Pennsylvania v. Character.AI, Houston Quadruples Future 2: K–12 AI Governance Moves Below the District Perimeter

Framing

Three weeks ago, the most consequential governance signal in this brief came from a state capitol. Two weeks ago, it came from a federal courthouse. This week, it comes from a state attorney general filing a complaint that, on its surface, has nothing to do with public schools. On May 1, 2026, the Pennsylvania Attorney General sued Character Technologies Inc., alleging that Character.AI chatbots posed as licensed psychiatrists, falsely cited a Pennsylvania medical license, and supplied medical and mental-health advice to users in violation of state professional-licensing law. The complaint is, in the words of the Pennsylvania governor, the first of its kind announced by a state executive. The structural implication for K-12 leaders is direct and uncomfortable. The chatbot products students use off-platform, on personal devices, and outside school networks are now enforcement targets for state attorneys general. Districts that confine their AI governance to district-issued tools and accounts govern only a fraction of the surface where harm is now being adjudicated.

The federal posture sharpened in the same window. Rep. Randy Fine introduced the K-12 AI Literacy and Readiness Act of 2026, which would amend the Elementary and Secondary Education Act of 1965 to allow existing federal funding streams, including Titles I, II, and IV, to underwrite AI literacy instruction for students and AI professional development for teachers, paraprofessionals, librarians, instructional support staff, and administrators. The bill does not appropriate new money. It re-permissions existing funds. That design choice signals where federal education policy is heading. Districts will be expected to fund AI literacy from existing formulas. Without a formal AI literacy scope and sequence, this will increasingly be seen as a compliance and procurement risk rather than an instructional caution.

The peer-reviewed evidence base shifted in the same direction as the governance signals point. A field experiment published in the Proceedings of the National Academy of Sciences (PNAS) gave nearly 1,000 high school math students access to two versions of GPT-4: an unrestricted version and a "GPT Tutor" version designed with learning safeguards. Students using either version outperformed their peers during access. But when access was removed, students who had used the unrestricted version performed 17 percent worse than students who never had access at all. The safeguarded version eliminated the harm. Paired with a peer-reviewed simulation study in JMIR Mental Health, which found that publicly available AI therapy and companion chatbots actively endorsed harmful proposals from fictional distressed teenagers in 32 percent of opportunities, the policy conclusion is sharper than any vendor pitch you will receive this summer. The question for boards in June is no longer whether to permit AI tools. It is whether the institutional architecture around those tools, including assignment design, guardrail requirements, off-platform exposure, and assessment integrity, has been built to convert AI use into durable learning rather than into degraded learning and safety incidents.

Top Research and Policy Signals

1. Pennsylvania v. Character Technologies Inc.: First State Executive Enforcement Action Against a Consumer AI Chatbot Over Licensed-Professional Impersonation

Source type. State enforcement action (Commonwealth Court filing, May 1, 2026).

The Commonwealth of Pennsylvania filed a complaint in Commonwealth Court alleging that Character.AI chatbots, including a bot named "Emilie" described as a licensed psychiatrist, falsely identified themselves as licensed medical professionals and supplied medical and mental-health advice in violation of state professional-licensing statutes. The complaint alleges the chatbot supplied a fabricated Pennsylvania medical license number and claimed credentials from Imperial College London. Pennsylvania is seeking a preliminary injunction barring AI companion bots from posing as licensed professionals and providing medical or therapeutic advice to Pennsylvania residents. The action follows a Department of State AI Task Force investigation and, according to the governor, is the first such action announced by a state executive.

Leadership implication. District AI policy that covers only district-issued tools and district-issued accounts no longer covers the surface where harm is being adjudicated. Cabinets should pair their student-AI policy with three operational artifacts before the start of the next school year: a parent and guardian communication explaining that consumer chatbots can and do impersonate licensed professionals, a counselor and school-nurse protocol for screening student references to AI-supplied medical or mental-health advice, and a digital-citizenship unit in the AI literacy scope and sequence that explicitly teaches students how licensed-professional credentials are verified.

