Connecticut Mandates AI Instruction While Florida and New York Move to Restrict It

Framing

For two years, the governing question in K-12 artificial intelligence has been who owns it. Last week's brief documented Maryland's answer to that question by statute, with a mandated district AI coordinator. This week, Connecticut moved the frontier again, past ownership and into instruction. Public Act 26-15, signed by Governor Ned Lamont on June 2 and developed jointly with the Attorney General and legislative leaders, adds computer science, explicitly including AI and emerging technologies, to required public-school instruction beginning in 2026-27. It establishes a Connecticut AI Academy to train students and teachers, expands teacher-certification programs to include emerging-technology instruction, and, in the same statute, imposes social-media age verification and a one-hour default feed limit for minors, and requires AI chatbot operators to detect and respond to expressions of self-harm. The structural shift is that the state is no longer only governing the AI that enters its schools. It requires that AI be taught.

The same week, it was revealed that the mandate now has two opposite faces. Florida's proposed AI Bill of Rights would have prohibited AI instructional tools before sixth grade; it passed the Florida Senate 37-1 before dying in the House. In New York, a sitting assemblymember has introduced legislation to keep most AI out of K-8 classrooms, with carve-outs only for diagnostic testing and disability support. Within a single national news cycle, one set of states is moving to require AI instruction, and another is moving to wall young children off from it. Both are governance acts. Neither rests on settled evidence about how AI affects a young learner's developing skills.

The research that surfaced this week maps almost exactly onto that uncertainty. A peer-reviewed meta-analysis found that AI agents shift skills and knowledge in K-12 classrooms, but not, at conventional significance levels, higher-order thinking. A peer-reviewed study in npj Artificial Intelligence and an arXiv preprint independently documented a grade-appropriateness gap, in which off-the-shelf models fail to adapt to younger learners and degrade on simpler material. A peer-reviewed scoping review found that the benefits of adaptive learning skew toward better-resourced learners. Read together, the pattern is uncomfortable for both camps: AI helps most where the task is concrete and the learner is older and well-resourced, and the evidence is thinnest exactly where the policy fight is loudest, which is younger children, higher-order thinking, and under-resourced settings.

For a cabinet, the implication is that a curriculum mandate is now itself a governance instrument, and it can take either polarity: a requirement to teach or a prohibition to wait. Either way, the districts that have already designed a developmentally staged, evidence-based, equity-checked AI scope and sequence will be able to implement the mandate without disruption. The district that has not will be retrofitting under a statutory clock. Connecticut and Florida are not really opposites. They are two timers running on the same unsolved problem.

Top Research and Policy Signals

1. Connecticut Enacts Public Act 26-15, Making AI-Inclusive Computer Science Required Instruction and Pairing It With Minor-Safety Rules

Source type. State legislation (enacted; Public Act 26-15, signed June 2, 2026).

Connecticut's new law is a sweeping, bipartisan statute that reaches consumer protection, employment, and the classroom. On the education side, it amends state law to add computer science, including AI and emerging technologies, to required public-school instruction starting in the 2026-27 school year. The state Department of Education has been careful to note what the law does not do: it is not a graduation requirement and does not compel a standalone AI course. The intent, in the words of the Connecticut Association of Boards of Education, is that students will encounter these concepts at one or more points in their academic careers, with districts retaining local flexibility on how. The act also creates a Connecticut AI Academy, which the Board of Regents must establish by the end of 2026, to deliver AI courses and responsible-use training to students aged 13 to 20 and to develop teacher training in consultation with educator unions.

The same statute pairs instruction with restriction. It requires social-media companies to verify age and obtain parental consent before minors can access addictive algorithmic feeds, bars notifications to minors between 9 p.m. and 8 a.m., sets a one-hour default daily limit on algorithmic feeds, and requires AI chatbot operators to make reasonable efforts to detect suicidal ideation and respond with resources. Connecticut has, in one law, told schools to teach AI and told AI products to behave around children.

