California Writes the Human Floor Into K-12 Law: AB 2148 Defines the School Employee as a Natural Person

Framing

For two years, the debate over artificial intelligence in schools has been argued in guidance documents, model policies, and district acceptable-use rules. This week the ground shifted. California moved the question from norm to law. On June 30, Governor Newsom signed Assembly Bill 2148, which amends the state Education Code to specify that an elementary or secondary public school employee and a contractor providing services in a public school are natural persons. The bill cleared the Assembly 76-0 and the Senate 38-0. It does not regulate a single classroom tool. It draws a line around a category: the work of a school employee cannot be performed by a system that is not a person.

Read structurally, AB 2148 is the first clear instance of a state writing a human floor into K-12 education code. It joins a wider 2026 movement. California has also advanced Senate Bill 928, which would require California State University instructors to be human, and Illinois has moved a measure barring community colleges from offering courses taught solely by AI. The common thread is not caution about a specific product. It is a decision to legislate the boundary of substitution, to say in statute what AI may not be, even before anyone has measured what AI does well.

That is where this week's research matters, because it points in the same direction as the law. The strongest K-12 evidence available is not a story of transformation. A preregistered randomized trial of an adaptive math tutor found no overall effect on achievement or motivation against business-as-usual instruction. A systematic review of intelligent tutoring systems finds positive but mitigated effects, mostly measured using weaker designs. And the evidence on student over-reliance suggests real cognitive cost, only partly offset by literacy. Lawmakers are drawing a line around the human at the exact moment the evidence says the human still carries most of the instructional weight. For district leaders, the practical question is no longer whether to adopt. It is where, in writing, a human must remain the decision-maker, and how to prove that line is held.

Top Research and Policy Signals

1. California AB 2148 Writes a Human Floor Into K-12 Education Code

Source type. State legislation, enacted (signed into law June 30, 2026).

California State Legislature. (2026). Assembly Bill No. 2148: School personnel (2025-2026 Regular Session). leginfo.legislature.ca.gov

Assembly Bill 2148, authored by Assemblymember Al Muratsuchi, amends the California Education Code so that a public school employee and a contractor providing services in a public school are defined as natural persons. In plain terms, an AI system cannot be the employee or the contracted provider of record for work that the code assigns to school staff. The measure passed both chambers unanimously and was signed on June 30. It does not ban instructional AI, nor does it tell teachers how to use any tool. It sets a categorical limit: the accountable actor in a school role must be a human being.

Leadership implication. Human-in-the-loop is becoming a compliance line, not a values statement. Boards outside California should not wait for a local version of this law. They should define now, in policy and in vendor contracts, which roles and decisions require a named human of record, and document how AI tools are supervised rather than substituted. The districts that write that boundary before they are told to will convert a legal risk into an operating standard.

2. A Preregistered Math-Tutor Trial Finds No Overall Effect

Source type. Peer-reviewed journal article (randomized controlled trial).

Feng, M., et al. (2025). Evaluating the efficacy of an intelligent tutoring system that integrates affective supports into math learning. Journal of Computer Assisted Learning. consensus.app

This preregistered randomized controlled trial tested MathSpring, a web-based intelligent tutor with affective supports, against business-as-usual instruction. The sample was 53 teachers and their 2,003 students, aged 10 to 12, in one United States state, across a full school year. Teachers were randomly assigned to use the tutor or continue ordinary instruction. The result: students who used the tutor showed no evidence of improved achievement, affect, or dispositions toward math compared with the control group. An exploratory analysis found that high-usage students outperformed the control group, which points to dosage rather than a general effect.

Leadership implication. This is the cleanest K-12 efficacy signal in the current window, and it is a null result at the classroom level. Districts should treat vendor efficacy claims as contract terms to be verified, not marketing to be trusted. Pilots should specify a minimum usage threshold, an independent outcome measure, and a fair comparison group before scaling. A tool that only works at high dosage is a professional-development and scheduling problem, not a purchase.

3. The Broader Tutoring Evidence Is Positive but Mitigated

Source type. Peer-reviewed journal article (systematic review).

Letourneau, A., et al. (2025). A systematic review of AI-driven intelligent tutoring systems (ITS) in K-12 education. npj Science of Learning. consensus.app

This review synthesized 28 studies covering 4,597 K-12 students. The overall finding is that intelligent tutoring systems have generally positive effects on learning and performance. Still, those effects are mitigated when the comparison is a non-intelligent tutoring system rather than no support at all. The authors also note that the evidence base relies on quasi-experimental designs with varying, often short, durations, and they call for longer interventions, larger and more diverse samples, and direct study of the ethical implications of using AI to teach.

Leadership implication. Adaptive and personalized are strong marketing words attached to a modest and uneven evidence base. When procurement teams evaluate these systems, the right question is the size of the effect against a fair baseline, not the presence of a dashboard. Ask vendors what their tool was compared against, for how long, and with which students. If the honest answer is a short pilot against nothing, the district is buying ahead of the evidence.