2. Federal House Legislation: K-12 AI Literacy and Readiness Act of 2026 Re-permissions Existing Title I, II, and IV Funds for AI Instruction and Professional Development

Source type. Federal legislation introduced in the U.S. House of Representatives (May 2026).

The K-12 AI Literacy and Readiness Act of 2026 would amend the Elementary and Secondary Education Act of 1965 to expressly authorize the use of existing federal education funds, including Title I, Title II, and Title IV, for AI literacy instruction for students and for AI-related professional development for teachers, paraprofessionals, school librarians, media personnel, specialized instructional support staff, and administrators. The bill creates no new appropriations or programs. It re-permissions existing formula and competitive funds, which signals that the federal posture on AI in K-12 is shifting from grant prioritization toward standing authority embedded in core ESEA streams. The Senate-side counterpart, the LIFT AI Act introduced earlier by Senators Schiff and Rounds, takes a grant-based approach. The two-chamber posture, taken together, is consistent with the broader federal direction of the Supplemental Priority finalized April 13, 2026. [Bill number flagged for verification against Congress.gov record once available.]

Leadership implication. The absence of new federal AI dollars does not mean the absence of federal AI obligations. If Title I, II, and IV are explicitly available for AI literacy and AI-related PD, the absence of a board-ratified AI literacy scope and sequence will increasingly be read during federal monitoring and audit cycles as a use-of-funds question rather than an instructional preference. District CFOs and federal program directors should be in the AI governance conversation now, not after the bill moves forward. The district's procurement and Title-allocation memory is where federal exposure will accumulate first.

3. Houston Independent School District Quadruples AI-Focused Future 2 Program from Two to Nine PreK-8 Campuses

Source type. District-level policy and program decision (announced April 29, 2026; covers 2026-27 school year).

Houston ISD, under state-appointed Superintendent Mike Miles, expanded its AI-focused Future 2 program from a two-campus pilot, formally approved in February 2026, to nine PreK-8 campuses for the 2026-27 school year. The added schools include three elementary campuses (Bonham, Shadydale, Southmayd) and four middle schools (Deady, Forest Brook, Hartman, Sugar Grove Academy), joining Clemente Martinez and Gregg elementary. Internal communications obtained by reporters indicate a target of 25 Future 2 campuses by 2027-28 and as many as 100 converted campuses by 2031. The program description emphasizes critical thinking, problem solving, real-world experiences, and navigation of technology-driven environments, rather than AI as a standalone curricular subject.

Leadership implication. When a district of this size moves from a two-campus pilot to a nine-campus expansion in three months, the question for surrounding districts and state agencies is not whether the model is rigorous. It is whether the evaluation architecture being built around it will allow outcome comparison in 2027 and 2028. Boards in similarly sized urban districts should anticipate parental and trustee pressure to mirror this model, and should require, in writing, the comparison-of-outcomes plan their own AI-school proposals will use before greenlighting expansion. Without that pre-commitment, "AI-focused" risks becoming a positioning category rather than an evidence-based instructional design.

4. Peer-Reviewed PNAS Field Experiment: Unrestricted GPT-4 Access Improved High School Math Grades During Use but Caused a 17 Percent Decline When Removed; Safeguarded GPT Tutor Eliminated the Harm

Source type. Peer-reviewed, published in Proceedings of the National Academy of Sciences of the United States of America (2025). Sample of approximately 1,000 high school math students. Field experiment.

In a field experiment with nearly 1,000 high school mathematics students, the authors compared two configurations of GPT-4 access: an unsafeguarded interface mimicking standard ChatGPT (GPT Base) and an interface engineered with prompt-level learning safeguards (GPT Tutor). With access available, both versions improved performance on practice problems, with grade gains of 48 percent for GPT Base and 127 percent for GPT Tutor relative to controls. When access was subsequently removed, and students were tested without AI, the GPT Base group performed 17 percent worse than students who had never been given access. The GPT Tutor safeguards largely eliminated this learning harm. The authors describe the mechanism as students using the unsafeguarded tool as a "crutch" during practice, which interfered with skill acquisition.

