Ten AI-in-Schools Laws Passed This Year. Only One Contains the Sentence That Matters.
Oklahoma wrote it into statute: AI may not be the primary basis for a grade, a placement or a suspension. Everywhere else, that sentence is still a draft.
A teacher I have known for two decades sent me her district's new AI policy in the spring and asked — in the tone people use when they already suspect the answer — whether it was any good.
I read it on my phone in a parking lot, in about four minutes, and told her it looked sensible. It used the phrase "human oversight" twice, said the right things about equity and integrity, and had clearly been written by people who care.
Here is the assumption I never checked, and I read documents for a living: I assumed a policy tells somebody what they may not do. That is what the word means when a bank writes one, or a hospital. So I read her district's document for tone, and never thought to read it for prohibitions.
When I finally did — hunting only for sentences that forbade something — I found none. Not one line said what the software was not allowed to decide. And once you see that shape you cannot stop seeing it: most of what is published about AI in schools has exactly the same shape.
Then I found one that doesn't.
What actually passed — and what "ten" is counting
Start with the arithmetic, because the number is the part everybody gets wrong.
FutureEd, the Georgetown-based tracker, is following 77 bills across 27 states that address artificial intelligence in classroom instruction. As of its July 13, 2026 update it reports ten bills enacted this session, naming six states among them: Alabama, Idaho, Maryland, Oklahoma, Utah and Virginia. Note the word: including. Six is illustrative, not a roll call. ExcelinEd's May 26, 2026 scorecard counted seven states with enacted laws and warned that "important gaps remain."
Underneath sits a bigger churn: Stateline's June 2026 survey found lawmakers filed more than 134 bills across 31 states this year, and MultiState reported in April that limiting high-stakes AI use is a recurring theme.
A theme is not a rule. Most of what passed tells the state education department to publish guidance and school boards to adopt a matching policy. K-12 Dive described that pattern across Idaho, Maryland, Oklahoma and Virginia in July 2026 — then added the line that sent me looking: at least one state has gone as far as to prohibit AI's use for grading, discipline or other high-stakes decisions.
One. Out of ten.
The sentence itself
It is in Oklahoma Senate Bill 1734, the Oklahoma Responsible Technology in Schools Act. From the enrolled text, at subsection E(3):
Artificial intelligence tools shall not be used as the primary basis for student grading, discipline, placement, promotion, retention, or other high-stakes educational decisions
Twenty-three words, in the grammatical mood policy documents almost never reach for: shall not.
It is narrower than the version people repeat. It does not ban AI from the classroom, or even from touching a grade. It bars AI from being the primary basis for one: a machine may inform the decision; a person must own it.
The statute does not leave that idea floating. The same subsection requires classroom AI tools to be educator-directed, says responsibility for AI-informed decisions "shall not be delegated to an artificial intelligence system or vendor," and requires outputs to be reviewed by an educator before they reach instruction, feedback, assessment or a decision. Quietly the most important part: it defines "human-in-the-loop" as an educator keeping "oversight, review, and final decision-making authority" over the tool's output. Not a glance. Not a checkbox. Authority.
Then two things a district policy generally cannot do. Parents get a written disclosure at least annually, naming every AI tool in use, its vendor and the student data collected. And a parent may opt a child out at any time — with students who opt out neither "academically penalized" nor "denied access to core instructional content."
This was not a partisan squeaker: the record shows it passed 42-0 in the Senate and 89-0 in the House, and was approved by the governor on May 12, 2026. Its author, state Sen. Ally Seifried, said the point was to set guardrails that "make sure it supplements, rather than supplants, the hard work of our classroom teachers". And Oklahoma Watch reported in June that nearly every other AI bill in the Oklahoma Legislature hit a dead end — this is the exception in a state that mostly decided not to regulate AI.
Four states told districts to write a policy. One said what the policy may not do.
Idaho is the useful comparison: good work, and still no sentence. Senate Bill 1227 passed 26-8 in the Senate and 62-6 in the House before reaching the governor in March 2026, and Idaho Education News described what it does: it instructs the state Department of Education to develop a statewide framework for integrating generative AI into classrooms, required to be "human-centered," transparent and safe. That reporting says nothing about grading, discipline or placement limits, and I will not invent one. Idaho built the scaffolding; somebody still has to hang a rule on it.
The sentence is not rare because nobody thought of it. It is nearly everywhere — in bills that have not passed.
Washington's SB 5956, still a bill in the 2025-26 session, would say an automated decision system may not be "the sole or determinative basis" for any student discipline decision. West Virginia's H.B. 5205, introduced on February 5, 2026 and referred to committee, would prohibit educators from using generative AI "to independently grade student work or issue high-stakes determinations". South Carolina's H.B. 5253, which would bar automated disciplinary or placement decisions "without meaningful human review," has sat in the House Education and Public Works Committee since February.
Even the trackers keep the tense honest — FutureEd's entries say those two would do these things. The sentence exists all over the country. In exactly one place, it exists as law.
