Washington Tells Districts to Buy EdTech on Evidence of Learning, and Its Own Clearinghouse Holds None for AI: Procurement Becomes the Federal Lever

Framing

This brief has spent 2026 tracking where AI governance attaches. Statutes told districts to write policies. State agencies issued model language. Boards adopted acceptable-use rules. In July, organized labor claimed a bargaining seat. Florida loaded AI into the internet safety policy districts already re-adopt every year. Two weeks ago, California AB 2656 moved the attachment point to the purchase order, requiring 45 days of union notice before a public employer deploys generative AI in represented work. Last week the attachment point moved again, in two directions at once, and the second one is federal.

The Department of Education did not issue a rule. It issued a letter, and the instrument matters as much as the content. A Dear Colleague Letter creates no obligation, carries no deadline, and cannot be enforced. What it does is establish a documented federal expectation, and expectations of that kind do their work later, in the places where a district has to explain itself. The letter tells states and districts to evaluate educational technology based on demonstrated learning outcomes rather than screen time alone, to incorporate evidence of effectiveness into procurement and renewal decisions, to seek products that have conducted randomized controlled trials, and to draw on the ESEA evidence framework already governing federal education spending. It asks providers to publish rigorous independent evaluations and to disclose product limitations. Read plainly, it moves the burden of proof from the district that questions a tool to the vendor that sells one.

The counterpoint arrived ten days earlier from inside the same Department, and the two documents have not been read against each other anywhere we could find. On August 10, the Institute of Education Sciences reported that a search of the What Works Clearinghouse, which the Department itself calls the most trusted repository of scientific evidence for education, identified no research studies on AI and student outcomes. The strongest available review located only 20 rigorous causal studies in the entire field. So the federal instruction is to buy AI on evidence of effectiveness, and the federal evidence repository holds none for AI. That is not a contradiction a district can resolve. It is a condition a district has to document. Two peer-reviewed papers published the same week explain part of why the shelf is empty: the field has been measuring competence with AI as technological fluency rather than pedagogical judgment, and one of those papers reports that its own validation could not establish that its four dimensions were distinct.

The calendar makes this the week to act rather than watch. Districts open the 2026-27 school year in the next two weeks, and most will renew AI and edtech contracts between now and October. California adjourns sine die at the end of business on Monday, August 31, with 24 AI-related bills near final passage. The federal letter is guidance today. The five questions it poses, in the Department's own words, are the questions a board member will read aloud in November and a plaintiff's counsel will ask in a deposition later than that. A district that can answer them for each major product this fall is doing cheap work. A district that answers them for the first time under examination is doing expensive work.

Top Research and Policy Signals

1. The Education Department Tells Districts to Buy EdTech on Evidence of Learning, and Asks Vendors to Produce Randomized Trials

Source type. Federal agency guidance. Dear Colleague Letter, Office of Elementary and Secondary Education, U.S. Department of Education, August 20, 2026. Not a regulation, not a rule, and not enforceable. No compliance deadline. Not research.

Baesler, K. (2026, August 20). Dear colleague letter: Education technology and screen time in schools. Office of Elementary and Secondary Education, U.S. Department of Education. ed.gov

The six-page letter, signed by Assistant Secretary Kirsten Baesler, opens by distinguishing between recreational technology and educational technology and argues that policy should not be driven by screen time alone. That framing is the letter's defensive half, and it is aimed at state phone and screen-time restrictions. The operative half is procurement. The letter states that states and districts should distinguish policies addressing recreational technology from those governing instructional technology, evaluate education technology based on demonstrated learning outcomes and instructional value rather than screen time alone, incorporate evidence of effectiveness into procurement and renewal decisions while looking for products that have conducted randomized controlled trials and demonstrated positive outcomes, establish processes for reviewing implementation and outcomes, and support professional learning.

