Issue 05 · Evidence briefing
Risk-tiered school and legal controls, state-policy gaps, university teaching tradeoffs, civic capacity, and measured creative-learning effects.
A decision-oriented read of what public institutions tried, what the evidence supports, and what leaders should design for next. Vendor claims are treated as claims, not outcomes.
Edition intelligence
Synthesis
DC's red-yellow-green staff policy and California's proposed duties for lawyers and arbitrators both translate abstract ideas such as human oversight into specific allowed, safeguarded, and prohibited actions. The new 40-jurisdiction K-12 policy analysis shows why that translation matters: privacy and ethics language is widespread, but continuing monitoring and stakeholder trust are less developed.
State education guidance increasingly addresses ethics and privacy, but the comparative study finds weaker attention to ongoing monitoring and stakeholder trust. New America's field review similarly describes rapid legislation and experimentation without a coherent implementation vision, while Macquarie's teaching deployment exposes how user volume and internal satisfaction do not resolve questions about learning quality, workload, or student value.
The creativity meta-analysis indicates that GenAI can improve a defined learning outcome, particularly in original and visual work. Yet DC prohibits AI from making high-stakes educational judgments, and Macquarie's experience shows that a grounded assistant can still create student and workforce concerns when introduced alongside reduced synchronous human teaching. The durable question is not whether AI is present, but which human capability or relationship the workflow preserves.
Macquarie reports nearly 80,000 questions in half of 2026 and New America identifies a broadening field of state sandboxes and city pilots, but neither establishes mission impact by volume alone. The creativity meta-analysis provides a useful contrast by measuring a specific outcome and effect size across 31 studies. Mature SLED evaluation needs adoption telemetry and outcome evidence, with each labeled correctly.
Full record
Showing 6 of 6 records · All
Updated September 1, 2026
OSSE released its first model policy for staff AI use after a February 2026 survey found that only 45% of DC local education agencies had established a staff AI policy. The voluntary template classifies uses as red, yellow, or green: it prohibits AI for student and staff surveillance, discipline, teacher evaluation, and IEP or Section 504 eligibility; permits guarded use for activities such as drafting IEP language and grading; and allows lower-risk drafting, customization, analysis, communication, and logistics with awareness and human review.
OSSE released its first model policy for staff AI use after a February 2026 survey found that only 45% of DC local education agencies had established a staff AI policy. The voluntary template classifies uses as red, yellow, or green: it prohibits AI for student and staff surveillance, discipline, teacher evaluation, and IEP or Section 504 eligibility; permits guarded use for activities such as drafting IEP language and grading; and allows lower-risk drafting, customization, analysis, communication, and logistics with awareness and human review.
The agency says the policy was built from a national policy review and consultation with local school leaders. It requires approved enterprise tools, human accountability, training before use, demonstrated AI literacy, and annual renewal; links personally identifiable information to FERPA, COPPA, CIPA, IDEA, HIPAA, and District privacy requirements; and announces two forthcoming educator courses. This is implementation guidance, not evidence that the controls have improved learning, privacy, or compliance.
This is a practical state-level pattern for giving educators usable decisions rather than a list of principles. It also recognizes that the same tool can move between risk levels depending on whether it drafts material, influences a review, or makes a consequential determination.
Enforce the stoplight model through identity-aware approved tools, data classification, blocked or approval-gated capabilities, logging, and workflow-specific configurations. Separate general drafting from systems that access student records or influence IEP, grading, discipline, monitoring, or employment decisions.
Require each LEA to tailor and formally adopt the policy, name accountable owners, define evidence for yellow-use approval, map training to permissions, document human review, and schedule updates as models and laws change. Extend the framework with separate student-use and procurement policies because OSSE explicitly leaves those areas out of scope.
Restrict sensitive work to enterprise tools with contractual data-use limits, retention and deletion controls, least-privilege access, audit logs, and vendor cybersecurity evidence. Treat disability, health, discipline, surveillance, and evaluation data as higher-risk even when AI only drafts a recommendation.
The policy is voluntary guidance and not legal advice. OSSE has not reported adoption, compliance, incident, equity, accessibility, or outcome data, and the release does not govern student use or establish a complete AI procurement standard.
