Michigan Defaults, Tennessee Restricts: K–12 AI Governance Hits Convergence

Framing

Michigan's Department of Education published a comprehensive Artificial Intelligence guidance package for the state's K-12 districts on May 12. Read in isolation, it is one more state document. Read against the running ledger of Boston, Columbus, Maryland, New York City, Idaho, Illinois, California, and now Michigan, it is a marker that state-level AI guidance has crossed from exploratory to expected. Districts that have been waiting for clarity from above are running out of cover. The pattern is becoming defaulted at the state level rather than locally constructed.

Two state actions this week sit in tension. Michigan's framework is permissive in posture, prescriptive in scaffolding. It encourages districts to engage AI but binds that engagement to four practices: purposeful and safe use, privacy and integrity, AI literacy, and human oversight. Tennessee's Grace Anne Sparks Coercive Suicide Prevention Law of 2026, by contrast, is restrictive on its face. It bars any AI system from being marketed or represented as a licensed mental health professional, with a private right of action attached. Districts are increasingly the institutional surface where permissive integration guidance meets restrictive safety law. That surface is where harm, liability, and reputational risk will land first.

This week's peer-reviewed research sharpens the procurement reading. A randomized controlled trial in Educational Psychology Review with 371 students in grades 7 to 9 finds that standard ChatGPT use, without structured pedagogical prompting, did not preserve students' utility value or strengthen strategy use. Only the conditions paired with motivational or cognitive strategic scaffolding showed measurable benefit. A new cross-national preprint examining teacher perceptions of AI across 55 countries reports that current large language model outputs do not faithfully reflect the benefit-risk balance teachers actually hold, drifting toward optimism even when steered. The signal to districts is direct. The tool alone is not the instructional intervention. The model alone is not aligned with your educator workforce. Governance architecture is what closes both gaps.

Top Research and Policy Signals

1. Michigan Department of Education releases comprehensive K–12 AI guidance

Source type. State-level institutional guidance (Michigan Department of Education). Released May 12, 2026.

The Michigan Department of Education published a starter guide and a longer, comprehensive guide for all of the state's traditional and charter districts. The framework is anchored on four practices: keeping AI purposeful and safe, protecting privacy and integrity, building AI literacy across the workforce, and maintaining human oversight in instructional and operational decisions. The package includes templates, practical examples, and administrative scaffolding meant to be used directly by district cabinets, technology directors, and board members rather than translated by a consultant. State Superintendent guidance is paired with a recommendation that each district begin or continue formal conversations about the use of AI tools.

Leadership implication. District cabinets in Michigan should treat this guidance as immediate. Compliance work this summer should map current vendor inventories against the four pillars, write or update an AI acceptable use policy that names a human owner for each high-stakes decision, and identify professional development gaps that will be visible at the start of the next school year. Districts outside Michigan should still read the framework. It is now one of the most useful default templates in the Midwest and will be referenced in procurement decisions in neighboring states.

2. Tennessee Grace Anne Sparks Law restricts AI from mental health roles

Source type. State legislation (Tennessee SB 1580). Signed by Gov. Bill Lee on April 1, 2026. Effective July 1, 2026.

The statute bars any person who develops or deploys an AI system from advertising or representing to the public that the system is, or can act as, a qualified mental health professional. The law is short, less than one page, and includes a private right of action. While framed as a consumer protection statute, it sits directly atop the rapidly expanding category of student-facing AI wellness chatbots and companion tools now being piloted in districts across the country. A growing number of vendors describe their products using therapeutic, counseling, or coaching language that would be subject to Tennessee's standard if marketed to or used by Tennessee minors.

Leadership implication. District counsel should immediately review the marketing claims of any deployed or contemplated student wellness AI tool against the Tennessee standard, even if the tool is used outside Tennessee. The statute is short, narrow, and likely to be copied by other states. Procurement should require vendors to certify in writing that the tool is not represented as a mental health professional and to indemnify the district against private right-of-action claims arising from such representation. Cabinets should also clarify staff escalation protocols so that no chatbot ever sits between a student in crisis and a credentialed human.

3. Cross-national audit finds LLMs misaligned with teacher risk perceptions

Source type. Preprint, not peer-reviewed. Submitted to arXiv May 8, 2026.

Using representative international survey data from teachers in 55 countries, paired with a systematic audit of large language model behavior, the authors find that LLM outputs do not reliably match the benefit-and-risk patterns teachers actually hold. Models default toward optimism about AI in education and do not adequately surface the country-specific or context-specific risks that teachers identify when surveyed directly. Steerability through prompting partially closes the gap but does not eliminate it, particularly for risk dimensions that teachers in lower-resource contexts emphasize more strongly than their peers in higher-resource contexts.

Leadership implication. Procurement teams should treat alignment with educator workforce perception as a measurable specification, not a marketing assertion. Add to RFPs a requirement that vendors demonstrate how their model surfaces risk dimensions in the language and context of the local educator workforce, and request published evidence of steerability testing. A model that drifts toward optimism by default is an institutional risk amplifier in a district whose educators are correctly more cautious than the model.

4. RCT shows GenAI improves student outcomes only with pedagogical scaffolding

Source type. Peer-reviewed research article. Educational Psychology Review (Springer Nature). Published online April 14, 2026.

