Utah Tested What 100 School Apps Actually Transmit and Most Contracts Were Wrong

Framing

Edition 30 of this brief argued that the federal lever on AI in schools is procurement rather than prohibition, and that the Department of Education had told districts what to require before paying rather than what AI may do. That argument holds. This week supplies the missing part. Procurement is a lever only if the procurement paper describes what the product actually does. The most important document to cross this desk in the last seven days is not a bill. It is a state education agency report showing that for a majority of tested apps under contract, the paper and the product disagreed.

Read the California instrument, not just its content. AB 2392 does not tell any campus which generative AI system to buy or ban. It orders a working group, a written procurement standard, mandatory training tied to the product, an annual review of that training, and a record of who completed it. That is an accountability architecture, and it travels. It travels because it is cheap. After all, it creates a document a legislator can ask to see, and because higher education has become the drafting room where states test language they are not yet ready to impose on 900 school districts at once. Read AB 2392 alongside SB 928, which the Governor signed on August 27 and which states that a California State University instructor of record must be a person. Both bills answer the same question: who is accountable when the system is in the room. Three days before that signature, and 2,000 miles east of it, the Fulton County Board of Education in Georgia carried a policy stating that AI output shall be limited to suggestions and recommendations and that human beings must make final decisions. A legislature and a school board reached the same rule in the same week without coordinating. When a rule arrives independently from two directions, it is not a trend. It is a settlement, and the drafting is effectively over.

The research counterpoint is uncomfortable because it is not about AI at all. It is about whether districts can verify anything they sign. The Utah investigation tested what 100 apps transmitted rather than what their privacy policies promised, and found that 61 percent shared data with third parties and 36 percent shared with advertisers, with three individual apps sharing with 32, 54, and 33 advertising entities. The most frequently shared element with advertisers was a unique user identifier that appears in no signed agreement as a permitted collection. The report's own authors state plainly that teachers and local privacy managers cannot be expected to run network traffic tests. If a district cannot verify a data privacy agreement, which is the simplest contract in the edtech stack, it has no realistic path to verifying a generative AI clause about model training, retention, or human review.

The timing argument is a calendar argument. Districts are in the first two weeks of the school year. Renewal decisions on learning platforms, monitoring tools, and AI assistants land between now and the winter board cycle. California's 30-day signing window closes at the end of September, and the procurement language in AB 2392 will be public and copyable the moment it is signed. A cabinet that adds a verification clause and an audit right to its next three renewals is doing quiet work in September. A cabinet that discovers the gap during a records request or a breach notification is doing expensive work in February.

Top Research and Policy Signals

1. California Orders Its Public Colleges to Write Generative AI Procurement Standards Before They Deploy

Source type. Pending legislation, California Legislature, passed August 26, 2026, awaiting the Governor's action.

Fong, M. (2026). Assembly Bill 2392: Public postsecondary education: generative artificial intelligence. California Legislature, 2025-2026 Regular Session. digitaldemocracy.org

AB 2392 requires the California Community Colleges and the California State University to convene a joint working group before providing a generative AI product to students, faculty, or staff. It requests that the University of California do the same. The working group must present recommendations for procurement standards and training, submitted on or before January 1, 2028. The bill further requires the systems to provide training to students, faculty, or staff on the use of the product, to review that training at least once per academic year, to update it as necessary, and to maintain records of completed training for each student, faculty member, and staff member who receives one. The Assembly Higher Education Committee analysis dated April 7, 2026, records that California State University campuses had adopted AI tools without consistent guidance or training, raising concerns around data privacy, academic integrity, and equitable use. The bill passed the Assembly 77 to 0 on May 26, 2026, passed the Senate 38 to 0 on August 26, 2026, and received Assembly concurrence 79 to 0 before being sent to the Governor.

Leadership implication. Assign your chief technology officer and your general counsel to pull the AB 2392 text the day it is signed or vetoed and to extract three elements into your own renewal template: the pre-deployment standard, the role-based training requirement tied to the specific product, and the completion record. Those three elements are what an auditor can check. Put them on the agenda of the cabinet meeting that precedes your next platform renewal, not on a summer planning retreat.

