The corroboration standard rests on a simple principle: no academic-integrity finding should rest on a single tool's output. Process evidence, i.e. the record of how a text actually came into being, is the strongest signal we have. When that record is rich, it can show convincingly whether a passage grew through an authentic writing process or was pasted in wholesale.

But process data is not always available. Sometimes a student writes offline; sometimes the available signals are simply thin for technical reasons. In those cases our engine does something deliberately honest: it highlights the passages that look copied or externally generated, and it tells you they are indications, not proof. A finding is not a verdict. Which leaves a real question on the table: a passage is flagged, the signal is suggestive but not conclusive. What now?

The gold standard that doesn't scale

The answer the academic world has trusted for generations is the oral defense. Sit down with the student, ask them to explain their own work, and authorship becomes self-evident within minutes. It is the one form of evidence that generative AI cannot fake on a student's behalf. While you can outsource the writing, you can not outsource the understanding.

The trouble is purely practical. A full viva for every questionable submission is expensive, slow, and hard to keep consistent. Few institutions can run them at that volume, so in practice the flagged passage usually just sits there, unresolved. The best instrument we have is the one we can almost never afford to use.

Five questions, not fifty

Here is what changes the economics. You don't need to re-examine the whole paper. Our engine already pinpoints which passages are suspicious, and why. That means the oral check no longer has to be a broad, open-ended interrogation. It can be narrowed to exactly the sentences/paragraphs in question. Instead of fifty questions about a thesis, you ask five about the three paragraphs that matter. The student's burden shrinks, the examiner's burden shrinks, and the conversation goes straight to the point.

This is where our exploration with Screeners AI becomes relevant. Screeners AI is a conversational assessment platform that conducts two-way, real-time, AI-led oral assessments. It’s human-like AI interviewer can bring out tailored questions, ask follow-ups when responses need more clarity, support multiple languages, and deliver structured evaluation records. Applied to academic integrity, this enables a focused oral check where students explain the specific passages flagged by our engine, in their own time, while examiners receive a clean, structured report for review.

Fair by design — and a human still decides

The combination is more than convenient; it is fairer. Every student gets the same focused, structured opportunity to demonstrate authorship, rather than an ad-hoc challenge that depends on who happens to be in the room that day. A student who wrote their work clears it in minutes. A student who cannot account for "their own" sentences reveals that just as quickly. And noone has to make an accusation first.

And there is one line we will not cross: the machine does not deliver the verdict. Our engine highlights the anomalies in passages. Screener AI's technology structures and scales the conversation about them. The examiner (a person) weighs the answers and decides. That is the corroboration standard working exactly as intended: process evidence and conversational evidence side by side, with judgment left to a human. Targeted questioning is the last resort in that stack, not because it is weak, but because you only reach for it when the lighter signals leave a genuine question open.

Generative AI made it simpler to disguise a text. It did not make it easy to fake an understanding of one. The oral defense has always known this. What has been missing is a way to make that defense small enough, fast enough, and fair enough to use at scale — and pointing it only at the passages that matter is what gets us there.

---

Further reading on Mentafy

• After Adelphi: Why "AI Detection Alone" Is No Longer a Legally Defensible Standard — where the corroboration standard comes from.
• Para-Plagiarism: Why the Most Expensive Plagiarism Software Misses It — the process- and source-evidence layer that flags the passages in the first place.