Leadership implication. This is the first widely reported instance of a state writing AI into a mandated curriculum and a minor-safety law in a single act, and it sets the template that other legislatures will copy. Superintendents should not wait for their own statute to read PA 26-15 as a planning signal: build a developmentally staged AI and computer-science scope and sequence now, decide deliberately where AI concepts attach in the existing curriculum rather than bolting on a course, and align student-device and chatbot-use rules to the same minor-safety logic the law encodes. A mandate designed locally is a capability; a mandate absorbed reactively is a compliance scramble.

2. Peer-Reviewed Meta-Analysis: AI Agents Raise K-12 Skills and Knowledge, but the Higher-Order-Thinking Effect Did Not Reach Significance

Source type. Peer-reviewed, published in Computers in Human Behavior Reports (2026). Meta-analysis of 73 effect sizes from 34 studies (2020-2025).

This random-effects meta-analysis synthesized 73 effect sizes from 34 studies of AI agents in K-12 classrooms published between 2020 and 2025. It found a statistically significant, moderate overall effect on students' cognitive learning outcomes (Hedges g = 0.404, p < 0.001). The effect was significant for skills-based outcomes (g = 0.391, p < 0.001) and for knowledge-based outcomes (g = 0.344, p < 0.05). For higher-order thinking, the point estimate was actually larger (g = 0.540), but it did not reach statistical significance (p = 0.066), a result the authors attribute in part to fewer studies and greater variability in that category. Effects were moderated by learner level and discipline, indicating that grade band and subject change how much AI agents help.

Leadership implication. The single most important number here is the one that did not clear the bar. Districts routinely justify AI instruction by appealing to critical thinking and higher-order reasoning, yet this is the strongest current K-12 synthesis, and it cannot confirm that effect at conventional significance levels. Leaders should keep AI where the evidence is solid, which is concrete skill and knowledge practice, and treat higher-order-thinking claims as a hypothesis to be tested locally, not a settled benefit to be marketed. Ask any vendor citing critical-thinking gains to show the grade band, the discipline, and the significance level behind the claim.

3. Peer-Reviewed (npj Artificial Intelligence): Off-the-Shelf LLMs Are Not Grade-Appropriate by Default, and the Gap Is Widest for Younger Learners

Source type. Peer-reviewed, published in npj Artificial Intelligence (2026). Framework development and human evaluation with 208 participants across six grade levels.

The authors show that large language models fail to provide grade-appropriate responses for students at different educational levels, then build a framework that finetunes models to generate age-appropriate content across six bands, from lower elementary to adult education, by integrating seven readability metrics. Evaluated across multiple datasets with 208 human participants, the finetuned approach improved grade-level alignment by 35.64 percentage points over prompt-based methods while preserving factual accuracy. The practical reading is that an unmodified consumer model, the kind students actually reach for, defaults toward explanations pitched above the youngest learners, and closing that gap took deliberate engineering, not a better prompt.

Leadership implication. This converts grade-appropriateness from a soft preference into a procurement specification. Any AI tool a district buys for elementary or middle grades should be required to demonstrate, with evidence, that its outputs are calibrated to the developmental level of the students who will use it, not merely that it is factually correct. The finding also tempers the Connecticut model and sharpens the Florida and New York caution: if the default tools are weakest for young children, then where AI enters the early grades, it has to be configured for them on purpose, with teacher mediation assumed rather than optional.

4. Peer-Reviewed Scoping Review: Adaptive-Learning and Analytics Benefits in K-12 Skew Toward Better-Resourced Learners

Source type. Peer-reviewed, published in The International Review of Research in Open and Distributed Learning (2026). PRISMA-ScR scoping review of 21 empirical studies.

