New York City Freezes Ed-Tech Purchasing While Illinois Forces AI Audits Into Existence

Framing

The largest school system in the United States stopped buying software this week.

On Monday, July 6, Chancellor Kamar Samuels asked New York City principals to hold off on purchasing educational software until the Education Department finalizes its artificial intelligence guidance later this summer. Programs required for mandated services or school opening are exempt. Everything else waits. A department spokesperson framed it as ensuring that any digital tool used in classrooms is properly assessed for safety and privacy. It is the most concrete action the city has taken since a draft AI policy in March drew roughly 6,500 public comments, packed hearings, and a letter from more than half the City Council urging a moratorium.

The reason for the freeze is more instructive than the freeze. At a City Council hearing on June 24, Education Department officials could not say how many schools were using AI-enabled products or which ones. Purchasing in New York City occurs at the school level and is not centrally tracked. A state comptroller audit in May had already flagged student data privacy gaps in city schools. The department has since sent schools a survey asking what software they run. Read that sequence plainly: the district is pausing procurement because it does not have an inventory, and it does not have a vetting standard to apply if it did. The pause is not a policy. It is a stall while the architecture gets built, and it is happening in July, after principals have already set budgets and planned interventions for the fall.

Now put Illinois next to it. On the same Monday, Governor JB Pritzker signed Senate Bill 315, the first state law in the country requiring independent third-party safety audits of the largest AI systems, along with public disclosure of safety practices and reporting of significant safety incidents. It says nothing about schools. It regulates the model developers whose systems sit underneath the products districts buy, and it forces an external record into existence where none was required before. One city froze procurement because it had nothing to vet against. One state began manufacturing exactly the material a district would vet with. Both events are procurement events, and that is the argument of this brief: AI governance in public education is not a values statement, a task force, or a guidance document. It is a purchase order, an inventory, and a standard. New York City just proved what happens to a district that has the third without the first two, and it proved it at scale, in public, in July.

Top Research and Policy Signals

1. New York City Freezes Ed-Tech Software Purchases Because It Cannot Say What Its Schools Are Running

Source type. District administrative action, reported by regional news outlet. Chancellor's email dated Monday, July 6, 2026; reported July 8, 2026.

Elsen-Rooney, M. (2026, July 8). Kamar Samuels asks NYC schools to pause software purchases until AI guidance is final. Chalkbeat New York. chalkbeat.org

Chancellor Kamar Samuels emailed principals on Monday, July 6, asking them to hold off on educational software purchases until the Education Department releases final AI guidance later this summer. Software needed for mandated services or school opening is exempt. Because New York City schools typically resubmit purchase orders annually, even long-standing programs, including tools used for grades, attendance, and academic intervention, are potentially caught in the freeze. Principals told Chalkbeat that budgets, interventions, and program decisions for next year were already built, and that changing direction in July creates real uncertainty about whether planned tools will be available.

The context is a governance failure, not a technology failure. The department's March draft AI policy used a traffic-light framework that barred AI for assessment and grading while green-lighting lesson-plan brainstorming, and it largely omitted student AI use. It drew roughly 6,500 comments and sustained public opposition. Samuels said in May that the draft had missed the mark. Final guidance promised for June was delayed at a June 24 Council hearing. At that hearing, officials could not tell the Council how many schools use AI products or which ones, because school-level purchasing is not tracked centrally, and a May state comptroller audit had already flagged student data privacy gaps. A survey asking schools what they use went out only afterward.

Leadership implication. Every superintendent should ask two questions this week and should not accept a soft answer to either. First: can we produce a complete list today of every software product in use across our schools, including any purchased at the building level with discretionary funds? Second, if we can, what standard did each of those products meet before they touched a student? If the honest answer to the first is no, the district is one Council hearing, one audit, or one incident away from New York City's position, and a purchasing freeze in July is a far more expensive instrument than an inventory in September.

2. Illinois Enacts the First State AI Law Requiring Independent Third-Party Safety Audits

Source type. State legislation, enacted. Signed July 6, 2026. Effective January 1, 2027. Not education-specific.

