Maryland Asks Whether Classroom AI Works, and Frederick County Asks Who Turned It On

Framing

No legislature, state board, or federal agency adopted a binding K-12 AI rule between September 20 and 26. After Florida's rule on September 16, that looked at first like a pause. It was not. This week, the standard for future rules moved in three directions at once: toward proof that AI improves learning, toward control over when AI gets switched on, and toward treating what AI remembers about a child as governed data.

The first direction came from Annapolis. Governor Moore's framework is a set of commitments, not a statute, and this brief reflects that. But one phrase in it matters more than the rest: tools must have demonstrated learning outcomes before use. Florida's rule, adopted ten days earlier, requires districts to review effectiveness evidence and says the absence of studies may not, by itself, block approval. Maryland's proposal points the other way. If it becomes law, the burden moves from the district to the product.

The second direction came from Frederick County, where the policy the governor describes met the district's operating reality. Gemini was enabled for middle and high school students before the board finished its AI policy. The district said Gemini was the one product it judged to meet Maryland's safety and privacy standards. Board members said parents were not told. Readers in Georgia should notice the parallel: DeKalb County announced a Gemini certification program for students with a September 21 release date, and this brief has found no board action or public data terms. In August, Google extended Gemini in Classroom to K-12 students of all ages in districts that had already granted student access to the Gemini app. The line between buying AI and turning it on is gone.

The third direction is the least visible and may matter most. A state CIO said on the record that AI memory is a new data category. Student privacy law was written for records a district creates. It was not written for inferences a model forms about a child and carries into the next session. The week's research makes the connection concrete: if a tutor behaves differently for a disengaged student, the system has, in effect, classified that student. Where that classification lives, who can see it, and whether it follows the student are governance questions no instrument enacted this season answers.

Top Research and Policy Signals

1. Maryland's Governor Proposed a "Do No Harm" Standard: Classroom AI Must Show Learning Outcomes, Supplement Teachers, Be Disclosed to Families, and Come With a Crisis-Response Protocol

Source type. State executive policy framework, Maryland. Administration commitments requiring agency and legislative action. Not law. Not research.

Office of Governor Wes Moore. (2026, September 22). Governor Moore outlines AI framework to protect Marylanders [Press release]. governor.maryland.gov Office of Governor Wes Moore. (2026). Protecting Marylanders in the age of artificial intelligence. governor.maryland.gov

The framework commits the Moore-Miller administration to actions across sectors, several aimed at young people and schools. It pledges to protect young people from unsafe chatbots and to curb AI-driven addictive design for users under 16. For K-12, it commits the administration to work with the Maryland State Department of Education and the legislature to extend the state's crisis-response safety net into the classroom by requiring schools to implement a protocol for any AI tool students interact with at school; to implement a "do no harm" policy for AI use in schools, under which AI tools have demonstrated learning outcomes, supplement educator instruction, and are disclosed to families before use; and to have the department work with districts to issue model K-12 lesson plans that help students understand AI, make safe and ethical choices, and prepare for an AI-shaped world. The framework builds on the AI Ready Schools Act, SB 720, which this brief reported in Edition 19 and which requires each Maryland district to adopt an AI policy and designate an AI coordinator.

Study context. A policy framework. It does not define demonstrated learning outcomes, name an evidence standard, set a date, or state whether the requirements would reach tools already in use. None of the K-12 items is in effect. Statutory elements would require action in Maryland's 2027 legislative session, which this brief will track.

Leadership implication. Put Maryland's proposal beside Florida's rule and give your board the choice they present. Florida: review the evidence, then approve without it if none exists. Maryland, as proposed: no demonstrated outcomes, no classroom use. Most districts are operating closer to Florida by default, without having decided to. Direct your chief academic officer to sort every student-facing AI tool into two columns, safety and privacy evidence in one and learning-outcome evidence in the other, because a product can be fully compliant and still show no evidence that it improves learning. Then assign an owner to the crisis-response question now: if a student discloses self-harm or a threat to an AI tool during the school day, who is notified, by what path, and within what time? Maryland is proposing to require that answer. You do not need to wait.