This scoping review analyzed 21 empirical studies of AI-driven adaptive learning and learning analytics in K-12 online and distance settings from 2020 to 2025. Most studies reported gains in engagement, motivation, and self-regulation. Two findings matter for governance. First, the reported benefits were unevenly distributed and often favored better-resourced learners, particularly where teacher mediation and institutional support were modest. Second, the review found that definitional inconsistency blurred the line between genuine intelligence and simple automated adaptation, so that products marketed as adaptive AI were not always doing what the label implied. The authors conclude that the value of these tools depends less on technical sophistication than on equitable, pedagogically informed integration.

Leadership implication. Adaptive technology tends to widen the gaps a district most wants to close, unless teacher support and institutional scaffolding are funded alongside the tool. Equity cannot be assumed from the product; it has to be engineered into the rollout. Two procurement actions follow: require vendors to define precisely what their system adapts and on what data, so the label is verified rather than trusted, and budget for the teacher mediation the evidence says is the actual mechanism of benefit, rather than purchasing the platform alone and hoping for equitable results.

5. Preprint: Frontier Models Show an Expert's Curse and a Foundational Fallacy, Degrading on Simpler, Lower-Grade Material

Source type. Preprint (not peer-reviewed), posted to arXiv (2026). Evaluation of four LLMs against a curriculum-aligned benchmark; non-U.S. context (Nepal, Grades 5-10). Transfer to U.S. K-12 is inferential.

This preprint evaluated four state-of-the-art models, including GPT-4o and Claude Sonnet 4, as AI tutors against Nepal's Grade 5 to 10 science and mathematics curriculum, scoring them on seven pedagogical metrics. Frontier models achieved roughly 97 percent aggregate reliability, but the authors identified two recurring failure modes that have a direct bearing on young learners. The Expert's Curse describes models that solve hard problems yet cannot explain them clearly to novices. The Foundational Fallacy describes performance that paradoxically degrades on simpler, lower-grade material because the models struggle to adapt to younger learners' cognitive constraints. The authors conclude that off-the-shelf models are not ready for autonomous classroom deployment and should be used only as assistants to human teachers who vet and adapt the content. As a preprint in a non-U.S. context, this is a signal to watch, not a settled finding, but its core observation about model behavior is consistent with the peer-reviewed npj result above.

Leadership implication. When two independent research efforts, one peer-reviewed and one a preprint, find the same weakness with the youngest and most foundational material, district leaders should treat human-in-the-loop as the default operating assumption for AI in the early grades, not as an enhancement. Procurement language and classroom protocols should both presume teacher vetting of AI explanations for younger students. This is precisely the design discipline that a Connecticut-style teach-it mandate will require and that a Florida-style restriction is implicitly worried about.

Emerging Strategic Themes

Theme 1. The Mandate Now Points Two Ways. State policy has split into opposite instruments operating in the same season: Connecticut requires AI-inclusive instruction while Florida and New York move to restrict AI for young children. The strategic point for districts is that both polarities impose a curriculum decision, and a scope and sequence designed to satisfy a teach-it mandate is largely the same document needed to satisfy a restrict-it regime. Building it now hedges against whichever timer reaches your state first.

Theme 2. Grade-Appropriateness Becomes a Procurement Specification. Two independent 2026 studies show that off-the-shelf models are not calibrated to younger learners by default. That moves developmental fit out of the realm of trust and into the realm of evidence that a vendor must produce. Districts buying AI for elementary and middle grades should require demonstrated grade-band calibration, not just factual accuracy, and should assume teacher mediation as a design condition.

Theme 3. The Higher-Order-Thinking Dividend Is Still Unproven. The strongest current K-12 meta-analysis confirms gains in skills and knowledge but cannot, at conventional significance levels, confirm the higher-order-thinking benefit that curricula most often invoke. Leaders should anchor AI instruction to the outcomes supported by the evidence and treat critical-thinking claims as locally testable hypotheses, thereby protecting the district when a board or a skeptic asks for proof.