It asks education technology providers to publish rigorous independent evaluations of product impact on student learning whenever feasible, share implementation guidance, provide transparent information about product capabilities, limitations, and performance, and continuously improve products using evidence from implementation and student outcomes. It ties this to existing law, stating that the Elementary and Secondary Education Act, as amended by the Every Student Succeeds Act, reflects a longstanding federal commitment to evidence-based decision-making, and that states and districts should draw upon the established evidence framework when making procurement and implementation decisions. It names Louisiana, Arkansas, Indiana, Michigan, and Texas as states exploring contracting models with shared performance measures and outcome-linked terms. It reaffirms the five principles from Secretary Linda McMahon's July 2025 AI letter that education technologies should be educator-led, ethical, accessible, transparent, and protective of student data, and extends them beyond AI to education technology generally.

Two of the most quotable items are in the Department's press release rather than in the letter itself and should be cited accordingly. The press release sets out five questions every education technology product should be able to answer: what learning problem does it solve, when should it be used, for whom should it be used, for how long should it be used, and what evidence demonstrates that it improves student learning. The press release also states that when evidence shows a tool is not improving learning, schools must be willing to change course, and that when repeated findings confirm persistent shortcomings, they should remove it altogether. That removal language does not appear in the letter.

Study context. What this cannot establish is any obligation. Guidance of this kind creates no duty, no deadline, and no cause of action, and the letter does not identify where randomized trial evidence for AI products exists, what evidence tier is sufficient, or what a district should do about a category where none exists.

Leadership implication. Put the five questions into your renewal workflow this month, before fall renewals close, and assign them to the budget owner for each contract rather than to your technology director. For every AI or edtech product over a dollar threshold your cabinet sets, require a one-page answer to all five, with question five, what evidence demonstrates that it improves student learning, answered with a citation or with the words none available. Both answers are defensible. A blank is not. Then add a single clause to your next renewal template requiring the vendor to furnish evidence of an independent evaluation and a written statement of product limitations, as this letter asks providers to supply. If a vendor will not, that refusal is now documented as a deviation from a published federal expectation.

2. California Writes Into the Brown Act That an AI System Is Not a Member of the Public

Source type. State legislation, pending. Passed both chambers, ordered engrossed and enrolled August 20, 2026. Not yet presented to the Governor and not law. Not research.

California Legislature. (2026). SB 1159: Artificial intelligence: Transparency and governance (2025-2026 Reg. Sess.). legiscan.com [Fallback ladder rung three. The official California Legislative Information pages at leginfo.legislature.ca.gov returned an empty response on repeated automated retrieval. Bill title, covered acts, vote tallies, and dates were independently confirmed against the Legislative Counsel's Digest as reproduced by CalMatters Digital Democracy at calmatters.digitaldemocracy.org.]

SB 1159, authored by Senator Cabaldon, adds Section 54951.5 to the Government Code among other amendments. The Legislative Counsel's Digest states that for purposes of the California Public Records Act, the Bagley-Keene Open Meeting Act, the Ralph M. Brown Act, the Legislative Open Records Act, the Administrative Procedure Act, the California Coastal Act of 1976, and the California Environmental Quality Act, the terms person, interested person, participant, and member of the public do not include artificial intelligence systems, autonomous agents, or robots, whether physical or digital.

The operative language that applies to school boards is new Government Code Section 54951.5, added within the chapter that is the Ralph M. Brown Act. Subsection (c) permits a local agency to use a disclosure verification tool to determine whether AI is present. Subsection (d) preserves the right of a natural person to use AI, including assistive technologies, to facilitate that person's own engagement with a governmental agency, provided the volume and frequency of the engagement are reasonably consistent with ordinary participation by a natural person. The bill also prohibits a person from knowingly using AI to falsely represent that a natural person appeared before or submitted information to a governmental agency. It passed the Assembly 74 to 0 on August 13, and the Senate concurred in Assembly amendments 37 to 0 on August 20.

Study context. What this cannot establish is whether the distinction is administrable. Nothing in the record shows how a district would determine, in the moment, that a submitted comment was machine-generated, or what the standard of proof would be.

