Why AI Broke Plagiarism Detection (and What Comes Next)

For roughly two decades, the industry's answer to academic dishonesty was the similarity score. Run a submission through a text-matching tool, see how much of it overlaps with something already indexed, and treat a clean report as evidence the student did the work. It was never a perfect system, but it was serviceable, until generative AI removed the exact signal it was built to read.

AI-written work is original by every measure a similarity checker can apply. There's nothing to match against, because nothing was copied. A student can submit a fully AI-generated essay and walk away with a pristine originality report, precisely because the tool was never designed to ask the question that now actually matters: did you do this yourself?

The burden shifted to faculty

When detection stopped working, the burden didn't disappear. It moved onto lecturers, who are now left making judgment calls with no real evidence behind them. Suspicion without proof is bad for everyone: it puts honest students under a cloud they can't clear, and it puts faculty in the position of accusing someone based on a feeling rather than a record.

That's the actual problem worth solving, and it isn't a better detector. Every generation of AI-detection tooling gets outpaced by the next generation of language models within months. Chasing that arms race indefinitely isn't a strategy.

Changing the question, not the arms race

We built Smart Defense around a different premise: instead of trying to determine who or what wrote a submission, ask the student to explain it. A short spoken or written defense, in the student's own words, on the same platform as the exam, verifies understanding directly, rather than inferring authorship indirectly.

It's a small addition to an assessment: a lecturer enables a defense step on an assignment or exam, the student responds when they submit, and the response arrives alongside the submission for review. No new grading pass, no rewritten rubric, no separate detection vendor to argue with.

Evidence, not accusation

The result isn't a probability score that faculty then have to interpret and defend. It's a timestamped record: the student's own explanation, tied to a verified identity, that a lecturer or an academic integrity committee can review directly. That's a real difference in an appeal or a hearing. A similarity percentage is an opinion about a document; a recorded explanation is evidence about a person.

AI didn't make academic integrity impossible. It made the old proxy for it obsolete. The institutions adjusting fastest aren't the ones buying a sharper detector. They're the ones changing what they actually verify.