2. A Georgia District Wrote the Human Decision Rule Into Board Policy Three Days Before California Wrote It Into Statute

Source type. District board policy, carried as an adoption action item on a consent agenda, Fulton County Board of Education, August 20, 2026.

Fulton County Board of Education. (2026, August 20). Board Policy IFBI: Artificial intelligence (adoption, action). Consent agenda item 5.3, Board Meeting of August 20, 2026. Fulton County Schools. simbli.eboardsolutions.com

Policy IFBI defines artificial intelligence as a machine-based system or tool that can generate, classify, predict, recommend, summarize, transcribe, analyze, or otherwise assist with content, communications, decisions, or tasks based on user inputs or available data, and applies to all persons who access district systems, technology resources, and data. Four provisions carry weight well beyond Georgia. Only district-approved AI tools may be used, and approval is contingent on a review confirming that the tool meets district data security requirements and that access is through district accounts. All instructional software and web-based applications must be reviewed and recommended by the District Technology Governance Committee. Before any use, all associated contracts must be thoroughly reviewed and a formal contract agreement must be executed. Users must not use AI tools to automate decision-making without human oversight. The policy states that any output from artificial intelligence or machine learning shall be limited to suggestions and recommendations, and that final decisions must be made by human beings, applying appropriate review, nuance, expertise, and context. And the policy acknowledges in its own text that many already-approved instructional, administrative, and productivity tools incorporate embedded artificial intelligence. It further treats AI recording and transcription as subject to the same consent rules as any other recording device, bars reliance on AI in ways that disadvantage students or employees based on disability, language, race, national origin, sex, religion, or other protected status, and commits the district to monitor usage and provide training.

Leadership implication. Copy the structure, not the wording. Fulton names a standing body that reviews tools, requires an executed contract before any use, writes the human-decision rule as a hard limit on what AI output may be rather than as a value statement, and concedes on the record that AI is already embedded in tools it approved before anyone called them AI. Ask your board attorney whether your own policy contains a named body, a contract precondition, and an output limit. If any of the three is missing, you have an acceptable use policy rather than a governance policy. Then ask the question that Fulton's own text raises but does not answer: who funds the monitoring the policy promises?

3. Utah Tested What 100 School Apps Actually Transmit and Found Most Contracts Overstated the Control

Source type. Institutional report, not peer-reviewed. Utah State Board of Education, dated August 20, 2025, in wide public circulation the week of August 24, 2026.

Keith, M., Giboney, J., LeVasseur, L., & Simpson, B. (2025). Utah EdTech app data collection and sharing: 2023-25 investigation. Utah State Board of Education, Brigham Young University, and Internet Safety Labs. schools.utah.gov

Researchers identified more than 3,000 unique education apps in use across Utah, selected 100 for testing, and captured live network traffic rather than reading privacy policies. In 44 of the 85 apps with signed agreements registered through the Student Data Privacy Consortium, or 52 percent, the app collected at least one data element the agreement did not permit. Eleven apps, or 13 percent, shared at least one element not named in their agreements with third parties. Across all apps tested, 61 percent shared data elements with third parties and 36 percent shared with advertisers; three apps shared with 32, 54, and 33 advertising entities, respectively. The element most frequently shared with advertising companies was a unique user identifier that appears in no signed agreement as a permitted collection, and which is expressly named as personal information under the Children's Online Privacy Protection Act at 16 CFR 312.2. Of 36 vendors contacted, 20 said advertising is absent from their student-facing or education-specific offerings, but the researchers could not verify which version districts were using. Vendor reconciliation added ten months to the project. Following the investigation, Utah enacted H.B. 55, Privacy Compliance for Education Technology Vendors, amending Utah Code 53E-9-309, effective July 1, 2026. The bill requires contract terms that protect student data, notice to vendors of unauthorized use, and termination if a confirmed violation is not remedied.

Leadership implication. Direct your data privacy manager to produce, within thirty days, a one-page count of how many approved applications your district has, how many have signed data privacy agreements, and how many of those agreements name a unique user identifier as a permitted collection. Then add a right-to-audit and a technical-verification clause to your standard renewal language. Utah's own report says teachers cannot run these tests. That argument supports pooling capability at the district, regional service agency, or state level, and it is a budget line your board can approve this fall.