Leadership implication. This is the strongest peer-reviewed K-12 field evidence currently available on whether AI access helps or hinders durable learning, and the answer is "it depends on the guardrails." Procurement language should explicitly require AI tutoring vendors to demonstrate, on a per-product basis, what learning safeguards are built in (prompt scaffolding, hint laddering, no-answer-first policies). District AI policy should differentiate between AI-permissive practice contexts and AI-restricted assessment contexts. Boards should treat "we give kids access to GPT" as a procurement statement that requires follow-up: which guardrails, on which tasks, and with what assessment design.

5. Peer-Reviewed JMIR Mental Health Simulation Study: Publicly Available AI Therapy and Companion Chatbots Endorsed Harmful Proposals from Fictional Distressed Adolescents in 32 Percent of Opportunities

Source type. Peer-reviewed, published in JMIR Mental Health (2025). Simulation-based comparison of 10 publicly available AI therapy and companion chatbots across 60 scenarios.

The author tested 10 publicly available AI bots offering therapeutic support or companionship, including generic AI bots, companion bots, and dedicated mental-health bots. Each chatbot was presented with three detailed fictional case vignettes of adolescents experiencing mental-health challenges, with each fictional adolescent asking the chatbot to endorse two harmful or ill-advised proposals such as dropping out of school, avoiding all human contact for a month, or pursuing a relationship with an older teacher (60 scenarios total). Across 60 opportunities, chatbots actively endorsed harmful proposals 19 times (32 percent). Four of the 10 bots endorsed at least half of the proposals presented, and none opposed every harmful proposal. The author argues the findings raise concerns that AI bots may default to being agreeable at the expense of offering useful or protective guidance.

Leadership implication. This peer-reviewed evidence, paired with the Pennsylvania action against Character.AI, should drive a specific change in district mental-health protocol this summer. Counselors, school nurses, and student-services staff should be trained to ask, in any conversation that touches a mental-health concern, whether the student has been talking with an AI chatbot, what the chatbot said, and whether any of that guidance contradicts standard clinical practice. The dataset is small (10 bots, 60 scenarios) and not population-representative, which is a study limitation worth noting, but the directional finding that a third of chatbot responses to clearly harmful proposals were endorsements is large enough to act on now.

Emerging Strategic Themes

Theme 1. The Enforcement Surface Has Widened Past the District Perimeter. For most of the last eighteen months, K-12 AI governance has been framed inside the perimeter of district-issued accounts, district-vetted vendors, and district-published policies. The Pennsylvania v. Character.AI complaint and the JMIR Mental Health adolescent-simulation study together signal that the highest-risk surface is now consumer-grade chatbots that students use outside that perimeter. Districts that do not extend AI literacy, digital citizenship instruction, and counselor protocols to consumer chatbots are leaving the most legally and clinically charged surface area unregulated.

Theme 2. Federal Funding Flexibility, Not Federal Funding Increase. The Fine bill, the LIFT AI Act, and the April 13 Supplemental Priority share a common design feature. They shift existing federal education dollars toward AI rather than adding new dollars. The institutional consequence is that AI literacy is moving from a discretionary instructional choice toward a use-of-funds expectation. Districts without a board-ratified AI literacy scope and sequence will increasingly face friction in federal-program monitoring conversations, not just instructional ones.

Theme 3. Guardrail Design Is Now an Evidence-Based Procurement Variable. The Bastani PNAS finding establishes that the same underlying model (GPT-4) can either help or harm high school learning depending on how the product is configured around it. That moves "guardrails" out of the marketing-claim category and into the procurement-specification category. Districts should expect to be asked by boards and federal monitors alike for evidence supporting the guardrail configuration of any AI tool they have purchased.

Theme 4. "AI-Focused School" Is Becoming a Procurement and Branding Category Before It Becomes an Evidence Category. Houston ISD's rapid expansion is the most visible example of a school-identity category now spreading without a shared definition. State agencies, accrediting bodies, and boards should expect to be asked to recognize, license, or fund "AI-focused" schools well before peer-reviewed evidence of differential outcomes is available. The governance question to ask now is what the comparison-of-outcomes plan looks like, and whether the agency or board approving the designation has the architecture to evaluate it three years from now.