A qualitative content analysis examined state-level K-12 AI guidance issued from 2023 through 2026 across 40 states and territorial jurisdictions. Ethics and data privacy dominated the policy landscape, equity and educator capacity received moderate attention, and stakeholder trust and ongoing monitoring were comparatively underdeveloped.
A qualitative content analysis examined state-level K-12 AI guidance issued from 2023 through 2026 across 40 states and territorial jurisdictions. Ethics and data privacy dominated the policy landscape, equity and educator capacity received moderate attention, and stakeholder trust and ongoing monitoring were comparatively underdeveloped.
The published article documents the included jurisdictions, issuing bodies, policy status, update status, inclusion rationale, and document length. Most entries are advisory or non-binding guidance, often explicitly described as living or evolving. The authors identify gaps in stakeholder trust and monitoring rather than measuring failures in deployed educational AI systems.
The study provides a cross-jurisdiction calibration point for education leaders writing or revising AI policy. It suggests that many systems have established an initial compliance and values layer but have not yet built the feedback, assurance, and legitimacy mechanisms needed for sustained operation.
Policy should map to operational telemetry: model and tool inventory, version changes, data flows, approval state, incident records, accessibility findings, user feedback, and outcome measures. Living guidance requires technical controls and inventories that can be updated without rediscovering the environment.
Add formal review cycles, stakeholder participation, complaint and appeal channels, monitoring responsibilities, and public reporting to ethics and privacy principles. Distinguish advisory guidance from mandatory district controls and define how state agencies verify local implementation.
Privacy prominence is useful but should be connected to real data-flow maps, vendor obligations, retention, model-training restrictions, sensitive-data access, incident response, and periodic control testing rather than policy language alone.
This is a content analysis of policy documents, not an evaluation of compliance or AI outcomes. The accessible abstract provides only high-level findings, policy documents vary greatly in length and authority, and coding judgments may not capture informal practices outside published guidance.
Both chambers of the California Legislature approved SB 574 and sent it to the governor. The measure would prohibit lawyers from delegating the practice of law to generative AI, require reasonable verification and correction of outputs and citations, require disclosure of AI use in court submissions, restrict entry of confidential and nonpublic information, and prohibit arbitrators from delegating decisions to AI.
Both chambers of the California Legislature approved SB 574 and sent it to the governor. The measure would prohibit lawyers from delegating the practice of law to generative AI, require reasonable verification and correction of outputs and citations, require disclosure of AI use in court submissions, restrict entry of confidential and nonpublic information, and prohibit arbitrators from delegating decisions to AI.
Reuters verified legislative passage on September 1 and linked the official bill text. The bill adds specific duties around confidentiality, accuracy, disclosure, and non-delegation and would require the Judicial Council to revisit its AI standard as the technology develops. It had not yet been signed or implemented, so there is no outcome evidence.
Courts, public defenders, prosecutors, attorneys general, municipal counsel, hearing officers, and administrative adjudicators all perform high-consequence knowledge work. The bill offers a capability-based control pattern that permits assistance while preserving professional responsibility and decisional authority.
Legal AI environments need protected matter workspaces, role and case-based access, source-grounded retrieval, citation provenance, immutable review records, disclosure support, and hard separation between research or drafting and the authoritative act of filing or deciding. Agentic tools should not possess unilateral submission or adjudication authority.
Translate professional duties into acceptable-use rules, mandatory verification workflows, training, disclosure criteria, sanctions or remediation, and periodic review. Procurement should require confidentiality protections, auditability, model-change notice, and the ability to retain evidence of human review.
Prevent confidential, personal, medical, financial, witness, victim, and sealed information from reaching systems without restricted access and enforceable confidentiality. Apply data-loss prevention, matter-level authorization, logging, retention, and incident handling to legal copilots.
SB 574 was awaiting gubernatorial action when reported and may change through signature, veto, litigation, or implementation. Some legal experts told Reuters that parts duplicate existing ethical duties, and the record does not show whether the proposed requirements reduce hallucinated filings or confidentiality incidents.
