Plagiarism checker vs AI detector: why schools probably need both

When schools ask about academic integrity tools, they often frame it as a choice: should we get a plagiarism checker or an AI detector? The honest answer is that these tools solve different problems, detect different forms of academic dishonesty, and are not good substitutes for each other. For most schools, the right approach isn't to choose between them.

What plagiarism is - and what a checker detects

Plagiarism, in the traditional academic sense, is the use of someone else's work without appropriate attribution. The most common forms involve copying text from a source - a textbook, a website, another student's work - with or without minor modifications. Plagiarism checkers detect this by comparing submitted text against a reference corpus and flagging passages that match existing content above a similarity threshold.

Good plagiarism checkers have substantial and well-maintained reference corpora covering academic publications, web content, and student submission databases. They're effective at catching direct copying, close paraphrasing, and recycled submissions. Their limitation is the corpus: they can only find matches for content already indexed. If the matching content isn't in the database, the similarity won't be found.

What AI generation actually is

AI generation is a different kind of academic dishonesty. When a student uses a language model to generate their essay, that essay doesn't copy from any existing source - it's constructed word by word by a probabilistic model. The output is genuinely novel text that has never existed before. There is no matching source for a plagiarism checker to find.

This is why running an AI-generated essay through a plagiarism checker typically produces a low similarity score. The tool will report that the essay is mostly original - which, in a narrow sense, it is. The problem isn't that it was copied; the problem is that the student didn't write it. These are different problems with different evidentiary requirements.

Where each tool fails the other's test

A plagiarism checker will miss AI-generated content in almost all cases. An AI detector won't catch content copied from existing sources: if a student pastes three paragraphs from a textbook, a text-based AI detector will likely score this as human-written, because it is. A process-based tool might flag the paste event, but it won't identify where the text came from.

The middle ground - AI-assisted paraphrasing of an existing source - can defeat both. A student who takes copied text and runs it through an AI paraphrase tool produces output that isn't similar to the original (defeating the plagiarism checker) and may not read as AI-generated (defeating text-based AI detection). This increasingly common technique is one of the reasons a layered approach to academic integrity is preferable to relying on any single tool.

The case for using both

Used together, plagiarism checking and AI detection cover substantially more ground than either does alone. Plagiarism checking addresses copying from sources; AI detection (particularly process-based) addresses generation shortcuts and paste-heavy workflows. The two tool types have different false positive profiles, which means using both tends to produce more targeted flags than using either in isolation.

The combination is also more defensible in formal proceedings. A submission that shows both low plagiarism scores and anomalous process signals - a very short session, a large paste - presents a more complete picture than either signal alone. Conversely, a submission with high similarity scores but a completely normal process record might be a citation error rather than intentional misconduct.

Process detection: covering both problems at once

Behavioural detection has an interesting property in relation to both forms of academic dishonesty. It doesn't distinguish between content pasted from a website and content pasted from ChatGPT - it observes that a large paste occurred and records the fact. A process-based tool can therefore flag both traditional plagiarism (paste from a source) and AI submission (paste from a generator) from the same underlying event type.

This doesn't replace dedicated plagiarism checking for situations where source identification matters - knowing that 300 words were pasted from a specific textbook chapter is more useful for assessing the severity of misconduct than knowing only that a paste occurred. But for initial triage, identifying which submissions deserve a closer look, process signals cover both problems simultaneously.

What a sensible layered approach looks like

A practical academic integrity approach for most schools combines three elements: a plagiarism checker for submitted documents to catch direct copying and close paraphrasing; a process-based AI detection tool integrated into the homework collection workflow to capture behavioural signals during submission; and a clear teacher process for using the outputs of both as starting points for conversation rather than automatic judgements.

The third element - the teacher process - is as important as the tools. The most defensible academic integrity cases are ones where the technology provided supporting evidence for a conversation, and the conversation produced clarity. Tools that bypass conversation entirely create the most legal and relational exposure. The best use of academic integrity software is to make human judgement better informed, not to replace it.