2. Frederick County Enabled Gemini for Middle and High School Students Before Its Board Finished the AI Policy State Law Requires by October 22, and Board Members Asked Who Has the Power to Turn It Off

Source type. District implementation and board deliberation, Maryland. Authoritative secondary reporting. Not research.

Rubinstein, S. (2026, September 22). As FCPS allows Google Gemini as AI tool, some board members raise questions. The Frederick News-Post. Republished as: Frederick County school board questions Google Gemini on student devices. Government Technology. govtech.com Original: fredericknewspost.com

Frederick County Public Schools enabled Google Gemini for middle and high school students starting this school year, under conditions set by the teacher and primarily for teaching AI literacy, according to curriculum director Scott Murphy. All other generative AI platforms remain blocked on school devices. Murphy said the district chose Gemini because it meets the state's safety and privacy standards under the AI Ready Schools Act. Several board members objected. Jaime Brennan said it was "unbelievable to have FCPS roll out AI implementation without any warning to parents." Karen Yoho said the district should disable the tool: "we don't want our kids being the guinea pigs." Janie Inglis Monier asked for regulations addressing the school system's ability to disable or disconnect AI tools. Dean Rose first supported disabling it, then softened his position pending community discussion. According to the report, the board must finalize its AI policy by October 22, 2026, to comply with state law.

Study context. One outlet reported it and a second republished it; the board video and district statement were not opened. The October 22 date is as reported for Frederick County; Maryland Matters reported in July that the AI Ready Schools Act gives each of Maryland's 24 districts 120 days after state guidance to adopt a policy, creating a fall deadline. Still, the state source did not confirm a single statewide date. [Flagged: rung two.]

Leadership implication. This is the governance failure most districts are one admin console away from. AI now arrives as a setting inside Google Workspace, Microsoft 365, the learning management system, and the reading platform, not as a new purchase that triggers procurement review. Direct your chief technology officer and chief academic officer to produce, together, an activation authority matrix: every AI capability available in systems the district already licenses, who may turn it on, what review is required first, whether parents are notified, and whether the board is told. Frederick's board member asked the right question in the wrong order: the power to turn AI off should be written before anyone turns it on. Georgia districts running Gemini programs should answer that question in writing before a board member asks it in public.

3. North Carolina's State Education CIO Said AI Memory Should Be Treated as a New Data Category and That Schools May Not Know Which of Their Products Contain AI

Source type. Trade reporting on an expert webinar. Not research. Not policy.

Gilban-Cohen, J. (2026, September 25). AI's memory capabilities have implications for K-12. Government Technology. govtech.com

Government Technology reported on an FETC webinar facilitated by the American Enterprise Institute with Matt Gee, director of U.S. program data at the Gates Foundation; Andrew Rice, CEO of Education Analytics; and Vanessa Wren, chief information officer of the North Carolina Department of Public Instruction. The panel examined AI systems that retain information from earlier interactions. Wren said, "AI tool adoption cannot be based on trust. We have to look at AI memory as a new data category," and, "Our schools have hundreds of products that they might not even know that has AI in it." Gee described the risk of a persistent record: "If that record contains things that they don't want in there, or that are really misunderstandings of who they are, that could hold them back." The panel identified procurement contracts as the primary governance tool, pointed to state student privacy laws that restrict targeted advertising and data sales, and called for data portability that lets a student or family review and edit a record before it moves between systems.

Study context. Expert commentary, not a study and not state policy. Wren's remarks are her statements on a panel and were not presented as North Carolina Department of Public Instruction guidance. The webinar date was not confirmed; the article is dated September 25.