4. State AI Guidance Is Split Evenly Between Caution and Opportunity, and Caution Is the More Brittle Choice

Source type. Peer-reviewed journal article, Educational Policy, first published online July 18, 2026.

McCune, J., Zarch, R., Childs, J., Yadav, A., Marshall, S. L., & Grooms, A. (2026). The fragile edge of innovation: Examining policy resilience in state-level K-12 artificial intelligence guidance. Educational Policy. Advance online publication. doi.org

The authors conducted a qualitative document analysis of AI education guidance from 35 U.S. states that had publicly published such resources between January 2024 and December 2025, coding 994 pages of documents and 4 state web pages, with individual documents ranging from 2 to 81 pages. A second researcher independently coded a 60 percent subsample. Using the CAPE framework of Capacity, Access, Participation, and Experience, the team classified each state's dominant policy posture. The two dominant postures came out even: 13 states wrote guidance dominated by caution, 13 wrote guidance dominated by opportunity, and 9 struck a balance. Caution-dominant guidance emphasized restrictive ethical rules, data privacy mandates, and prohibitions of misuse. The authors argue that this posture narrowly defines institutional capacity as compliance and produces what they term policy fragility, in which the pace of reform exceeds the infrastructure needed to sustain it. They further find a rhetorical-implementation gap on access, where states issue broad equity statements and descriptive tasks rather than prescriptive strategies, shifting the burden of equitable integration onto districts. They name teacher capacity as the primary structural equity issue and warn of a dual-system risk in which only well-resourced districts deliver conceptually grounded AI learning.

Leadership implication. Pull your own state's AI guidance and classify it yourself against these three categories before your next board policy reading. If it is caution-dominant, your district is inheriting a compliance checklist and no instructional model, and the gap is yours to close with local dollars. Name the person who owns AI professional learning, distinguish training on tools from instruction about AI, and put both lines in the budget you are building now rather than waiting for the state to fund what its own guidance did not specify.

5. Vendor Self-Disclosure Documents Cannot Carry the Weight Procurement Puts on Them

Source type. Peer-reviewed conference proceedings, ACM Conference on Fairness, Accountability, and Transparency 2026.

Kuehnert, B., Johnson, N., Dotan, R., & Heidari, H. (2026). Disclosure or marketing? Analyzing the efficacy of vendor self-reports for vetting public-sector AI. In Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency. doi.org

The authors studied the GovAI Coalition FactSheet, a widely adopted transparency document that supports AI procurement and governance in government. They combined semi-structured interviews with vendors and public-sector practitioners with a systematic analysis of completed FactSheets. They find that FactSheets are asked to serve multiple and conflicting purposes at once: showcasing vendor offerings, supporting evaluation and due diligence, and opening early-stage dialogue between vendors and agencies. Those competing expectations, combined with the structural constraints of voluntary and public self-disclosure, limit the ability of FactSheets to function as standalone evaluation or risk-assessment tools. The authors do not conclude that the documents are worthless. They conclude that the artifacts work when treated as relational instruments that establish shared understanding and sustain dialogue over time, and that they fail when treated as a substitute for independent evaluation.

Study context. Qualitative, interview-based, and focused on government agencies generally rather than school districts specifically. It does not quantify how often disclosure documents are inaccurate. It establishes what practitioners report the documents can and cannot do.

Leadership implication. Stop treating a completed vendor questionnaire as the vetting step. Route it to the person who will ask follow-up questions, and include at least one requirement in your evaluation rubric that the vendor cannot satisfy by self-report, such as a named subprocessor list, a third-party security attestation, or a right to technical testing. Give your procurement team explicit permission to score an unanswered follow-up as a deficiency rather than a formality.

Emerging Strategic Themes

Theme 1. The contract is not the control. Districts have spent three years building AI governance on documents: signed agreements, vendor questionnaires, approved tool lists, and acceptable use policies. Utah measured the gap between one of those documents and the product it governs and found that the gap is the norm, not the exception. Treat every unverified document in your stack as an assertion about the world rather than a constraint on it, and rank them by how much harm an inaccurate one would cause.