Office of Governor JB Pritzker. (2026, July 6). Gov. Pritzker signs nation-leading artificial intelligence safety law [Press release]. gov-pritzker-newsroom.prezly.com

Senate Bill 315, the Artificial Intelligence Safety Measures Act, passed with bipartisan support and was signed on July 6. It requires developers of the largest advanced AI systems to identify, disclose, and mitigate risks; publicly disclose safety practices; report significant safety incidents; and maintain compliance processes. It creates confidential reporting channels and whistleblower protections. Illinois is the first state to require regular independent third-party safety audits of covered systems by qualified experts without financial conflicts of interest. The state Attorney General framed the law as filling a gap left by the absence of federal oversight. It takes effect January 1, 2027.

For districts, the law is an upstream accountability instrument. It will not tell a principal whether a reading program works. What it does is create a documentary trail that did not previously have to exist: disclosed safety practices, reported incidents, and independent audit results tied to the frontier models underneath a large share of instructional AI products. New York City's stated reason for pausing was that tools must be properly assessed for safety and privacy. Illinois just started producing the raw material for that assessment.

Leadership implication. This is the procurement clause to write now, and it costs nothing to adopt. Require any AI vendor to disclose, on request, the safety documentation, incident reports, and third-party audit results of the underlying model or models its product relies on, and to notify the district of material safety incidents. Districts outside Illinois can hold vendors to the standard voluntarily, because another state is already creating the vendor's compliance burden for them.

3. Institutional AI Readiness Reaches Students Through Teacher Capability, Not Teacher Enthusiasm

Source type. Preprint. Not peer-reviewed. arXiv, submitted March 20, 2026. Cross-sectional; associational, not causal.

Guan, X., Zheng, M., Gasevic, D., Guo, W., Liu, Y., Han, X., Gasevic, D., Ma, R., Wu, Q., & Yan, L. (2026). From school AI readiness to student AI literacy: A national multilevel mediation analysis of institutional capacity and teacher capability [Preprint]. arXiv:2603.20056. arxiv.org

Using linked national survey data from 1,007 vocational institutions, 156,125 teachers, and 2,379,546 students across all 31 provincial-level regions of China, the authors test whether school-level AI readiness is associated with student AI literacy, and through what mechanism. It is, and the mechanism is specific. Aggregated teacher-perceived AI capability partially mediates the relationship between institutional readiness and student AI literacy. General attitudinal acceptance, meaning how positively teachers feel about AI, shows no stable transmission effect. The pathway held across regions at very different levels of AI development and under alternative model specifications. The authors conclude that infrastructure investment must be aligned with sustained professional capacity development.

Two limits belong in the room. The setting is Chinese vocational education, not United States K-12, and the design is cross-sectional, so these are associations, not causal effects. What survives translation is the structural claim, and it is consistent with what the RAND panels have shown domestically: readiness that is not converted into teacher capability does not reach students. It also reframes what New York City is pausing. Freezing purchases stops the spending. It does not build the capability that would have made the spending worth anything.

Leadership implication. This is the empirical case against the device-and-license approach to AI adoption. Any AI investment brought to a board should have to name the teacher capability it builds, how that capability will be measured, and who owns it. Adoption rates, license utilization, and enthusiasm surveys are attitudinal proxies, and the largest available study finds they do not translate to students.

4. Language Models Give Eighth Graders Different Writing Feedback Based on Who They Are Told the Student Is

Source type. Preprint. Not peer-reviewed. arXiv, March 2026. Stanford University.