Why one sentence is worth this much fuss
Because grading is where the machines are most confidently oversold, and the evidence says so.
A University of Cambridge-led study published on May 22, 2026 ran 761 undergraduate essays from 125 students past frontier models and found they matched human examiners' broad grading bands "between 35-65% of the time". Hold the scope on that: undergraduate essays, not school work — nobody in that study tested a fifth-grader. But the warning travels. Leaning on current models, they concluded, would produce grading that is "homogenised, underestimates brilliance, and favours linguistic style over the substance of sound academic judgement," and Dr. Deborah Talmi's bottom line was that AI alone is "far too shallow and inconsistent" to grade undergraduate work.
Discipline is worse, because the errors are not evenly distributed. The ACLU's survey of the edtech surveillance industry — older fieldwork now, published October 3, 2023 — found a quarter of students already worried the systems would be used to discipline them or their friends, with LGBTQ students, students of color, students with disabilities, low-income and undocumented students especially exposed.
Meanwhile the district layer — where my friend's document lives — is being filled by templates. The Mississippi Artificial Intelligence Network publishes one, and its authors are careful about what it is: a planning framework, "not a final policy". An honest disclaimer — and the whole problem in one line. A framework you may adapt is not something anyone can be held to.
The strongest case against everything I have just said
The argument is that it is far too early to legislate best practice, and the person making it best is Justin Reich of MIT, whose Teaching Systems Lab argues for "an ethos of humility" and for racing "to answers that are right, not first." Speaking to Stateline, Reich said states are regulating at a moment so unstable that writing a guide for AI in 2026 "is like writing a guide for aviation in 1905." Run the AI literacy and readiness documents against a checklist, he added, and "there would be a lot of boxes in the 'we're making this up' column and not a lot in the 'we have evidence' column."
There is a practitioner version in the same reporting, from the School Superintendents Association's Noelle Ellerson Ng: "I don't have a calculator policy. Why would I have an AI policy? I have a technology policy." And there is enforcement: writing in Forbes on July 17, 2026, Lance Eliot argued that state AI statutes nobody enforces become "paper laws" that "exist on paper but lack any teeth".
The sharpest objection is the threshold itself: who decides what "primary basis" means? The Brennan Center flagged that shape of loophole in April 2024 — writing about federal agency AI policy, not Oklahoma — noting that when a rule lets an institution opt out by deciding the AI was not a "principal basis", the institution tends to decide it wasn't.
All fair. Here is why the sentence still beats the framework: an aspiration cannot be violated, so it can never be argued about. A prohibition can be evaded, litigated, narrowed — and, crucially, invoked. A parent can hold up twenty-three words at a hearing; nobody has ever held up a posture.
Estonia came at this from precisely the other end
If Oklahoma wrote the rule and skipped the rollout, one small country did the reverse.
Estonia's AI Leap is a national initiative developed by the President's Digital Council and the Ministry of Education and Research, launched on September 1, 2025 to give students and teachers free access to leading AI applications — starting with 20,000 high school students in grades 10-11 and their 3,000 teachers. Education Minister Kristina Kallas said Estonia would "invest strongly in teacher training as part of the AI Leap," and the European Commission's Eurydice network confirms the initiative is led by the Ministry of Education and Research. Not a district experiment. A national one.
Euronews reported in May 2026 that the program runs on a 50-50 public-private split, aimed at "empowering teachers as the primary guides for students," and that by the AI Leap site's own account 64-90% of Estonian students were using AI tools before it began. Its March 9, 2026 progress report describes a nationwide app reaching 10th- and 11th-graders from the end of January, with teachers on ChatGPT and Gemini and more than 60% using these tools weekly. The design resists the obvious failure mode: as Siim Sikkut of the launch team told IEEE Spectrum, the tool "starts to tutor you more than give you an answer". In March 2026 President Alar Karis said there is "no way back"; Kallas added that Estonia will need more teachers, not fewer.
So where is Estonia's version of Oklahoma's sentence?
Harno, the Education and Youth Board, is discussing the possibility of using AI to grade mother tongue final language exams in the future with Tallinn University — the precise use Oklahoma's law forecloses. Read that carefully: a discussion, not a rollout. In that February 2026 report from Estonia's public broadcaster, Tallinn University's Merilin Aruvee said that in 60 percent of cases a language model stays within the range in which human markers already vary. A month later, Harno's Nele Toime said the question will be settled only after extensive testing and pilot trials, that there is no timeframe for introduction — and stated the principle: "automated grading can never become the only method of assessment. The role of the human grader will always remain central." An assurance, from an official, to a reporter. What legally stands between an Estonian student and a machine-set grade, per that February report, is that "European Union directives and general good practice" do not allow a machine to make the final decision on a person's text.