Theme 2. Verification is the next procurement specification. The instruments now in play share a common shape: a written pre-deployment standard, a named reviewing body, a contract executed before use, a training record, an audit right, and a termination trigger. Utah's H.B. 55, California's AB 2392, and Fulton County's Policy IFBI were written by two legislatures and one school board for three different sectors and converged on the same architecture. Expect right-to-audit and technical-testing clauses in state model contracts within two legislative sessions. Districts that add them voluntarily this fall will not renegotiate under the deadline later.

Theme 3. Guidance written as compliance produces brittle districts. Thirteen states have adopted a cautionary posture, and peer-reviewed findings indicate that this choice defines capacity as compliance, leaving systems unable to absorb the next change. A district in a caution-dominant state should not read that guidance as complete. It should read it as the floor and the budget, separately, for the instructional capacity the state did not describe.

Theme 4. The human decision rule is no longer being debated, only drafted. California tested instructional personhood on the state university with SB 928 and is testing procurement standards on its colleges with AB 2392, and neither binds a school district. But Fulton County reached the same limit on its own three days earlier, without a state mandate and without waiting for one. The rule is arriving from the statehouse downward and from the board table upward at the same time. Track both. A superintendent who watches only K-12 bill numbers will encounter this language only after it is settled, and a superintendent who watches only the legislature will miss that peer districts have already written it.

What Was Not Found

This section exists this week because the window produced a strong measure of vendor behavior but no measure at all of whether governance changes that behavior.

No peer-reviewed K-12 AI outcome study published in the fourteen-day window met the selection standard. Searches of academic indexes for the period August 17 through August 30, 2026 returned meta-analyses and reviews published earlier in the year, several of which were strong, and no new peer-reviewed K-12 outcome study dated inside the window. Every research item in this brief carries a publication date outside the fourteen-day window, and each one states that date in its source line.

No study measures whether an audit right, a review committee, or a contract term changes vendor behavior. Utah demonstrated that agreements are breached. Utah did not test whether districts with stronger contract language, named review bodies, or termination triggers experience fewer breaches. Fulton County's new policy commits the district to monitoring usage, which is the right commitment and the only clause in the document without a published method, a named owner, or a dollar figure attached. This brief recommends audit clauses based on structural reasoning rather than outcome evidence, and that this limitation be stated to your board when you propose them.

No equivalent network-traffic audit exists for any state other than Utah. Searches for comparable state education agency technical investigations returned none. Forty-nine states are operating approved-tool lists that have never been tested against live traffic. The absence is not evidence that other states are cleaner. It is evidence that nobody looked.

No data separates English learners, students with disabilities, or high-poverty districts in any of this week's sources. The Utah investigation reports app behavior, not student subgroup. The Educational Policy analysis reports what states wrote about equity, not what any student experienced. Districts are making consent-default and approved-tool decisions that fall hardest on families least able to contest them, with no subgroup evidence in the record.

No evidence addresses elementary literacy or non-STEM subjects. Nothing located in this window examines AI use in early reading, writing instruction, social studies, or the arts. The instrument states are subject-neutral. The evidence base is not, and it is thinnest exactly where the youngest students are.

No governance study tests policy efficacy. The Educational Policy analysis classifies state guidance and theorizes about resilience. It does not measure whether caution-dominant or opportunity-dominant states produced better district implementation, better teacher preparation, or better student outcomes. The framework is useful. It is not yet an evidence claim.

The pattern is consistent with what this brief has reported since spring: mandates and instruments are outrunning evidence, and the gap is widening rather than closing. The correct response is not delay. It is monitoring, written documentation of what you decided and why, and contractual exit ramps that let you reverse a decision without litigation.

Novo Executive Summary

This week, California moved a generative AI procurement standard to the Governor's desk; a Georgia school board wrote the human decision rule into its own policy without waiting for a state to require it; and a state education agency published evidence that the procurement documents districts already hold do not describe what their products do. Those three facts belong in the same paragraph, because the third one determines whether the first two matter. A procurement standard is only as strong as a district's ability to verify compliance with it, and the peer-reviewed record now shows that vendor self-disclosure cannot supply that verification and that most state guidance defines capacity as compliance rather than building it. The differentiator for districts is not which AI tool they select. It is whether they have decision rights written down, procurement specifications that require something a vendor cannot self-assert, an evaluation cadence that survives leadership changes, and literacy pathways that are role-based rather than generic. That architecture is what Novo Innovative Pathways builds with district leaders, and it is built before the renewal, not after the incident.