Tan, M., Phalen, L., & Demszky, D. (2026). Marked pedagogies: Examining linguistic biases in personalized automated writing feedback [Preprint]. arXiv:2603.12471. arxiv.org

The researchers took 600 eighth-grade persuasive essays from the PERSUADE dataset and asked four widely used language models, GPT-4o, GPT-3.5-turbo, Llama-3.3 70B, and Llama-3.1 8B, to generate feedback. The essays were held constant. What changed was a prompt describing the student: gender, race or ethnicity, learning needs, achievement level, motivation. Feedback shifted in systematic, stereotype-aligned ways. Feedback tied to students described as high-achieving skewed toward argumentation, reasoning, and development of ideas. Feedback tied to students described as English learners or as Hispanic skewed toward grammar, spelling, and formality. For students described as low-achieving or as having a learning disability, models tended toward positive reinforcement with limited critical content, a pattern the authors characterize as withholding.

This is not a factual error a content review would catch. The model is not wrong about the essay. It is making a pedagogical judgment about what this student is capable of receiving, from a label. That is a judgment a school system is legally and professionally obligated to make deliberately, and it is being made silently, at scale, inside a feedback tool. It is also the concrete answer to what safety and privacy vetting is supposed to catch, and neither a privacy review nor a data-protection addendum would have caught it.

Leadership implication. Any district using AI for writing feedback, formative comments, or rubric scoring should assume differential treatment until it has evidence to the contrary. Require vendors to state whether student characteristics, including special education status, English learner status, or achievement data, are passed to the model, and prohibit that transfer unless the district has explicitly approved it. Then run the test locally: identical student work, varied labels, compare the feedback. It takes an afternoon, and it is now a defensible standard of care.

5. A Peer-Reviewed Scoping Review Names the Risk K-12 Adoption Keeps Skipping: Developmental Harm

Source type. Peer-reviewed journal article, scoping review. Computers and Education: Artificial Intelligence, 2026. Author names could not be confirmed from full-text access and are therefore not asserted.

[Authors unconfirmed]. (2026). Potential risks of generative artificial intelligence integration into K-12 education: A scoping review. Computers and Education: Artificial Intelligence. sciencedirect.com

The review synthesizes 22 empirical K-12 studies to map which risks the research literature actually reports, which mitigations have been proposed, and where the gaps lie. Its central argument is that prior reviews have largely overlooked developmental risk to children and adolescents specifically. Offloading cognitive work to a chatbot is associated with metacognitive passivity and disengagement, and generative tools can erode voice and authorship in adolescents at exactly the stage when those capacities are forming. The mitigations the authors compile are pedagogical, not technical: use the tool to generate hints, examples, and counterarguments while requiring students to show their reasoning; shift assessment toward process, including in-class explanations; and embed generative AI within AI literacy and critical digital citizenship instruction. Safe integration, the authors argue, requires shifting focus from technical adoption to ecological protection.

This is the peer-reviewed anchor of the week, and it exposes the limit of what New York City's pause can accomplish. A procurement freeze governs what a district buys. It does not govern what a student stops doing for themselves.

Leadership implication. Assessment design is an AI governance instrument, and most districts have not treated it as one. Curriculum and assessment leaders, not just technology directors, belong on the AI governance body. If a district's AI policy addresses tools and devices but never defines what counts as evidence of student thinking, it has governed the vendor and left the classroom unprotected.

Emerging Strategic Themes

Theme 1. You Cannot Govern What You Cannot Enumerate. New York City's freeze traces to a single missing artifact: a list of what software its schools actually run. Building-level purchasing, discretionary funds, and free tools adopted by individual teachers mean most districts are in the same position and have not been asked yet. The inventory is not an IT chore. It is the precondition for every other governance step, and it is the cheapest.

Theme 2. The Audit Trail Moves Upstream. Illinois did not regulate schools. It regulated the developers whose models sit underneath the products schools buy, and it forced independent verification into existence. A district's leverage in procurement just increased without any district doing anything. The documentation will exist. Districts that ask for it will get it. Districts that do not ask will continue to sign on vendor assurance.

Theme 3. Personalization Is Where Bias Hides. Bias in AI has been discussed as a data problem. The Stanford finding reframes it as a pedagogical one. When a model personalizes, it must infer what the learner needs, and inference from a label is precisely how differential expectations get encoded. Every district that has celebrated personalization as the core promise of instructional AI now has to govern personalization as its core risk.

