AI detector vs plagiarism checker: what's the actual difference?

People regularly use the terms 'AI detector' and 'plagiarism checker' as though they're two versions of the same thing. They're not. They detect different problems, use different methods, and have different failure modes. Understanding the distinction helps you choose the right tool - and understand why you might need both.

What a plagiarism checker actually does

Plagiarism checkers work by comparing submitted text against a reference corpus - a large database of previously published content, including web pages, academic papers, and often a repository of previously submitted student work. When significant textual similarity is found, the tool flags the matched passages alongside the source. The key output is a similarity score: what percentage of the submission matches content already in the database.

This approach is well-suited to catching direct copying and close paraphrasing from existing sources. It works reliably when students lift text from textbooks, Wikipedia, published papers, or recycled assignments. The limitation is the reference corpus: the tool can only find matches for content that's already indexed. Content that doesn't exist in the database won't be detected, regardless of how it was produced.

What an AI detector actually does

Text-based AI detectors take a different approach. Rather than comparing text against a corpus, they analyse its internal statistical properties - primarily how predictable each word choice is relative to what a language model would generate. AI-generated text tends toward higher predictability; human writing tends toward more variation. The output is a score reflecting how 'AI-like' the text reads, not how similar it is to any specific source.

Process-based AI detection - a separate and more reliable approach - records the writing behaviour that produced the text: timing patterns, paste events, session duration. This doesn't analyse the text at all; it analyses the actions of the person writing it.

The key distinction: source versus authorship

Plagiarism is fundamentally about source: did this student copy content from somewhere it shouldn't have come from? AI generation is fundamentally about authorship: did this student actually write this themselves? These are related concerns, but they're not the same question, and the same evidence doesn't answer both.

Here's the critical implication: AI-generated content is 'original' by definition. When a student asks ChatGPT to write their essay, the result doesn't copy from any database - it generates novel text. A plagiarism checker will not flag a fully AI-generated essay, because there is no matching source to find. The essay is technically original content. The problem isn't that it was copied; the problem is that the student didn't write it.

Where they overlap - and where they don't

The two categories overlap in one area: contract cheating services that recycle previously submitted essays or copy from essay mills. A plagiarism checker will catch some of these; an AI detector won't. But there's very little overlap in their core detection targets, which is why treating them as alternatives is a mistake.

Paraphrasing tools sit in an interesting middle ground. AI-assisted rewriting of an existing source - taking copied content and running it through a paraphraser - may defeat both a plagiarism checker (the text is now substantially different from the source) and a text-based AI detector (the output may not read as AI-generated). Process-based detection is more likely to catch this, because the paraphrasing activity typically involves pasting content and making minimal revision.

What most schools actually need in 2026

For most school contexts today, the more pressing concern is AI generation rather than traditional plagiarism. Students who would previously have copied from Wikipedia now have access to tools that generate apparently original text on demand - and this passes plagiarism checks without triggering any flag. A school running only a plagiarism checker is effectively detecting yesterday's problem whilst the current one goes unnoticed.

That said, traditional plagiarism hasn't disappeared. Contract cheating, copying from online sources, and recycling of previously submitted work are all still common. The right answer for most schools is not to choose one tool but to have a layered approach: plagiarism checking catches source-copying; AI detection (preferably process-based) catches generation shortcuts.

The process layer that covers both

Behavioural detection has an interesting property in relation to both forms of academic dishonesty: it flags both. A student who copies content from a source and pastes it into an assignment produces the same kind of process anomaly - a large paste event - as a student who copies AI-generated content and pastes it in. The paste doesn't carry a source label; it's just a paste.

This doesn't mean process detection replaces dedicated plagiarism checking. Plagiarism checkers can identify the specific source that was copied from, which may be relevant for assessing severity and for formal proceedings. But for initial triage - identifying which submissions are worth examining more closely - process signals catch both problems at once, regardless of origin.

Questions worth asking when choosing tools

When evaluating AI detection and plagiarism tools, the questions worth asking are quite different for each. For plagiarism checkers: how large is the reference corpus? Does it include student submission repositories? How quickly is new online content indexed? For AI detectors: what is the false positive rate for non-native English writers? Does it use text analysis, process analysis, or both? What data protection guarantees apply to submitted content?

A school that has both a reliable plagiarism checker and a process-based AI detection tool is in a considerably stronger position than one relying on any single approach. The two tools cover different ground, and together they provide much more complete coverage of the academic integrity risk landscape than either does alone.