Leadership implication. This lands on your board secretary and your public records officer, not your technology director. Two documents need work before your first board meeting of the year. First, the public comment procedure: who is authorized to determine that a submission is not from a natural person, on what evidence, and what notice the submitter receives. Second, the public records request protocol, applying the same question to a requester. Write the volume-and-frequency standard from subsection (d) into both, because the statute protects a person using assistive technology and does not protect a person operating an agent at machine scale. Your staff will need a written line between those two. Districts outside California should draft the same two paragraphs now.

3. An Instructor of Record Must Be a Person, and the Carve-Out Is the Point

Source type. State legislation, pending. Enrolled and presented to the Governor August 18, 2026 at 3 p.m. Not law. Applies to the California State University, not to K-12. Not research.

California Legislature. (2026). SB 928: California State University: Faculty employees (2025-2026 Reg. Sess.). legiscan.com [Fallback ladder rung three. Enrolled bill text, roll calls, and the August 18, 2026 presentment action confirmed in LegiScan's mirror. The leginfo.legislature.ca.gov bill status page did not render on automated retrieval across repeated attempts.]

SB 928, authored by Senator Cervantes, adds Section 89500.3 to the Education Code. Subsection (b) provides that the instructor of record for a course of instruction shall be a person who meets the rule to serve as a faculty employee teaching credit or noncredit instruction established pursuant to Section 89500. Subsection (c) provides that a California State University faculty employee shall be a person who meets the rule to serve in that position. Subsection (d) provides that the section does not prohibit California State University employees from using artificial intelligence tools to assist in the operations of the university or in providing services to students. The bill passed the Senate 37 to 0 on April 23, 2026, and the Assembly 74 to 0 on August 13, 2026.

Study context. The design limit is jurisdictional and should be stated plainly. This governs a higher education system, and no K-12 obligation follows from it. Its relevance is as drafting architecture, not as authority.

Leadership implication. Several states already prohibit AI from replacing a teacher. This bill does something structurally cleaner, and your general counsel should read the three subsections together. It does not regulate how much instruction AI may deliver, a line no one can police. It defines who may hold the role, then separately protects the use of the tool. That converts a policy question into a personnel question. Ask your human resources director one question this month: for every course, intervention, and credit-recovery pathway your district offers, including anything delivered through a third-party online provider, can you name the person of record and produce the credential? Districts running vendor-delivered credit recovery are the ones most likely to find they cannot do so.

4. The Federal Evidence Clearinghouse Contains No Studies on AI and Student Outcomes

Source type. Federal agency research synthesis. Institutional, not peer reviewed. Published August 10, 2026.

Young, J. M., & Peterson, K. (2026, August 10). AI in K-12 education: The good, the bad, and the guardrails to consider. Regional Educational Laboratory Northeast and Islands, Institute of Education Sciences, U.S. Department of Education. ies.ed.gov

The REL Northeast and Islands Governing Board, a panel of state, district, and school leaders from each of the region's nine states and jurisdictions, requested an update on the state of the research on AI and student outcomes. The response team searched the What Works Clearinghouse, which the piece describes as the most trusted repository of scientific evidence for education, and reports that the search did not identify any research studies on AI and student outcomes. Instead, it located a 2026 Stanford review that found only 20 rigorous education research studies producing causal evidence about AI's impacts, with most research conducted in postsecondary settings and causal studies more common in high school than in middle or elementary settings.

From what does exist, the synthesis reports three patterns. Teacher-mediated and AI-augmented tutoring, meaning tools used by or alongside a teacher who reviews, guides, or approves what the AI does, may promote learning. Student-facing tools show mixed effects: access alone did not improve exam scores but did improve homework performance, and tutoring chatbots giving hints rather than answers produced performance equal to traditional study methods, not better. General-purpose chatbot use during learning is associated with lower exam performance, shallower learning processes, reduced brain activity, and weaker recall, even when students found the tools helpful. The piece also notes emerging evidence that students may perceive AI-mediated feedback as less caring and supportive than teacher feedback.

