates and selected sources. These controls refine the calculation or view; they do not alter the submitted document. Exclusions can be restored, so the defensible practice is to disclose them rather than present a filtered score as though it were the only result.

Do not expect every exclusion to lower the percentage by the visible amount. One passage can match overlapping sources. Excluding one source may simply reveal another source for the same words, and Turnitin warns that the overall number may or may not change. This is another reason to preserve the report settings and reason from the passages.

If the problem is attribution, use the source match to correct the quotation or citation. Humanizing is not a citation remedy. HumanPen's citation format correction is for making existing citations and references follow a required style; it cannot invent evidence for an uncited claim.

Before uploading or revising anything, identify the document. Do not rely on a coloured percentage in a screenshot. Look for the report title and what the pages contain: a Similarity Report is organised around matches and sources; an AI Writing Report is organised around an AI indicator, qualifying text and AI classification highlights. The current access guide shows where the separate AI report opens in the new and classic interfaces.

• If you have a Similarity Report: inspect the source list, active filters and meaningful overlaps. Correct missing attribution and citation formatting in the document. Do not upload this PDF to HumanPen's Turnitin-report workflow; it does not contain the AI highlights that workflow needs.
• If you have an AI Writing Report: read the highlighted passages in context and compare them with drafts, notes and permitted AI use. If revision is allowed and no review is pending, the full report can define a narrow set of passages to revisit instead of rewriting the entire document.
• If you have only a percentage or screenshot: ask for the full report used in the decision, along with the report date and relevant settings. A number detached from its denominator, passages and source details is not enough to diagnose the issue.
• If you have both reports: keep them separate in your notes. Resolve source attribution from the Similarity Report and review AI classification from the AI Writing Report. Record which change answers which finding.

If an academic-integrity review or appeal has already started, preserve the submitted file and original reports unchanged. Work only on a copy after the applicable process permits revision. Version history, outlines, source notes and feedback can show how the text developed; overwriting them to chase a new percentage can destroy the evidence most useful to you.

HumanPen deliberately rejects Similarity Report PDFs in the AI-report upload flow. The product needs the AI Writing Report because that file identifies the passages selected for AI review. With the document and compatible AI report, the workflow maps those highlights back to the source document and limits rewriting to that selection. It does not treat the Similarity Score as an instruction to paraphrase source matches.

That narrower workflow is still not a guarantee about a future score. Detector models, versions and institutional settings change, and human prose can be misclassified. The defensible goal is to make permitted revisions you understand, preserve meaning and citations, and keep ownership of the argument - not to convert either report into a promise of clearance.

One submission, two diagnostic tools, two evidence trails: sources and attribution for similarity; prose context, process evidence and policy for AI classification. Start with the right trail.