Leadership implication. Your data inventory lists what vendors collect. It almost certainly does not list what their AI remembers or concludes. Direct your privacy or data governance lead to add four questions to every renewal for a tool with AI features: what prompts, histories, profiles, and inferred attributes persist; for how long; whether the district and the family can inspect, correct, and delete them; and whether they can move to another system or influence a later automated action. Put the answers in the contract, not the vendor's privacy page. Wren's second point is the harder one: before you can govern AI memory, you need an inventory of which products have AI at all. Assign that inventory a date.

4. A Federal Bill Would Bar Training AI on Student Data and Require Parental Opt-Out, While California's Governor Has Until September 30 on Two Bills That Reach Public Bodies and Public Workers

Source type. Introduced federal legislation and pending state legislation. Neither is law. Not research.

Office of Congresswoman Suzanne Bonamici. (2026, September 22). Bonamici introduces human-centered framework to provide guardrails for AI in education and workforce [Press release]. bonamici.house.gov Office of Governor Gavin Newsom. (2026, September 20). Governor Newsom issues legislative update 9.20.2026. gov.ca.gov CalMatters Digital Democracy. (2026). SB 1159 and AB 2656 bill pages. calmatters.digitaldemocracy.org ; calmatters.digitaldemocracy.org

Rep. Suzanne Bonamici introduced the Artificial Intelligence Education and Workforce Readiness Act on September 22 with cosponsors Ted Lieu, Valerie Foushee, Josh Gottheimer, Raja Krishnamoorthi, Jimmy Panetta, Bennie Thompson, Frederica Wilson, and Greg Landsman. According to her office, the bill would set risk-assessment standards for classroom AI tools, prohibit using student data to train AI models, require parental notice and opt-out rights, and create a breach-reporting portal enforced by the Department of Education and the Federal Trade Commission. It would require human oversight, bias testing, data security protocols, and independent audits for AI systems in federally funded education. It would fund in-service training for school leaders and open-access research on AI literacy and on AI's classroom effects. In California, as of the Governor's September 20 update and his signing announcements through September 26, no action had been taken on SB 1159, which would specify that AI cannot be a person under the Public Records Act and the open meeting laws, including the Ralph M. Brown Act that governs school boards, and would bar knowingly using AI to pose as a natural person engaging with a government agency; or on AB 2656, which would require state and local public employers to give a recognized employee organization 45 days' notice before deploying generative AI to perform work within its scope. The deadline is September 30.

Study context. The federal provisions are as described in the sponsor's release; the bill text and number were not opened, and no Republican cosponsor is listed, which bears on its prospects. California bill content is confirmed at CalMatters because the Legislature's pages did not render. Whether AB 2656 reaches K-12 districts, which bargain under a separate statute, depends on text not opened. [Flagged: rung two for both.] H.R. 10315, the other federal AI education bill this brief is tracking, has seen no action since September 8.

Leadership implication. Read the federal bill as a checklist, not a forecast. Its core provisions already appear in instruments this brief has reported: a bar on training AI with student data in California's AB 1159 and Florida's rule, parental notice and consent in Florida's rule, and a human decision requirement in Florida's rule and Atlanta's draft. When the same provisions recur across a statute, a rule, a board draft, and a federal proposal, they are becoming the market's minimum whether or not this bill moves. Check your standard data privacy agreement against the bill's list this month: no training on student data, parental notice and opt-out, human oversight, bias testing, independent audit. On SB 1159: if your board accepts public comment electronically, ask your board secretary how the district would know a comment came from a person.

5. AI Tutors Held Their Rankings Against Simulated Disengaged Students, but Performance Varied by Learner State and Automated Scoring Did Not Fully Match Human Judgment

Source type. Preprint, not peer-reviewed. Simulation and evaluation study. Posted to arXiv in September 2026.