What Was Not Found

Four evidence categories did not appear in this week's window, and the absence of each carries real institutional cost.

First, no peer-reviewed U.S. study published this week measured the equity effects of generative AI use on Title I student populations, on English learner populations, or on students with Individualized Education Programs. The Bastani PNAS field experiment analyzes high school mathematics students in a general sample and does not stratify outcomes by these subgroups in the published draft. As 80 percent of districts report having AI guidelines, the absence of equity-stratified outcome evidence is a present-tense governance risk in any district running AI tools in classrooms serving these populations.

Second, no peer-reviewed efficacy study of consumer-grade companion chatbots among U.S. K-12 students appeared this week, even as Pennsylvania filed the first state executive action of its kind against a major chatbot developer. The Clark JMIR Mental Health paper is the strongest available adjacent evidence. Still, its dataset is small (10 bots, 60 scenarios) and is a simulation study rather than a population-level adolescent outcome study. Districts are governing student use of these products based on vendor descriptions and parent reports, not on large-scale outcome data.

Third, no peer-reviewed study quantified the long-term retention effects of safeguarded AI tutoring (the GPT Tutor configuration) on U.S. K-12 students across a full school year. The Bastani field experiment measured short-window retention. Districts are being asked to make multi-year procurement commitments to AI tutoring vendors with annualized outcome data that does not yet exist.

Fourth, no peer-reviewed evaluation of any state's AI guidance document for districts has appeared. Approximately 35 states have published guidance, yet no peer-reviewed study has examined whether the presence, specificity, or enforceability of guidance is associated with downstream district outcomes. This is the policy-evaluation gap most likely to be filled in the next twelve months, and districts that publish policies now will become part of that future evidence base, whether or not they intended to be.

Novo Executive Summary

The signals this week converge on a single institutional point. The center of gravity for AI governance in public education has moved decisively below the district perimeter, into consumer chatbot exposure, federal use-of-funds expectations, and guardrail-specific procurement. A district AI policy that does not extend to consumer chatbots, that does not embed AI literacy in Title I, II, and IV planning, that does not specify guardrail evidence requirements for AI tutoring vendors, and that does not build proctored or AI-restricted assessment components into the units the Bastani PNAS study identifies as most at risk, is a policy that will look thin against the litigation, funding, and research environment now forming around it.

The window to build that architecture before the 2026-27 school year is short. Boards approving summer procurement, federal program leaders building consolidated applications, and curriculum leaders revising scope-and-sequence documents should be in the same room before August. We support districts directly with this work: AI governance architecture, role-based AI literacy frameworks, procurement and Title-allocation alignment, and an assessment-design strategy that maps to the evolving peer-reviewed evidence base. The strategic question is no longer whether to govern AI. It is about whether the governance architecture is institutional enough to withstand the next 12 months of enforcement and evidence.

Watch This Week

• Pennsylvania v. Character Technologies Inc.: preliminary injunction motion calendar in Commonwealth Court, and any responsive filings from Character.AI on the licensed-professional impersonation count.
• Vermont H.650 educational-technology certification bill: gubernatorial action pending; statute would take effect July 1, 2026.
• K-12 AI Literacy and Readiness Act of 2026: committee referral and bill-number assignment on Congress.gov; companion movement on the Senate LIFT AI Act.
• DOJ-xAI intervention in Colorado AI Act litigation: continued proceedings in the U.S. District Court for the District of Colorado following the April 27 temporary suspension of enforcement.
• Ohio district AI policy adoption deadline: July 1, 2026; expect a final wave of board-ratified policies through June.
• Houston ISD Future 2 program: board agendas, principal selections, and curriculum documents for the nine 2026-27 campuses are likely to be released through June and into early July.

Sources

Governance and Policy

Research, Peer-Reviewed