4. Student AI Dependence Is Linked to Weaker Critical Thinking

Source type. Peer-reviewed journal article (survey, moderated-mediation model).

Tian, J., et al. (2025). Learners' AI dependence and critical thinking: The psychological mechanism of fatigue and the social buffering role of AI literacy. Acta Psychologica. sciencedirect.com

Using survey data from 580 university students, this study found that greater AI dependence was associated with lower critical thinking, and that cognitive fatigue partially explained the link. Information and AI literacy buffered the negative effect, but the same literacy amplified fatigue when reliance on AI was already high. The population is higher education, not K-12, and the design is correlational and self-reported, so it establishes association rather than cause. It is the clearest recent signal on why substitution carries a cognitive cost, and why literacy is a partial answer rather than a complete one.

Leadership implication. This is the evidence under the human floor. Role-based AI literacy for students, teachers, and leaders belongs in the governance stack, but literacy alone will not offset heavy substitution. The instructional design that matters most is teaching students when not to use AI, and building assignments that require original reasoning. Districts should pair any literacy rollout with usage norms, because a literate student who offloads everything is still at risk.

Emerging Strategic Themes

Theme 1. The Human Floor Becomes Statute. States are moving from guidance that encourages human oversight to law that defines what AI may not be. AB 2148 writes that boundary into K-12 code, and parallel bills reach higher education. Expect human-in-the-loop language to migrate from mission statements into compliance obligations and contract clauses.

Theme 2. Substitution Versus Augmentation Is the Fault Line. The new laws police substitution. The research shows modest, dosage-dependent augmentation gains. Policy and evidence are converging on the same instruction: keep the human as the decision-maker and treat AI as a support that must earn its place.

Theme 3. Literacy Is Mitigation, Not Immunity. Dependence research shows that literacy reduces cognitive cost without eliminating it and can increase fatigue at high reliance. Role-based literacy is necessary, but it must be paired with usage design that specifies when the tool should be set aside.

Theme 4. The Procurement Gap Is Widening. Vendors sell personalization and adaptivity. The strongest K-12 trial this cycle showed no main effect. The gap between what is marketed and what is measured is now a governance problem that lives in the contract, not the classroom.

What Was Not Found

• No new peer-reviewed causal K-12 study appeared in this week's window. The strongest causal evidence available, the math-tutor trial above, is a null main effect and is not new, yet adoption accelerated this cycle again.
• No causal evidence shows that human-floor laws improve student outcomes or safety. AB 2148 encodes a value, human accountability, not a measured effect. Districts are complying without an evidence base on what the required human must actually do differently.
• No K-12-specific causal evidence connects AI dependence to critical thinking. The available study is in higher education, is correlational, and is not United States-based. We are governing adolescent cognition using adult, self-reported data.
• No disaggregated evidence exists for English learners, students with disabilities, or high-poverty districts on either tutoring effects or dependence risk. Equity claims, both the promise and the harm, remain unmeasured at the subgroup level.
• No outcome data ties human-requirement statutes to instructional quality, cost, or staffing feasibility. The law draws a line. No one has yet measured what it costs a district to hold it.

Novo Executive Summary

The signal this week is a category shift. The human role in schools is shifting from professional norm to statutory requirement, while the strongest evidence indicates that AI's instructional gains are modest and its cognitive risks are real but unmeasured at grade level. That distance, between a legal line and an evidentiary void, is exactly where district liability now lives. The answer is not to adopt faster or slower. It is architecture: written definitions of where a human must decide, procurement that treats efficacy as a contract term rather than a claim, and role-based literacy that teaches when not to use the tool. Districts that build that scaffolding now will comply by design and defend their decisions with evidence rather than enthusiasm. This is the work Dr. Reginald Griffin and Novo Innovative Pathways do with district cabinets: governance architecture, role-based AI literacy, and implementation strategy that turns a moving legal landscape into a standard a board can hold. We don't sell AI. We govern it.

Watch This Week

• California Legislature returns from summer recess in August with roughly 30 AI bills still moving, including AB 2392, which would require a generative-AI procurement and training working group before public colleges deploy such systems.
• Illinois: watch for the governor's signature on the measure barring community colleges from offering courses taught solely by AI.
• Ohio: the first compliance window after the July 1 deadline requiring every district to adopt an AI-use policy. Watch enforcement posture and how many districts adopted the state model versus a local policy.
• New Jersey: A 5184 and S 4469, which would require boards of education to adopt AI policy and direct the Department of Education to publish a model, remain in committee.
• Federal: K-12 AI literacy and readiness bills remain pending, with Congress still in a funding-and-permission posture rather than setting a national standard.
• Publication windows at npj Science of Learning, the Journal of Computer Assisted Learning, and Computers and Education for any late-summer K-12 randomized trials.

Sources

Governance and Policy

Research, Peer-Reviewed

Research, Preprint (Not Peer-Reviewed)

AI in Public Education Brief is published weekly by Novo Innovative Pathways. For district advisory engagements, contact Dr. Reginald Griffin through Novo Innovative Pathways.