Read against signal 1, this is the week's central fact. The Department's guidance arm asked districts to buy on randomized trial evidence on August 20. The Department's research arm reported on August 10 that its own clearinghouse contains no AI studies.

Study context. This is an agency research synthesis, not a study. It reports the state of a literature and does not itself test anything.

Leadership implication. You now have two federal citations that work together, and they belong in the same board slide. The Department has told you to buy on evidence of effectiveness, and that its clearinghouse holds none for AI. That pairing changes who carries the burden of proof: a vendor claiming a learning effect is claiming to hold evidence the federal repository does not have, and should be asked to produce the study, the design, and the comparison condition. Use the three patterns as procurement categories: price teacher-mediated tools as the defensible purchase; price student-facing tools as a pilot with a written exit ramp; and treat general-purpose chatbot access as an instructional risk that requires a written instructional purpose before deployment.

5. Two Peer-Reviewed Papers in One Week Find the Field Is Measuring Teacher AI Competence as Technical Fluency

Source type. Two peer-reviewed articles, Frontiers in Education. One critical integrative review published August 18, 2026. One mixed-methods original study published August 14, 2026. Design-stage evidence, single institution, self-report, neither isolated to K-12.

Badoi-Hammami, M., Matei, S.-R., Ivan, S., Colareza, C., Paraschiv, R., Baiceanu, C.-M., Popa, C.-S., Casangiu, L.-I., & Lungu, I. (2026). The illusion of teaching competence in AI-mediated education. Frontiers in Education, 11, 1872857. doi.org/10.3389/feduc.2026.1872857

Hu, Y., & Wang, L. (2026). AI literacy and subject specialization in pre-service teacher education: A mixed-methods study of dimensional profiles and perceived pedagogical challenges. Frontiers in Education, 11, 1876627. doi.org/10.3389/feduc.2026.1876627

The nine authors of the review, based at Ovidius University of Constanta and Titu Maiorescu University in Romania, screened 1,876 initial records, reducing them to 15 studies published between 2018 and 2025, following PRISMA 2020 guidance. They name a construct they call the illusion of teaching competence, defined as the overestimation of pedagogical expertise derived from technological fluency and instrumental performance. Their operational finding concerns the literature itself: the reviewed studies evaluate teaching competence in AI-mediated contexts largely through quantitative self-report indicators, chiefly frequency of AI tool use and self-assessed digital competence, and perceived digital competence may overestimate actual pedagogical competence. They examine AI-TPACK and Intelligent-TPACK and argue such frameworks can emphasize technical dimensions over pedagogical ones.

The second paper supplies the measurement evidence from the other direction. Hu and Wang surveyed third-year students in a primary teacher education program at Wuchang Institute of Technology in Wuhan, China, retaining 161 valid responses of 175 targeted, comprising 81 Chinese specialization, 58 Mathematics, and 22 English. Using a 36-item AI literacy scale measuring AI Perception, AI Knowledge and Skills, AI Application and Innovation, and AI Ethics, they report an overall mean of 3.67, with Mathematics students highest at 3.94, compared with 3.54 for Chinese and 3.48 for English. The finding that matters is in their own limitations. Harman's single-factor test explained 52.30 percent of the variance on the first factor, above the conventional threshold, and the confirmatory fit was only partial, with an RMSEA of .095. The authors write that the results should not be treated as definitive evidence that the four dimensions are empirically distinct.

Study context. Both papers state their limits directly. The review cannot eliminate interpretive subjectivity, excludes grey literature, and spans pre-service through university teachers rather than K-12 alone. The study is cross-sectional, entirely self-report, drawn from a single program in Central China, with an English subgroup of 22 participants and no preregistration.

