Who shaped US school AI guidance? A 36-jurisdiction study found shared sources but few student or parent authors
An unreviewed analysis coded 1,539 pages from 35 states and Puerto Rico. Privacy, integrity and bias appeared almost everywhere, while budgets, roadmaps and named community needs were less consistent; no document credited parents as contributors.
The short version
- What happened
- Researchers coded 36 guidance sets totalling 1,539 pages, covering 35 states and Puerto Rico as available by 31 August 2026; two coders reached 86.3% agreement and Cohen's kappa of 0.73.
- Who it affects
- Teachers may receive strong warnings about privacy and integrity without equally specific budgets, roles or review cycles.
- What remains unproven
- The manuscript is unreviewed and covers only 36 jurisdictions that had issued guidance by 31 August 2026.
Keep in mind
The manuscript is unreviewed and covers only 36 jurisdictions that had issued guidance by 31 August 2026.
Binary codes capture whether a topic appears, not its depth, quality, tone, enforceability or effect.
Documents span different dates, and citations or shared text do not by themselves establish causal influence.
One author helped write Washington's guidance; AI tools assisted coding checks and dictionary drafting under disclosed human review.
Source analysis · 1 linked source. Primary analysed source date: 4 November 2026.
Read the full report · 8 min The evidence, methods, implications and limitations
The answer: states converged on risks, but not on who implements the guidance
A new unreviewed study of K-12 artificial-intelligence guidance finds striking agreement on familiar legal and integrity concerns, alongside much thinner treatment of budgets, accountability and specific communities. Researchers analysed the latest guidance they could verify by 31 August 2026 from 35 US states and Puerto Rico. Across 1,539 pages, academic integrity, data privacy and algorithmic bias appeared in 35 of the 36 jurisdictions. Yet practical questions—who evaluates vendors, who pays, how implementation is reviewed and which groups helped write the rules—were far less consistently answered.
The result matters because state guidance can operate like soft law even when it is not formally binding. Districts, teachers and technology teams may treat a state document as the authoritative starting point for procurement, assessment and acceptable use. The study does not score which state has the best policy, test legal compliance or observe a classroom. It maps what documents mention, who they address, where their citations lead and who is publicly credited. A mention can be a sentence or a detailed framework, so coverage should not be confused with quality or depth.
The corpus came from Ballotpedia's list of state education-department guidance, Puerto Rico material located through AI for Education and manual checks of official state sites. University, district-only and other non-state documents were excluded. Interactive websites were converted to PDFs. The researchers also linked eight state variables on demographics, broadband, school spending, poverty and legislative composition. Those sources covered different years, an accepted limitation in their rank-based comparisons. States without published guidance are absent, so the study describes adopters rather than the whole country.[1]
How 36 documents were coded and checked
Two authors first reviewed three documents and built 28 binary questions in seven families: document metadata and structure, audience, pedagogy and instruction, actionability, workforce readiness, and responsible use and equity. They then independently coded all 36 guidance sets. Agreement was 86.3%, with Cohen's kappa of 0.73, and disagreements were resolved through discussion. An AI system acted as a secondary consistency check, flagging 145 of 1,080 cells. Two authors reviewed every suggestion and accepted 60, or 41.4%, of those proposed changes.
The study separately extracted 5,351 reference mentions and resolved them into 328 distinct sources, producing 898 state-source pairs. It identified 61 citations from one jurisdiction to another and coded whether a citation was a reference or explicit adoption. Contributor credits produced 565 named people, 49 organizations and 30 group statements. One author reviewed the source dictionary and contributor mappings, some of which were initially drafted with coding assistants. The paper discloses those uses and says every substantive code change received human review.
That workflow is more transparent than an automated scrape, but it still contains judgment. Binary coding does not capture tone, specificity or enforceability. Citation counts do not prove influence, and shared phrasing can result from legitimate adoption rather than misconduct. Documents were issued and updated across roughly three years, giving early states more time to be cited and late documents more opportunity to address emerging issues. The first author also served on Washington's AI advisory group and co-wrote that state's guidance, a relevant disclosed connection because Washington is among the most cited states.[1]
The distinction that matters
What the study counted
The unit was the public guidance document, not a school, classroom or student outcome.
Corpus
Verify 36 jurisdictional guidance sets
1,539 pages
Content
Code 28 yes-or-no questions
Coverage, not depth
Sources
Resolve 5,351 mentions
328 distinct sources
Contributors
Read public credits
Few students, no named parents
What the documents emphasize—and whom they overlook
Workforce preparation or upskilling appeared in 86% of the guidance, often framed as an urgent response to an AI-driven future. Personalized learning also appeared in 86%, though the authors say it was often loosely defined. Data privacy appeared in 97%, helped by established laws such as FERPA and COPPA. Accessibility and disability appeared in 75%. Newer issues were less common: half discussed deepfakes or synthetic media and 22% mentioned AI companion applications. Equity of access appeared in 86%, but rural or underserved contexts appeared in 42% and tribal or Indigenous communities in 14%.
