Plagiarism detection answers a question about origin: has this text appeared somewhere before? It looks backward at an index and returns matches. If a sentence in a student paper appears verbatim in a published article, the plagiarism checker flags it.
Quick Answer
Plagiarism detection compares submitted text against an index of published sources, web pages, and prior student work, and returns matches. AI detection uses a statistical classifier to estimate the probability that a passage was generated by a large language model. A submission can be zero-plagiarism and still fully AI-generated. A submission can score high on plagiarism and be entirely human-written. The two systems answer different questions.
Understanding why the two are different starts with what each system was built to detect in the first place.
The Core Difference
AI detection answers a different question about authorship: does this text look like it was generated by a language model? It looks at statistical properties of the writing itself, without reference to any index. It returns a probability, not a match.
The two questions are related in the classroom sense that both bear on academic honesty, but the underlying methods share almost nothing. One is essentially a lookup problem. The other is a pattern classification problem. That is why a single number cannot tell you both things, and why tools that bundle them are running two separate engines under one interface.
Teachers who rely on plagiarism reports alone will miss most AI-assisted submissions. Teachers who rely on AI detection alone will miss lifted passages that a decent plagiarism engine catches instantly.
The Working Educators editorial position
How Plagiarism Detection Works
Plagiarism detection is old technology by AI standards. Turnitin has been doing it since 1998. The system indexes a large corpus, published journals, web pages, and in Turnitin's case the archive of prior student submissions, and compares the submitted document against that index using fingerprinting techniques that identify overlapping strings of a certain length.
What comes back is a similarity score and a set of highlighted matches. A 15 percent similarity report might be quotations properly cited, might be a common phrase, or might be lifted text. The number itself is not a verdict. The instructor has to look at what was matched and decide.
The important limit: plagiarism checkers can only find what is in the index. Text that has never been published, never been submitted, and does not appear on the crawlable web is invisible to them. That includes almost everything generated by a language model, because model output is novel text produced on demand.
How AI Detection Works
AI detection systems are classifiers. They are trained on examples of human writing and examples of model-generated writing, and they learn statistical signals that distinguish the two. Perplexity, burstiness, token distribution patterns, and semantic coherence are among the features detectors use, though methods vary by vendor.
The output is a probability. A detector returning 87 percent AI does not mean 87 percent of the words were generated. It means the model estimates an 87 percent likelihood that the passage was model-produced. Different vendors calibrate this probability differently, and thresholds for flagging are set by the platform or the instructor.
Accuracy claims in this category vary widely and depend heavily on what was tested. The Stanford study by Liang et al. found detectors flagged non-native English writing as AI-generated 61.3 percent of the time on TOEFL essays known to be human-written. The Krishna et al. DIPPER paper showed detector accuracy collapsing from 70.3 percent to 4.6 percent after light paraphrase. Vendor accuracy claims and independent testing rarely land in the same place.
We have written elsewhere on how to read those figures. Our guide on evaluating detector accuracy covers the questions teachers should ask before trusting a percentage.
Why a Clean Plagiarism Report Doesn't Mean Not AI
This is the practical trap. An instructor runs a paper through Turnitin, sees a two percent similarity score, and concludes the work is original. It might be. It might also be an entirely ChatGPT-generated essay.
Model-generated text is novel. It is written on the fly in response to a prompt. It does not exist in any index the plagiarism checker consults. From the checker's perspective, the paper is clean, because there is nothing to match against.
The reverse is also true. A student who quotes heavily from primary sources and cites correctly may pull a high similarity score. That is a citation issue, not a plagiarism issue, and it says nothing at all about whether AI was used.
Teachers who rely on plagiarism reports alone will miss most AI-assisted submissions. Teachers who rely on AI detection alone will miss lifted passages that a decent plagiarism engine catches instantly. The two are complements, not substitutes. Our piece on handling a denied AI accusation covers the messy classroom case where the tools disagree.
The Bundled-Tool Trend
Vendors have noticed the confusion, and the market has moved toward bundling both engines in one interface. Turnitin added AI detection to its existing plagiarism product in 2023. Copyleaks runs both engines and returns two separate scores. Proofademic launched with a sentence-level AI detector and now includes a plagiarism checker on its paid plans, folding both into a single submission workflow.
The bundling matters for workflow reasons more than technical reasons. Instructors do not want to run a paper through two separate systems, reconcile two separate reports, and remember which one flagged what. A single report that shows a plagiarism score and an AI score, each with methodology behind it, is easier to defend in an appeal and easier to explain to a student.
A caution: bundling does not mean the two engines are equally strong. Some vendors have a mature plagiarism engine and a weaker AI detector, or the reverse. The right question is what each score means and how each was validated, not whether both appear on the same page. For teachers weighing options, our 2026 detector roundup looks at which bundles hold up.
Frequently Asked Questions
Can Turnitin detect both plagiarism and AI?▼
Yes, if your institution's Turnitin license includes the AI writing detection feature. Both scores appear on the Similarity Report, but they come from different engines and mean different things.
Will a plagiarism checker catch ChatGPT?▼
Rarely. Model-generated text is novel and does not appear in any plagiarism index. A plagiarism checker only catches AI output if the student submitted something that was later indexed, or if the model happened to reproduce indexed text verbatim, which is uncommon.
Does Proofademic do both?▼
Yes. Proofademic launched as a sentence-level AI detector and now bundles a plagiarism checker on paid plans. Both scores appear together in the report.
Which score matters more in an academic integrity case?▼
It depends on the allegation. Copied text calls for plagiarism evidence. AI-generated authorship calls for detection evidence plus classroom context. The two rarely substitute for each other.
Can a student paper score zero on plagiarism and 100 percent on AI detection?▼
Yes, and this combination is common with AI-assisted submissions. The paper is novel text that matches no source, and the detector reads it as model-generated.
The Bottom Line
Plagiarism detection and AI detection answer different questions with different methods on different evidence. Treating them as interchangeable leads to blind spots on both sides.
For most teachers the practical takeaway is that you need both signals, ideally in one report, and you need to know what each score actually means before you take it to a student. Our guide on reading a Turnitin AI score covers the interpretation side, and the Proofademic academic-detection approach is one of the options worth knowing if you are procuring a bundled tool.