Meng, X., & Lin, J. (2026). Simulating disengaged students to evaluate LLM-based tutors. arXiv. arxiv.org

Xianghui Meng and Jionghao Lin of the University of Hong Kong start from a problem every district pilot shares: AI tutors are usually evaluated with students who cooperate. Real classrooms include students who game the system, spin their wheels on the same error, or go off task. The authors built student simulators conditioned on five engagement states, engaged, gaming, wheel-spinning, off-task, and mixed, and validated the labels with human coders, who agreed on 84 percent of sessions (Cohen's kappa of 0.78). Conditioning the simulator on the intended state brought simulated behavior closer to authentic student logs, cutting the correctness-rate gap from 0.54 to 0.20 for gaming and from 0.51 to 0.18 for wheel-spinning. They then evaluated five AI tutors from the Claude, Llama, Gemini, Qwen, and GPT model families. Relative rankings remained stable across learner states and interaction lengths, but absolute performance varied. The authors report that automated tutor evaluation was not fully aligned with human judgment and that a stable tutor ranking can conceal state-specific differences in tutor support.

Study context. A preprint with no named venue. It tests tutors against simulated students and measures tutor behavior, not student learning. Its arXiv identifier places its posting in early September, near the start of the research window; the exact day could not be confirmed because arXiv's abstract page did not render. Edition 34 did not report it.

Leadership implication. If Maryland's demonstrated-outcomes standard spreads, the next question is demonstrated for whom. Before renewing any AI tutor, ask the vendor in writing: which students did you test against, did that include students who are disengaged, guessing, or stuck, and did a person review the tutor's responses or did another model grade them? Ask for results broken out by learner behavior, not only an aggregate score. A leaderboard tells you which tool wins on average. It does not tell you what happens to the student you bought it for.

Emerging Strategic Themes

Theme 1. Safety evidence is not effectiveness evidence. Florida requires an effectiveness review but permits approval without studies. Maryland proposes that tools show learning outcomes before use. Instruction Partners, in Edition 34, found seven of sixteen products with independent achievement studies across student groups. Boards are routinely shown privacy compliance and told a tool is vetted. Ask to see the two kinds of evidence in separate columns, and expect the second one to be mostly empty.

Theme 2. Activation has replaced procurement as the point of decision. Google extended Gemini in Classroom to K-12 students in August for districts that had already enabled student access. Frederick County turned it on under teacher conditions before its policy was final. DeKalb has a certification program with no board record. When AI arrives as a feature, the procurement gate never opens. Districts now need an activation authority rule, not another purchasing checklist.

Theme 3. The next student record is one the district did not write. A state CIO called AI memory a new data category; a Gates Foundation data director warned that a persistent record built on misunderstandings could hold a student back; a research team showed tutors behaving differently by inferred engagement state. The chain runs from what a system sees, to what it remembers, to what it infers, to what it recommends. Student privacy law governs the first link. Contracts will have to govern the rest.

Theme 4. Districts are building standing bodies, not one-time policies. Chico Unified voted unanimously on September 23 to create a five-to-seven-member AI and technology work group with teachers, administrators, parents, students, and IT staff that will hold public forums. Santa Barbara Unified's AI task force opened its year on September 25 with four working groups and eleven students. Los Angeles Unified's committee has meetings calendared into April 2027. WBUR reported September 22 that Massachusetts districts, under state guidance that leaves implementation local, range from Boston approving Gemini for teenagers while blocking ChatGPT to districts with no clear policy. AI changes too quickly to govern by a memo written once. The capacity to teach it is its own problem: a Purdue preprint this month proposes borrowing mentors from nearby colleges to reach rural schools, projecting statewide reach in Indiana only over forty years.

What Was Not Found

No binding K-12 AI rule was adopted by any state or federal body between September 20 and 26. Maryland's framework is a set of commitments; the federal bill is an introduction; California's pending bills are unsigned. New York's S9051 and A6578 remain unsigned against a December 31 deadline. H.R. 10315 has seen no action since September 8. No Kansas State Board of Education action was located. No new settlement, ruling, or filing in the PowerSchool, Character.AI, or New Mexico Amira Learning matters was found in the window.