Theme 4. Adaptive Tools Default Toward the Already-Advantaged. Adaptive-learning benefits in the peer-reviewed record skew toward better-resourced learners and depend on teacher mediation that under-resourced settings often lack. Without deliberate equity engineering and funded teacher support, an adaptive platform can widen the very gaps a district adopts it to close. Equity is a property of the rollout, not the software.

What Was Not Found

Five evidence categories did not appear in this week's window, and each absence carries a present-tense cost while states legislate and districts plan.

First, no study, peer-reviewed or preprint, has tested whether a mandated AI-inclusive curriculum, the mechanism Connecticut just enacted, improves any student outcome, AI-literacy measure, or workforce-readiness result. States are now requiring AI instruction with no evidence that the requirement produces what it promises.

Second, no peer-reviewed study establishes the correct age floor for AI instructional tools. This is the exact question Florida legislated at sixth grade, New York is legislating at K-8, and Connecticut is answering in the opposite direction by mandating exposure. Three states are setting three different developmental thresholds, and none can cite causal evidence for the line it drew.

Third, the higher-order-thinking effect did not reach significance in the strongest available K-12 meta-analysis. The cognitive capacity most often used to justify teaching AI is therefore the one without demonstrated causal support in the K-12 record, even as curricula are being mandated around it.

Fourth, the developmental evidence that surfaced concerned model behavior, not student learning. The npj study and the arXiv preprint show that off-the-shelf models are miscalibrated for young learners, but neither demonstrates that grade-aligned AI content actually improves outcomes for young children. Districts configuring tools for the early grades are acting on a plausible mechanism rather than a proven result.

Fifth, no peer-reviewed K-12 study this week produced equity-stratified causal outcomes for English learners, students with Individualized Education Programs, or students in high-poverty schools. The scoping review shows adaptive benefits skewing toward better-resourced learners, but the causal subgroup evidence needed for a district to protect its most vulnerable students is still absent, even as procurement decisions are being made now.

Novo Executive Summary

This was the week the AI mandate split in two. Connecticut wrote AI instruction into the required curriculum and paired it with minor-safety rules; Florida and New York moved in the opposite direction to keep AI away from young children; and the research that surfaced says the tools help older, well-resourced learners on concrete skills, while the developmental case and the higher-order-thinking case both remain thin. The governing variable is no longer whether to govern AI. It is whether a district's curriculum architecture is staged by development, bounded by the evidence that actually exists, and checked for equity before a state imposes either timer. A scope and sequence that can satisfy a teach-it mandate and a restrict-it regime is, in practice, the same document, and the districts that build it now will not be retrofitting it in August. Novo Innovative Pathways works with district cabinets to design the architecture: a developmentally staged, role-based AI literacy scope and sequence; a grade-appropriateness standard that AI tools must meet before procurement; and a governance model that turns a curriculum mandate, in either polarity, into defensible institutional practice.

Watch This Week

• Ohio's statutory deadline for every district, community school, and STEM school to adopt an AI use policy is July 1, 2026, with a final wave of board ratifications expected through June.
• Connecticut implementation of Public Act 26-15: the Board of Regents standup of the AI Academy by the end of 2026, State Department of Education guidance, and teacher-certification program changes that may begin in July.
• New York Assembly bill A9190, which would restrict most AI in K-8 classrooms, and any refiling of Florida's AI Bill of Rights after its House defeat.
• New York City Public Schools' comprehensive AI playbook, expected in 2026, with differentiated guidance for K-5, 6-8, and 9-12 grade bands.
• The 2026 state legislative pipeline tracked by MultiState and FutureEd, with more than 130 AI-in-education bills filed across more than 30 states.
• Replication of the grade-appropriateness finetuning result against U.S. curricula, which would convert this week's developmental signal into actionable procurement evidence.

Sources

Governance and Policy

Research, Peer-Reviewed

Research, Preprint (Not Peer-Reviewed)