Understanding Your Similarity and AI Report: What the Percentages Actually Mean

What a similarity score really measures

A similarity score is the percentage of your text that matches sources in the detector's index — other student papers, journals, books, and the open web. It is not a plagiarism score. A correctly cited quotation will show up as a match. Your own bibliography will show up as a match. Common phrases like "the results suggest that" will show up as a match. Plagiarism is about intent and attribution; similarity is about overlap.

That distinction is why every serious university reads the report, not just the number. A 25% report that is all quoted, cited material is fine. A 6% report where the 6% is an uncited paragraph copied from another thesis is not.

What counts as a "safe" similarity score

There is no universal threshold — every institution sets its own — but the bands below reflect what most departments treat as routine.

• Under 10% — typical for original work with normal citations. Rarely investigated.
• 10–20% — common for literature-heavy chapters, methods sections, and STEM papers with shared terminology. Usually fine if the matches are short, distributed, and properly cited.
• 20–30% — borderline. Markers will check whether the matches are quoted, paraphrased correctly, or clustered in one section.
• Above 30% — almost always triggers a closer review. Not automatic misconduct, but you should be able to explain every block.

Long, technical documents tend to score higher than short essays, because methodology, references, and standard definitions repeat across the field. A 22% dissertation can be cleaner than a 9% op-ed.

How AI-detection scores work

AI detectors — including Turnitin's AI checker and Turnitin AI detector — estimate the probability that a stretch of text was generated by a large language model. They look at statistical fingerprints: how predictable each word is given the previous ones (perplexity), how uniform the sentence rhythm is (burstiness), vocabulary choices, and punctuation patterns. Human writing tends to be uneven; LLM output tends to be smoother and more average.

The output is a probability, not a verdict. A 60% AI score does not mean 60% of your paper was written by AI. It means the detector's model is, in aggregate, fairly confident some portion was AI-generated. Some tools also highlight the specific sentences they flagged.

What an AI score actually tells you

• 0–20% — reads as human-written. Most original drafts land here.
• 20–50% — mixed signal. Often triggered by heavy editing, translated passages, or AI used for grammar polish.
• 50–80% — the detector thinks substantial sections look AI-generated. Worth reviewing sentence by sentence.
• Above 80% — high-confidence flag. Expect a conversation if your institution checks.

False positives are real. Non-native English writing, highly formulaic methods sections, and text that has been paraphrased through a tool can all push the score up without any LLM involvement. False negatives are real too — short prompts, careful editing, and newer models all reduce detection accuracy. Treat the AI score as a flag for review, not a confession.

How to read the report, not just the numbers

1. Open the full report, not the summary tile. Both numbers hide the detail you need.
2. Filter the similarity match list. Exclude the bibliography, quotes under a set length, and small matches — the percentage usually drops sharply once standard noise is removed.
3. Look at where matches cluster. A single dense block is more concerning than the same percentage spread across the whole document.
4. For AI flags, read the highlighted sentences out loud. If they sound like you and match your draft history, you have a clear answer for any follow-up.
5. Keep your drafts and notes. Version history is the strongest defence against a false-positive AI flag.

When you need a report before submitting

Most universities only run the detector after you submit, which means you don't see the report until it's too late to change anything. Running your own check beforehand lets you find the unintentional matches, fix the citations, and rewrite anything that reads as AI-generated — before the version that counts is graded.

That's what similarityreports.com does: a pre-submission similarity and AI report on the same Turnitin systems your institution uses, delivered in as little as one hour.