Watch This Week

• Monday, August 31. The California Legislature adjourns sine die. Governor Newsom then holds a 30-day window to sign or veto. AB 2392 and AB 2656 are both on his desk.
• Resolved and carried forward from Edition 30. Governor Newsom signed SB 928 on August 27, 2026, making it law that a California State University instructor of record must be a person. Watch for the first K-12 bill in any state that copies the definitional language.
• Carried forward from Editions 29 and 30. AB 2656, requiring public employers to give a recognized employee organization 45 days' written notice before deploying generative AI in represented work, passed the Senate 39 to 0 on August 24 and received Assembly concurrence 74 to 2 on August 25. It is with the Governor. School districts are local public employers.
• Carried forward from Editions 27 through 30. AB 1159, extending California student privacy protections to digital operators marketing to schools, cleared Senate Appropriations 5 to 0 on August 13 and was ordered to third reading. It requires a floor vote before adjournment Monday.
• Carried forward from Editions 27 through 30. Florida's proposed AI amendment to Internet Safety Policy Rule 6A-1.0957 has still produced no formal Notice of Proposed Rule following the August 5, 2026 development workshop. The draft district compliance date remains January 1, 2027.
• Fulton County Schools, Georgia. The minutes of the August 20, 2026 board meeting are scheduled to appear on a subsequent consent agenda and will confirm whether Policy IFBI was adopted. Watch also for the implementing regulation and for any published charge to the District Technology Governance Committee, because the policy assigns that body the tool review and contract execution work without stating its membership, cadence, or budget.
• Still unresolved and carried forward. New York Senator Andrew Gounardes introduced the FOCUS Act on August 21, 2026, which would impose age-based restrictions on screens, AI, and edtech in New York public schools; it falls outside this edition's governance window and awaits committee action. Also outstanding: the GAO study of AI in K-12 requested by Senate letter in June; North Carolina H 301 in conference committee; New York S 9051 and A 6578 awaiting the Governor's action by December 31; New York City's promised AI playbook; Michigan SB 760 with the House Communications Committee following the Legislature's August 25 return; and the October publication of Nagashima and colleagues' CSCW study of teacher and student misalignment over classroom AI control, first flagged in Edition 26.

Sources

Governance and Policy

Fong, M. (2026). Assembly Bill 2392: Public postsecondary education: generative artificial intelligence. California Legislature, 2025-2026 Regular Session. digitaldemocracy.org

Fulton County Board of Education. (2026, August 20). Board Policy IFBI: Artificial intelligence (adoption, action). Consent agenda item 5.3, Board Meeting of August 20, 2026. Fulton County Schools. simbli.eboardsolutions.com

Office of Governor Gavin Newsom. (2026, August 27). Governor Newsom signs legislation 8.27.2026. State of California. gov.ca.gov

Transparency Coalition. (2026, August 28). AI legislative update: August 28, 2026. transparencycoalition.ai

Utah State Legislature. (2026). H.B. 55, Privacy compliance for education technology vendors. le.utah.gov

Research, Peer-Reviewed

Kuehnert, B., Johnson, N., Dotan, R., & Heidari, H. (2026). Disclosure or marketing? Analyzing the efficacy of vendor self-reports for vetting public-sector AI. In Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency. doi.org

McCune, J., Zarch, R., Childs, J., Yadav, A., Marshall, S. L., & Grooms, A. (2026). The fragile edge of innovation: Examining policy resilience in state-level K-12 artificial intelligence guidance. Educational Policy. Advance online publication. doi.org

Institutional Report, Not Peer-Reviewed

Keith, M., Giboney, J., LeVasseur, L., & Simpson, B. (2025). Utah EdTech app data collection and sharing: 2023-25 investigation. Utah State Board of Education, Brigham Young University, and Internet Safety Labs. [Verified at source, August 30, 2026: authors, institutions, the August 20, 2025 date, and every figure cited above appear in the report text.] schools.utah.gov

Arntz, P. (2026, August 26). Popular school apps may be sharing student data with advertisers. Malwarebytes Labs. malwarebytes.com