No peer-reviewed K-12 AI study with a publisher-verified online date between September 13 and 27 was located, and the search was incomplete. This is the fifth consecutive edition without one. The Consensus database, this brief's primary research source, reached its monthly limit during this build and resets October 1; ten queries ran before the limit, and the rest of the sweep relied on web and journal searches. Most 2026 items the database returned carried publication dates from February through July once checked at the publisher. The 371-student Educational Psychology Review trial in Edition 34 was dated April 2026 on this week's check and was removed from the hold list. [Flagged: rung two.] A September survey of 428 teachers in Jordan in the Journal of Educational and Social Research was located but could not be opened, and it is not reported. Digital Promise announced the second cohort of grantees in its $26 million, four-year K-12 AI Infrastructure Program on September 21, which funds openly licensed datasets, models, and benchmarks; the announcement page did not render for this build, so the grantee list and project descriptions are not reported.

No definition of "demonstrated learning outcomes" exists in any instrument this brief has tracked. Maryland's framework uses the phrase without a standard. Florida's rule names ESSA studies as examples and then permits approval without them. The federal Dear Colleague letter of August 20 asked districts to buy based on demonstrated outcomes with no federal evidence base behind it. No state has said what study design, sample, or effect size would count, or whether outcomes must be shown separately for English learners, students with disabilities, and students in poverty.

No evidence exists on what AI systems remember or infer about students, or how tutors perform with disengaged students in real classrooms. No audit, study, or vendor disclosure in any window of this brief has inventoried persistent memory or inferred attributes held by K-12 AI products. The one study this week on disengaged students used simulations.

No Georgia board acted on AI in the window, and DeKalb's certification program has no public record. DeKalb County's Gemini AI Certification Program passed its September 21 release date with no enrollment notice, board item, vendor agreement, or data privacy terms found. A similar program in Douglas County, Colorado, surfaced repeatedly in searches and is a different district. Atlanta's draft Policy IFBI awaits October committee review and a reported November decision. Fulton County's adoption of Policy IFBI remains unconfirmed in board minutes. No action was located in Gwinnett, Cobb, Cherokee, Forsyth, Clayton, or Henry.

No instrument this week separates elementary literacy, English learners, or students with disabilities. Maryland's framework, the federal bill as described, and the week's research are silent on outcomes by student group. The gap this brief has named since spring persists.

The pattern held with a new edge. For months, policy asked whether AI was safe. This week, a governor asked whether it works, a board asked who switched it on, and a state CIO asked what it remembers. Each question arrived before the evidence or the rule to answer it. The districts that write their own answers now (an evidence column, an activation authority, a memory clause) will not be rewriting them when the rule arrives.

Novo Executive Summary

No binding K-12 AI rule was adopted this week, but the standard that future rules will be measured against moved. Maryland's governor proposed that classroom AI demonstrate learning outcomes, supplement teachers, be disclosed to families, and come with a crisis-response protocol. A Maryland district showed how that standard collides with practice when Gemini reached students before its board finished the policy state law requires by October 22. A state education CIO called AI memory a new category of student data, and a federal bill proposed barring AI training on student data with parental opt-out. The week's preprint showed AI tutors performing differently for disengaged students while holding their rankings. The strategic conclusion is that governance now runs along a chain: what AI sees, what it remembers, what it infers, who can switch it on, and whether any of it improves learning. Novo Innovative Pathways builds that chain with district leaders, from activation authority and evidence review to memory and inference clauses in contracts and the role-based AI literacy that puts a trained adult at every link.