The statistical comparisons found little alignment between a state's measured characteristics and what its guidance mentioned. States that discussed rural learners did not have significantly larger rural populations, and mentions of English learners or disability were not associated with the corresponding student shares. The exception was tribal or Indigenous coverage: states that mentioned it had median Native enrolment of 10.3%, compared with 0.4% among those that did not, with p=.001. Broadband access, free-lunch share and per-pupil spending were not significantly associated with the selected capacity-building topics. These small group comparisons cannot prove that context played no role.
Partisanship also showed no statistically significant association with the varying equity and responsible-use topics. Three issues—bias, integrity and privacy—had almost no variation because 35 documents covered each. The more revealing gap was participation. Of the 22 documents that named contributors, teachers were represented in 15 jurisdictions, administrators in 14 and IT staff in 12. Students appeared in five. No document named parents or families as contributors even though 83% addressed them as stakeholders. Credits may be incomplete, but the public record does not show broad co-authorship by people governed by the guidance.[1]
Shared sources can help states move faster—and narrow the frame
Five sources appeared in at least half the jurisdictions: FERPA in 29, TeachAI and the US Department of Education in 28 each, COPPA in 27 and the International Society for Technology in Education in 20. TeachAI appeared in 78% and Code.org in 44%. Two-thirds of the 328 distinct sources were cited by only one state, but the shared core clustered around federal law and connected national nonprofits. A pair of states shared, on average, about one in eight sources. States in the same Census region were no more similar than states in different regions in the authors' permutation test.
Twenty-one jurisdictions cited another state, generating 61 peer citations. Fifty-three were references or resource listings; eight were coded as explicit adoption. Oregon, North Carolina and Washington—early publishers—were prominent sources. North Carolina and Ohio used Washington's Human-AI-Human model, and North Dakota built on North Carolina's roadmap. Text comparisons found 6,239 overlapping ten-word phrases and 88 highly similar sentences across documents; Wyoming reproduced a TeachAI model and West Virginia and Utah adapted it. The authors frame this as policy diffusion, not plagiarism.
For schools, the practical lesson is to treat state guidance as a starting point that needs local ownership. Districts should assign accountable roles, publish vendor and data checks, state what evidence justifies a tool, fund teacher learning and include students, families and staff in revision. The study cannot show whether any recommendation changed practice, reduced harm or improved learning. Confidence would rise with longitudinal versions, interviews with authors, district implementation data and comparisons with statutes and union policies. Until then, the evidence shows a common vocabulary and uneven operational detail—not a national consensus that every school should adopt the same AI tools.[1]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring ai governance: compare the finding and its evidence limits.
What do universities need before their AI rules become real governance?
A new Asia-Europe Foundation collection proposes a 48-indicator model and tests it against public documents from 12 universities. Scores ranged from 30 to 71, but the purposive, English-only exercise did not test whether higher scores produce safer or fairer decisions.
Read the next reportExplore the full source and evidence record
What this means for people
- Teachers may receive strong warnings about privacy and integrity without equally specific budgets, roles or review cycles.
- Students and families are addressed by many documents but rarely visible in the public authorship record.
- District technology teams need local procurement, accessibility and incident procedures rather than assuming a state mention is an implementation plan.
Global context
US states have unusual authority over education, and this corpus is not a template for national systems elsewhere. The method—tracking content, citations and contributors—can be applied internationally, but legal duties, school capacity and community participation need local interpretation.
What the evidence does not yet show
- The manuscript is unreviewed and covers only 36 jurisdictions that had issued guidance by 31 August 2026.
- Binary codes capture whether a topic appears, not its depth, quality, tone, enforceability or effect.
- Documents span different dates, and citations or shared text do not by themselves establish causal influence.
- One author helped write Washington's guidance; AI tools assisted coding checks and dictionary drafting under disclosed human review.
What to watch next
- Peer review and release of the coded corpus.
- Revisions after 31 August 2026.
- District-level implementation and spending evidence.
- Named participation by students, families, unions and disability groups.
Living evidence record
Impact record IAI-0MJ6JU9
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
6 November 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what arXiv published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 6 November 2026
Research question, topics and regions
What US state education-department AI guidance covers, whom it addresses, which sources it cites and who is credited with writing it
Topics: Schools · AI governance · Policy diffusion · Student voice
Regions: United States
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.