Leadership implication. This is a specification defect in most district AI professional development and in the diagnostics sold alongside it, and it is fixable at low cost this fall. If your professional development evaluation asks teachers how confident they feel using AI tools and how often they use them, you are measuring the thing the review says is mistaken for competence. Add one item to whatever instrument your provider uses: require a teacher to justify an instructional decision made with AI assistance, in writing, against a pedagogical standard, and have an instructional leader score it. Before you buy a readiness diagnostic, ask the vendor for the validation study and look for two numbers: the common method variance figure and the confirmatory fit indices. Under signal 1, that request is now aligned with published federal expectation rather than being an unusual ask.

Emerging Strategic Themes

Theme 1. The federal lever is procurement, not prohibition. Washington did not tell districts what AI may do. It told them what to require before paying for it, and pointed at the ESEA evidence framework that already governs federal education spending. Guidance is not enforceable, but it establishes what a reasonable district was told to ask. Expect the five questions to migrate into state procurement templates and grant assurances long before they appear in any regulation, and expect them to be quoted back at districts that cannot answer them.

Theme 2. The evidence shelf is empty and now officially so. Until August 10, a superintendent could reasonably say the evidence was unsettled. The Institute of Education Sciences has now stated that its own clearinghouse holds no studies on AI and student outcomes, and ten days later the Department asked districts to buy on evidence of effectiveness. That converts an ambiguity into a documented condition, and documented conditions change legal exposure. Boards voting on AI tools this fall should be told both sides of the argument in the same motion.

Theme 3. Personhood is the new drafting tool. Legislatures spent three years writing rules about what AI may do in schools and discovered that conduct rules require line-drawing, enforcement, and constant amendment. A definitional amendment requires none of those. Once a statute states that an AI system is not a member of the public, every rule referencing members of the public automatically updates. Expect the next wave of state bills as amendments to definitions sections rather than as new AI chapters, which means they will not surface in your legislative tracker under the search term artificial intelligence.

Theme 4. Measurement validity becomes a procurement specification. Two peer-reviewed papers in one week said the same thing from different directions: the field measures teacher AI competence as technical fluency, and the instruments used to measure it cannot establish that their own dimensions are distinct. Districts are buying readiness diagnostics and professional development packages priced against those profiles. Ask for the validation study the way you would ask for a security audit, and treat an instrument without one as an unvalidated assessment being used to allocate staff development dollars.

What Was Not Found

No federal guidance identifies where randomized trial evidence for AI products actually exists. The August 20 letter directs districts toward products that have conducted randomized controlled trials and demonstrated positive outcomes. It does not name a product, a registry, an evidence-tier threshold, or a clearinghouse that a procurement officer could consult. Ten days earlier, the Department's own clearinghouse reported holding nothing on AI and student outcomes. Districts are being asked to apply a standard without the instrument during the fall renewal cycle.

No guidance addresses what a district should do about products it has already bought. The letter speaks to procurement and renewal decisions in the future. It says nothing about the installed base. The press release states that schools should remove tools that repeatedly fail to improve learning. Still, no document addresses contract terms already executed, sunk costs, instructional dependency, or the evidence threshold that would justify mid-term termination. Districts with multi-year AI contracts signed in 2024 and 2025 have no federal instruction that applies to them.

No study measures what happens when a public body is authorized to disregard AI-generated public comment. SB 1159 is close to enactment, and no empirical work exists on the operational question it raises, including error rates in identifying machine-generated submissions, effects on participation by disabled residents who use assistive technology, or the volume-and-frequency standard the statute leaves undefined.

No study measures whether requiring a human instructor of record changes any student outcome. SB 928 encodes an intuition most educators share. We found no research testing it, in either direction, at any level of schooling. The template is likely to be copied into K-12 bills before evidence exists either way.

The evidence gap for specific populations remains unaddressed and is now compounded. Nothing this week addressed English language learners, students with disabilities, or high-poverty districts. This matters more under the new guidance than it did before it, because a district told to procure on demonstrated outcomes has no outcome evidence for the student populations whose services carry the highest legal exposure. Elementary literacy and non-STEM subjects remain similarly bare, and the IES synthesis confirms causal studies are rarer in elementary and middle grades than in high school.