Theme 4. Deadlines Produce Documents; Architecture Produces Protection. Ohio's district AI policy deadline passed July 1. Maryland districts are on a 120-day clock from the release of state guidance, which requires structured vendor vetting, continuous review of approved tools, and educator professional development. New York City has a draft policy, 6,500 comments, and no inventory. The compliance artifact and the governance capability are two different projects, and only one of them protects a student.

What Was Not Found

The most important thing not found this week was found by New York City, not by a researcher: the nation's largest district could not produce a list of the AI products in its own classrooms. That is not a research gap. That is an operating gap, and it is almost certainly not unique.

On evidence, nothing closed. There is still no causal study showing that any state AI mandate, model policy, or district AI plan improves any student outcome. Ohio, Maryland, Illinois, California, Connecticut, and Idaho have produced deadlines, model policies, guidance, and curriculum requirements. Not one has been evaluated against student learning, teacher workload, or equity outcomes. New York City is about to run the largest natural experiment in the country: a midsummer procurement freeze in a system of roughly 900,000 students, and no one has announced a plan to measure its effects on instruction.

There is no K-12 AI audit standard. Illinois now requires third-party safety audits of frontier developers. There is no equivalent requirement or accepted methodology for auditing an instructional AI product against the population a district actually serves. A superintendent who wanted to demand an audit today could not say what the audit should test. New York City has paused purchasing and has not published the standard it intends to vet against.

The subgroup evidence remains one-directional. This week produced strong evidence that models treat English learners and students with disabilities differently in written feedback. It produced no evidence about whether those students learn more or less when the tools are used. We now know more about how AI systems behave toward vulnerable students than about what happens to those students. The harm case is documented. The benefit case is asserted.

And the large mediation study is a Chinese vocational education study, cross-sectional and associational. There is no United States K-12 equivalent linking district readiness, teacher capability, and student AI literacy. Elementary grades and non-STEM subjects are again absent; the strongest work in this window is eighth-grade persuasive writing and vocational institutions, while adoption proceeds in early literacy, elementary mathematics, and the arts regardless.

Novo Executive Summary

New York City stopped buying software this week because it could not say what it had already bought, and Illinois began forcing an audit trail to be created for the models underlying those purchases. Both are procurement events, and together they make the case this brief has been making all year: AI governance in public education lives in the purchase order, the inventory, and the vetting standard, not in the guidance document. The research reinforces it from three directions. Institutional readiness reaches students only through teacher capability, not enthusiasm. Personalization encodes differential expectations that a privacy review will never catch. And the developmental risks identified in the peer-reviewed literature show up in what students stop doing for themselves, which no procurement freeze can address. A district that treats its AI policy as the deliverable has produced a document. A district that treats the policy as the first artifact of an architecture, one that names decision rights, a tool inventory, a vetting standard, role-based capability requirements, and a review cycle with an owner, has produced protection. Dr. Reginald Griffin works with district leadership teams through Novo Innovative Pathways to build that architecture: governance structures, role-based AI literacy, and an implementation strategy that can survive an audit, a City Council hearing, and a school board meeting.

Watch This Week

• New York City's final AI guidance. Officials have indicated a release this summer, with reporting pointing to September. Check whether it includes a vetting standard and an approved tools list, or only usage rules.
• The NYC school software survey. The Education Department asked schools what products they run. If the results are made public, it will be the first comprehensive inventory of AI penetration in a major district.
• Whether the roughly 6,500 public comments on New York City's March draft are released, as officials told the City Council they would be.
• Illinois State Board of Education K-12 AI guidance, statutorily due July 1, 2026. Confirm publication and whether it addresses vendor documentation.
• Illinois SB 315 rulemaking ahead of the January 1, 2027 effective date, particularly the scope of covered systems and the audit standard, and whether other states copy the audit provision.
• Maryland district board calendars through the fall as the 120-day clock runs, and whether any district publishes its vendor-vetting criteria.

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.