Which is not nothing: Annex III of the EU AI Act lists systems intended to "evaluate learning outcomes" among its high-risk categories, and Article 14 requires that high-risk systems be designed so they can be effectively overseen by natural persons. But look at the form — a continental design duty plus institutional good practice, applied from above. Oklahoma has nothing resembling Estonia's rollout, and one sentence in its own statute naming grading, discipline and placement out loud. Same principle, opposite architecture — and which holds up in ten years, after a change of vendor and a change of minister, is what I cannot stop turning over.
Now run the clock forward
Just imagine it is 2031, and you are in a small room with your daughter's teacher.
Nobody here will say the word "algorithm." The platform the district bought does not grade anything. It produces a readiness indicator, drawn from four years of assignments, attendance and keystroke cadence, and the recommendation for next year's math track arrived pre-selected before any human opened the file. Your daughter's teacher retains oversight, review and final decision-making authority; she exercises it about three percent of the time — not from laziness: agreeing takes one click, and disagreeing requires a written justification, a second signature, and a conversation with someone who will ask her, kindly, what she sees that the system doesn't.
Notice that nobody here broke Oklahoma's sentence. The software was never the primary basis for anything. It was merely the first thing in the room, and everything a human did afterward was, technically, a correction.
That is the failure mode I would bet on, and it is why the rest of the statute matters more than the twenty-three words that make the headline. The disclosure means you know the tool exists and who sells it. The penalty-free opt-out means "no" is a real option, not a costly one. And defining human-in-the-loop as final decision-making authority is the hook someone's lawyer hangs an argument on, when a district insists a teacher who clicks "accept" 97% of the time was exercising judgment.
Who else is uneasy, and from which direction
This is not a left-right fight.
From the market-oriented flank, the R Street Institute argued back in 2024 that states should prefer narrow, targeted frameworks to "heavy-handed AI regulatory measures" — which, read carefully, favors Oklahoma rather than opposing it: twenty-three words about high-stakes decisions is roughly as narrow as legislation gets. From civil liberties, the ACLU and the Brennan Center have spent years documenting what automated judgment does to students when nothing hard sits behind it.
And the school-technology world's response to the federal push to clear state AI laws away was strikingly cross-ideological. When Education Week surveyed the K-12 reaction, the Consortium for School Networking's Keith Krueger said preemption is appropriate "only when it establishes a clear, well-designed, widely accepted federal alternative"; the American Federation of Teachers' Randi Weingarten called the directive "outrageous and likely illegal"; ISTE+ASCD's Joseph South said any approach "should preserve schools' ability to protect students."
That fight is the live threat to all of this. The executive order signed December 11, 2025 sets out to check "the most onerous and excessive laws emerging from the States" — while carving out child safety protections and state government procurement and use of AI from what it asks Congress to preempt. Whether a school's rules on grading and discipline shelter inside those carve-outs is an open question nobody has answered credibly. Nor is the order's own legality settled: NPR reported that Florida Gov. Ron DeSantis said an executive order "doesn't/can't preempt state legislative action", that Utah Gov. Spencer Cox preferred a version without the bar on state laws, and that Public Knowledge's John Bergmayer thinks the legal theories "don't work very well."
What does this mean for you?
If you have a child in school — or pay taxes toward one, which is everybody — here is what I would do.
Read your district's AI policy for verbs. Search it for "shall not," "may not," "prohibited." If those words never appear, you are holding a statement of values, not a rule — and values are not enforceable.
Ask the question that separates the two: which decision about my child may this software not make? Ask it at a board meeting; ask for the answer in writing.
Ask who has final decision-making authority — and what it costs them to disagree. If overriding the system takes more paperwork than accepting it, the override is decorative.
Ask for the tool list. Every vendor, every category of student data, how far it travels. Oklahoma's parents get that annually by law; everyone else has to ask.
Ask whether opting out costs your child anything. If opted-out students miss instruction or lose points, it is not an opt-out.
Look up your own state's bill. In South Carolina, Washington and West Virginia a version of this sentence is already written and has not passed. Bills die quietly and revive when a handful of people email a legislator. Not a metaphor — the mechanism.
The lesson, as I see it
I spent years reading these documents wrong — judging them on whether they sounded thoughtful, when the only question that has ever mattered about a rule is whether breaking it means anything.
Ten AI-in-education laws passed this year, and most will probably do some good — guidance beats silence. But only one wrote down what the software may not decide, in language a parent can quote at a hearing without a lawyer. That is a small thing, small enough that a determined vendor and a tired teacher can route around it, and I would rather say so than oversell it.
It is also the only kind of sentence that can grow teeth. You cannot enforce a framework, appeal to a posture, or read a school board a value. But twenty-three words beginning with shall not — those you can read aloud, and everyone will know exactly what has been promised.
My vote? Copy the sentence. It has already been drafted in at least three other statehouses and passed in none of them — the hard part is done; the easy part is still outstanding.
Ask your school district for the sentence, not the statement — a document that forbids nothing has promised you nothing. Sorting out which is which is most of what the HAIA Foundation does all day.