No evidence exists on whether guidance on evidence-based procurement changes district purchasing behavior. The Department names Louisiana, Arkansas, Indiana, Michigan, and Texas as exploring outcome-linked contracting and describes emerging research as suggestive. No evaluation establishes that these models improve student outcomes or reduce ineffective spending. The recommended policy instrument has roughly the same evidentiary status as the products it is meant to screen.

The pattern this week is not that evidence is missing. It is that a federal expectation was published against an evidence base the same agency has documented as empty for this product category. That does not argue for delay, because the tools are already in classrooms and the contracts are already signed. It argues for writing down what you do not know, at the moment you buy, in the file where a future reader will look. Monitoring, documentation, and exit ramps are the correct response. A district that records none available in response to the Department's fifth question has complied with the spirit of the guidance. A district that leaves it blank has not done so.

Novo Executive Summary

In one week, the federal government told districts to judge educational technology by demonstrated learning outcomes, to incorporate evidence of effectiveness into procurement and renewal decisions, and to expect vendors to produce independent evaluations, while its own What Works Clearinghouse reported having no studies on AI and student outcomes at all. California moved in a different direction entirely, excluding AI systems from the categories of person, participant, and member of the public under the Brown Act, and requiring that an instructor of record be a person. Neither the federal letter nor the California bills create a deadline a district can put on a calendar. All of them raise questions that a district will be asked to answer in public, without warning. The districts that answer well will not be the ones that picked better tools. They will be the ones who build the architecture first: decision rights that specify who determines what, procurement specifications that state the evidence standard and record its absence when it is absent, evaluation designs written before deployment rather than after, and role-based literacy pathways that distinguish fluency from judgment. That architecture is what Novo Innovative Pathways builds with district leaders.

Watch This Week

• Monday, August 31. The California Legislature adjourns sine die with 24 AI-related bills near final passage. The Governor then holds a 30-day signing window. SB 1159 awaits presentment. SB 928 has been with the Governor since August 18.
• Federal. The Department of Education has not stated whether a supplemental grant priority, technical assistance, or an evidence registry for AI products will follow the August 20 Dear Colleague Letter. Watch the Federal Register and Department grant priority notices through the fall, because the letter's operative force depends entirely on whether it is attached to money.
• California AB 2656. Carried forward from Edition 29, it cleared Senate Appropriations 7 to 0 on August 13 and still requires a Senate floor vote before August 31. It would require 45 days of written notice to a recognized employee organization before a public employer deploys generative AI in represented work.
• California AB 1159. Carried forward from Editions 27 through 29, it would extend California student privacy protections to digital operators marketing to schools. It cleared Senate Appropriations 5 to 0 on August 13 and awaits a floor vote. AB 2392 passed 7 to 0 the same day.
• Tuesday, August 25. The Michigan Legislature returns to Lansing. SB 760, the kids chatbot safety bill, passed the Senate April 29 and sits with the House Communications Committee. Pennsylvania's House returns September 9 and its Senate September 28.
• Florida. The proposed AI amendment to Internet Safety Policy Rule 6A-1.0957, carried forward from Editions 27 through 29, has still produced no formal Notice of Proposed Rule following the August 5 development workshop. The district compliance date in the draft remains January 1, 2027.
• Carried forward and still unresolved. 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 Governor Hochul's action by December 31; New York City's promised AI playbook; and the October publication of Nagashima and colleagues' CSCW study of teacher and student misalignment over classroom AI control, flagged first in Edition 26.

Sources

Governance and Policy

Research, Peer-Reviewed

Research, Preprint (Not Peer-Reviewed)

A direct review of the arXiv listings for Computers and Society, Human-Computer Interaction, and Computer Vision for August 17 through 21, 2026, produced no K-12 preprint with governance or procurement consequences. The single K-12 item in that listing was a program description of a mentorship model with no outcome data.

Institutional Report (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.