How-ToGuide
Plagiarism Checkers vs AI Detectors: How to Read Both Reports
Many schools and publishers now run submitted writing through two different tools and hand back two different numbers. They sound similar and are often shown on the same screen, but they answer unrelated questions. One looks for text that already exists somewhere else. The other guesses how a text was produced. Reading them well starts with knowing which is which.
On this page
What a plagiarism checker compares
A plagiarism checker is really a text-matching engine. It breaks your document into overlapping strings of words and looks for the same strings in its database: published articles and books, web pages and, in many institutional systems, earlier student submissions. Crossref, the scholarly publishing organisation, describes its Similarity Check service this way: editors upload a manuscript, it is compared against a large corpus of published academic and general web content, and the result is a similarity report with a score and a highlighted set of matching text. Crossref is clear about the next step: editors review the matches and make their own decision about originality.
That last point matters. A similarity score is the share of your text that matches something else. It includes properly quoted and cited passages, common phrases, the title of a law or a standard, a reference list, and your own earlier work if it was submitted before. None of those is plagiarism. The report is a map for a person to read.
What an AI detector estimates
An AI detector has no database of sources to compare against. It runs your text through a statistical model and estimates how likely it is to be machine-written, usually based on how predictable the word choices are. The method is set out in the longer explanation of AI detector scores. Independent testing has been unkind to these tools. A 2023 study led by Debora Weber-Wulff of HTW Berlin tested fourteen detection tools and concluded that they were neither accurate nor reliable, with a tendency to call AI text human, and that disguising the text made them worse. A Stanford study the same year found that several detectors flagged a majority of essays by non-native English writers as machine-written.
| Plagiarism or similarity checker | AI detector | |
|---|---|---|
| Compares against | A database of existing text | Nothing; it models word patterns |
| Output | Similarity percentage plus highlighted matches with sources | A probability or percentage, sometimes with highlighted passages |
| Can you check the result? | Yes, open each matched source | Not directly; there is no source to open |
| Typical false alarm | Quotes, references, common phrases, your own earlier work | Formal, formulaic or second-language writing |
| Typical miss | Paraphrased or translated copying | Edited machine text |
Why the two often disagree
Because they measure different things, disagreement is normal rather than suspicious. An essay made mostly of carefully cited quotations can show a high similarity score and a low AI score. An original essay written in plain, careful English can show almost no matches and a high AI score. Machine-generated text is usually new wording, so it often matches nothing in a similarity database, while copied human writing can pass an AI detector easily. Neither tool can tell you about the other's question.
Read both reports step by step
Open the similarity report, not just the score
Click through the highlighted matches and look at each source. A score alone says nothing about whether any match is a problem.
Sort the matches
Separate quoted and cited passages, references, standard phrases and your own earlier work from anything that looks like uncredited borrowing.
Check paraphrases against sources
For each remaining match, compare your wording with the source. Close paraphrase without credit is the real issue; a citation fixes most of it.
Read the AI score as an estimate
Check the text length and genre. Short, technical or second-language writing is where false positives cluster.
Answer with your writing record
If either report raises questions about your own work, the guide on what to do after an AI flag explains how to present drafts and version history.
A note on privacy and watermarks
Pasting a draft into an online checker sends it to someone else's server, and some institutional systems keep submitted work in their database for future comparisons. Before using a free online tool, read what happens to text you paste into online tools. And neither report is a watermark test; the guide to what a text watermark can and cannot prove explains that separate idea. For more guides, see other how-to walkthroughs.
Common questions
What is a normal similarity score?
There is no universal threshold. Reviewers look at what matched, not the percentage; a long reference list alone can raise the number.
Can a plagiarism checker detect AI writing?
Not by matching. Generated text is usually new wording, so it rarely matches a database. Some products add a separate AI score, which is a different measurement.
Can an AI detector detect plagiarism?
No. It does not compare your text with any source, so it cannot show copying.
Should I run my work through both before submitting?
Only if the rules allow it, and only with tools whose data policy you have read. Keeping good notes and citations is a better safeguard.
Sources
- Crossref: Similarity Check: comparison against published and web content, reviewed by editors.
- arXiv: Testing of Detection Tools for AI-Generated Text: fourteen detectors found neither accurate nor reliable.
- arXiv: Study finding AI text detectors are biased against non-native English writers: false positives on non-native english essays.