A theory-informed randomized controlled trial with 371 students in grades 7 to 9, conducted across regular physics and English class sessions, compared three conditions: a GenAI utility-value reflection condition, a GenAI cognitive strategy prompting condition, and a control condition using standard ChatGPT without pedagogical prompting. Across six 45-minute sessions, the structured conditions preserved students' utility value and supported strategy use, while the standard ChatGPT control did not produce comparable benefits. The authors are explicit that the gains depend on the pedagogical scaffolding surrounding the tool, not on exposure to the tool itself.

Leadership implication. Districts that purchase GenAI tools without funding the pedagogical scaffolding that makes them effective are likely to see no measurable instructional return. RFP rubrics should require vendors to specify, with evidence, the cognitive or motivational scaffolding embedded in their product, and should treat unscaffolded tools as supplementary rather than instructional. Professional development cycles should explicitly train teachers in prompting strategies that mirror the utility-value or cognitive-strategy conditions documented in the published trial.

5. Federal K–12 AI Literacy and Readiness Act introduced in House

Source type. Federal legislation introduced in the U.S. House of Representatives by Rep. Randy Fine (R-FL).

The bill would amend the Elementary and Secondary Education Act of 1965 to explicitly allow existing federal education funds to be used for student instruction on the safe and responsible use of AI, and for teacher training and professional development on AI. The bill does not create new appropriations. The Computer & Communications Industry Association issued a public statement of support, calling it a way to reduce friction in deploying federal funding for K-12 AI literacy.

Leadership implication. Federal funding flexibility for AI literacy may arrive faster than district planning. If your organization views AI literacy as a discretionary professional development line, please re-examine federal Title formula funds and prepare narratives that anticipate eligibility expansion. Treat the bill as a planning signal, not an authority. The signal value is strong because it aligns directly with the Department's April Final Priority and Definitions on Advancing AI in Education, which became effective May 13, 2026.

Emerging Strategic Themes

Theme 1. The four-pillar pattern is consolidating at the state level.

Michigan's structure of purpose and safety, privacy and integrity, AI literacy, and human oversight echoes prior frameworks from New York City, Idaho, California, and Maryland. The vocabulary is converging. Districts that adopt these four anchors directly will move faster through audits, board approval cycles, and counsel review than districts that invent local taxonomies.

Theme 2. Mental health AI has become a litigation vector.

Tennessee's Grace Anne Sparks Law is unlikely to be the last. Other states are reading the same evidence of teen chatbot use and harm. Districts that have piloted student wellness AI or that allow personal companion apps on issued devices should expect that procurement and acceptable use language will face plaintiffs' attention. Counsel-led vendor audits this summer are appropriate.

Theme 3. LLM alignment with the educator workforce becomes a procurement specification.

The cross-national preprint by Tao and colleagues turns a quiet concern into a written specification. If your educator workforce is more cautious than the model, your model is misaligned. Procurement should treat steerability and risk surfacing as evaluable line items rather than as aspirational vendor claims.

Theme 4. GenAI without pedagogy has no effect size.

The Educational Psychology Review trial is the clearest school-based evidence to date that unstructured exposure to GenAI does not produce learning gains. This is no longer a controversial claim. Tool acquisition without scaffolded pedagogy is an expense, not an instructional intervention.

What Was Not Found

Five gaps remain visible in the peer-reviewed and policy record this week.

First, no peer-reviewed causal evaluation yet exists for any state-level AI guidance framework, including Michigan's new four-pillar package or New York City's red, yellow, green model. Districts are adopting these structures without evidence of outcomes.

Second, the strongest school-based RCT this week, with 371 students in grades 7 to 9, is implemented outside the United States. No equivalent rigorous U.S. K-12 RCT on GenAI for self-regulated learning is available. Districts cannot generalize without local replication.

Third, there is no published evidence on the harm rates associated with AI mental health and wellness chatbots in school settings. Tennessee is regulating a category for which the empirical baseline is largely anecdotal.

Fourth, no peer-reviewed evidence speaks to LLM alignment for the specific risk patterns of high-poverty U.S. districts, English learners at scale, or students with disabilities. The cross-national preprint highlights the alignment gap, but the demographic composition of that gap remains unstudied in U.S. K-12.

Fifth, no causal study has yet evaluated whether AI fluency graduation requirements, of the kind in Boston Public Schools that are rolling out in September, change postsecondary or workforce outcomes. Districts are building graduation pipelines without the data to defend them.

Novo Executive Summary

The state action this week confirms that districts no longer have the option of waiting for federal clarity. Michigan has consolidated a four-pillar framework that will be referenced across the Midwest. Tennessee has shown how restrictive laws will land on the same tools that integration guidance encourages. Peer-reviewed and preprint research together confirm that the tool alone is neither an instructional intervention nor a trustworthy interface. Districts now sit at the convergence of permissive guidance, restrictive law, and uneven research. The work this summer is governance architecture, role-based AI literacy, and procurement language that survives both audits and litigation. Novo Innovative Pathways helps districts and cabinets build that architecture, train cabinet and educator roles to the convergent standard, and align policy with federal and state regimes that will not wait.

Watch This Week

• Illinois State Board of Education statewide AI guidance due July 1, 2026, under SB 1920.
• Tennessee Grace Anne Sparks Law (SB 1580) effective July 1, 2026, with private right of action.
• U.S. House K-12 AI Literacy and Readiness Act of 2026, introduced by Rep. Randy Fine; monitor referral to House Education and Workforce Committee.
• Michigan Department of Education guidance rollout; watch for accompanying technical assistance and timelines.
• New York City Public Schools' comprehensive Playbook drafting window approaching June 2026 release.
• Boston Public Schools AI fluency programming; teacher training enrollment for September 2026 launch.