Watch This Week

• Wednesday, September 30, 2026. Governor Newsom's deadline on SB 1159, AB 2656, and AB 2392. Carried from Editions 29 through 34.
• Wednesday, October 21, 2026. Los Angeles Unified Board Generative AI Committee, second meeting, confirmed through LAist's published calendar; recommendations aimed at a full-board vote in 2027.
• Thursday, October 22, 2026. Frederick County Board of Education's reported deadline to finalize its AI policy under Maryland's AI Ready Schools Act. Watch whether the policy addresses the power to disable AI tools and whether the Gemini pilot continues. Other Maryland districts face fall deadlines under the same law.
• Maryland 2027 session. Legislation implementing the governor's "do no harm" standard and classroom crisis-response protocol, and Maryland State Department of Education model K-12 AI lessons. No dates announced.
• Georgia. Atlanta's Policy Review Committee returns to draft Policy IFBI in October; decision reported for November. Fulton County's October minutes. Any DeKalb County board item, contract, or data terms for the Gemini certification program.
• Also tracking. Santa Barbara Unified AI task force, October 29. Chico Unified's special session on its new AI work group, not yet scheduled. Microsoft's National AI Safety and Privacy Standard, November 1. Governor Hochul on S9051 and A6578 by December 31. Congressional referral and any Republican cosponsor for the AI Education and Workforce Readiness Act. Florida Administrative Register filing of Rule 6A-1.0957.

Sources

Governance and Policy

Office of Governor Wes Moore. (2026, September 22). Governor Moore outlines AI framework to protect Marylanders [Press release]. governor.maryland.gov

Office of Governor Wes Moore. (2026). Protecting Marylanders in the age of artificial intelligence. governor.maryland.gov

Rubinstein, S. (2026, September 22). Frederick County school board questions Google Gemini on student devices. Government Technology (republished from The Frederick News-Post). govtech.com

Nachman, M. (2026, July 6). Maryland school districts face fall deadline to set AI policies. Maryland Matters. marylandmatters.org

Gilban-Cohen, J. (2026, September 25). AI's memory capabilities have implications for K-12. Government Technology. govtech.com

Office of Congresswoman Suzanne Bonamici. (2026, September 22). Bonamici introduces human-centered framework to provide guardrails for AI in education and workforce [Press release]. bonamici.house.gov

Office of Governor Gavin Newsom. (2026, September 20). Governor Newsom issues legislative update 9.20.2026. gov.ca.gov

CalMatters Digital Democracy. (2026). SB 1159: Artificial intelligence: transparency and governance; AB 2656: Public employees: notice: artificial intelligence performing service within scope of work. calmatters.digitaldemocracy.org ; calmatters.digitaldemocracy.org

Jiang, G. (2026, September 1). Schools caught flat-footed after Google makes Gemini chatbot available to all students. Education Week. edweek.org

Myers, E. S. (2026, September 24). Chico Unified mints AI & tech advisory committee. ChicoSol News. chicosol.org

Wallace, S. (2026, September 26). Students, parents, teachers want their voices heard on artificial intelligence in Santa Barbara schools. Santa Barbara News-Press. newspress.com

Lee, S. (2026, September 22). School districts, slowly, build guidance around AI learning in classrooms. WBUR. wbur.org

Dale, M. (2026, September 1). LAUSD AI committee first meeting. LAist. laist.com

Research, Peer-Reviewed

No peer-reviewed study with a publisher-verified online date between September 13 and 27, 2026 met the selection standard. The search was incomplete; see What Was Not Found.

Research, Preprint, Not Peer-Reviewed

Meng, X., & Lin, J. (2026). Simulating disengaged students to evaluate LLM-based tutors. arXiv:2609.12331. arxiv.org [Flagged: exact posting day not confirmed. Content verified through the arXiv HTML full text.]

Jacobson, M. J., Rodriguez-Rivera, G., Drineas, P., & Xue, Y. (2026). Teaching AI, robotics, & community: A hubs-based K-12 education framework for reaching rural schools. arXiv:2609.18072. arxiv.org [Flagged: exact posting day not confirmed. Content verified through the arXiv HTML full text.]

Institutional Report, Not Peer-Reviewed

Digital Promise. (2026, September 21). New grantees awarded by the K-12 AI Infrastructure Program. digitalpromise.org