Stop detecting AI. Start assessing what it can't fake.
Detection is an arms race educators are losing — with honest students as collateral damage. AI-resistant assessment takes the other path: verify understanding in live conversation, where no chatbot can stand in.
Why the arms race can't be won
With roughly 90% of students using AI on coursework, the instinctive institutional response has been detection: run submissions through a classifier, flag the suspicious ones, prosecute the flagged. Every part of that pipeline is broken.
False positives punish honest students
Detectors flag human writing — disproportionately that of non-native English speakers — and a single false accusation can derail a student's academic career. No accuracy rate is high enough when the penalty is an integrity violation.
Evasion is trivial
Paraphrasing tools, "humanizer" services, and light manual editing defeat detection reliably. Each detector improvement is answered within weeks. The students most willing to cheat are the least likely to be caught.
Surveillance poisons the classroom
Treating every submission as a suspect exhibit turns instructors into prosecutors and students into defendants. That relationship damage persists long after any individual case.
The alternative isn't a better detector. It's assessment that doesn't depend on knowing who — or what — wrote the text.
Three properties that make assessment resistant by design
No gap to fill with AI
Spoken conversation moves at speaking pace. There is no window to paste a question into a chatbot and read back the answer without the delay — and the secondhand phrasing — being obvious.
Nothing to pre-generate
Every question depends on the student's previous answer. There is no question bank to leak, no essay prompt to feed a model in advance, no script that survives contact with a follow-up.
Their submission, their defense
When the exam probes the student's own submission — why this argument, why this method, what breaks if we change this — outsourced work exposes itself. You can't defend reasoning you never did.
AI-resistant doesn't mean anti-AI
The graduates your programs produce will use AI daily in their careers. Banning it from coursework prepares them for a world that no longer exists — and is unenforceable anyway. The question worth answering isn't "did this student use AI?" but "does this student understand the material?"
Verification by conversation decouples those questions. Students can use modern tools in their process; the grade rests on what they can explain, defend, and extend in a live exchange. That standard is fair, transparent, announced in advance — and it changes behavior upstream, because studying for a conversation means actually learning.
"You can't defend reasoning you never did. That single fact does more for academic integrity than every detector on the market."
Frequently Asked Questions
What makes an assessment AI-resistant?
An assessment is AI-resistant when using AI to complete it either doesn't help or is immediately visible. Real-time oral examination qualifies on both counts: questions adapt to each answer so there is nothing to pre-generate, and the conversation moves at speaking pace, leaving no practical way to relay questions through a chatbot without obvious gaps and secondhand answers.
Why not just use AI detection tools?
Detection tools have documented false-positive problems — flagging honest students, disproportionately non-native English speakers — and are trivially defeated by paraphrasing tools or light editing. They also create an adversarial dynamic between instructors and students. Assessment design avoids the arms race entirely: instead of trying to prove a text was AI-written, you verify understanding in a format AI can't perform on the student's behalf.
Does AI-resistant mean banning AI from coursework?
No — often the opposite. Once verification happens in conversation, students can be free to use AI in their process, the way professionals do. What Prova verifies is that learning happened along the way: the student can explain, defend, and extend the work they submitted, whatever tools were involved in producing it.
Can't a student use AI to prepare for the oral exam?
Yes — and that's the point. Preparing for an adaptive oral exam means learning the material well enough to discuss it. If a student uses AI as a study partner and then demonstrates genuine understanding in conversation, the assessment worked. What they can't do is have AI attend the conversation for them.
How do we roll this out without redesigning courses?
Think of Prova as a plugin for your existing course rather than a replacement for it. Keep your assignments and add a short oral verification layer to the ones that matter: students submit as usual, then complete a 10–15 minute Prova session about their work. Grading weight can shift gradually as your confidence in the format grows — and the final grade always remains the instructor's call.