Macquarie's educator-configured Virtual Peer became part of weekly learning in two mandatory psychology units offered online. The AI activities were optional and used professor-supplied, checked material, while paid tutors still offered optional feedback sessions. The online format no longer included the prior optional weekly Zoom tutorials, prompting some students and staff to question whether AI was supplementing or displacing human teaching.
Macquarie's educator-configured Virtual Peer became part of weekly learning in two mandatory psychology units offered online. The AI activities were optional and used professor-supplied, checked material, while paid tutors still offered optional feedback sessions. The online format no longer included the prior optional weekly Zoom tutorials, prompting some students and staff to question whether AI was supplementing or displacing human teaching.
The university reported that Virtual Peer answered almost 80,000 questions during 2025 and nearly as many in the first half of 2026, mostly administrative, and said surveyed users found it valuable. The reporting also documents two units with about 400 and 700 students, roughly one optional human feedback session per 70 students, student dissatisfaction, staff concerns, and broader labor negotiations. No controlled learning, retention, equity, or cost evaluation was reported.
The case shows how an apparently bounded, grounded assistant can become an operating-model and labor issue when deployment coincides with fewer structured human interactions. Higher-education leaders need to evaluate the complete service design, not the chatbot in isolation.
Ground assistants in faculty-curated course material, preserve clear provenance, monitor conversations under a defined privacy policy, and provide reliable escalation to instructors. Instrument question type, unanswered or low-confidence interactions, escalation, accessibility, learning outcomes, and the effect of AI on demand for human support.
Evaluate AI together with class modality, staffing, workload, student fees, accessibility, and human-contact commitments. Establish faculty ownership, student notice and alternatives, labor consultation, pedagogical review, and predefined evidence for whether the tool supplements or replaces teaching activity.
Clarify what student conversations are monitored, who can access them, how long they are retained, whether they train models, and how sensitive wellbeing or academic information is escalated. Protect course intellectual property and student records while avoiding surveillance-like use of interaction logs.
The source is independent reporting rather than a formal evaluation. Usage and satisfaction figures are university-reported, student concerns are illustrative rather than representative, the activities were optional, and the reporting does not establish that AI caused staffing or modality decisions or changed learning outcomes.
New America's review combined more than 40 practitioner and expert interviews, pilot work, literature review, legislation analysis, and a field scan of state and city activity. It found rapidly expanding legislative and pilot activity, with states favoring enterprise sandboxes or walled gardens and cities favoring stand-alone service pilots, but local capacity, coherent strategy, trust, infrastructure, budgeting, and evidence of return remained major constraints.
New America's review combined more than 40 practitioner and expert interviews, pilot work, literature review, legislation analysis, and a field scan of state and city activity. It found rapidly expanding legislative and pilot activity, with states favoring enterprise sandboxes or walled gardens and cities favoring stand-alone service pilots, but local capacity, coherent strategy, trust, infrastructure, budgeting, and evidence of return remained major constraints.
The review reports more than 1,600 state AI bills proposed since 2019, with 77% categorized as controlling legislation; at least seven states with sandbox or pilot structures; and 12 city use cases across permitting, employee productivity, public safety, resident services, public engagement, and infrastructure analytics. Interviews characterized actual usage as pragmatic and modest, concentrated in drafting, translation, summarization, and basic automation, while cities reported shortages of technical talent, infrastructure, project-scoping capacity, and clear ROI.
This is a useful operating-model snapshot of why visible experimentation does not automatically become durable public capability. It also highlights the distinct paths of states and cities and the potential role of universities, professional associations, and shared institutions in filling capacity gaps.
State shared platforms can provide secure model access, common data services, evaluation, and reusable components, while cities may need narrower service integrations and external capacity partners. Architecture decisions must align with budget, staffing, data readiness, and the ability to operate and monitor systems after pilots end.
Pair guardrails with an affirmative portfolio strategy, use sandboxes to generate reusable evidence and operating standards, fund implementation capacity, and establish partnerships with universities and civic institutions. Track which pilots graduate, stop, or remain experimental and why.
Walled gardens reduce uncontrolled use only if identity, data boundaries, approved models, logging, evaluation, and incident response are centrally operated. Local partnerships require clear data stewardship, access, confidentiality, intellectual-property, and exit responsibilities.
The review describes its sample as representative but not exhaustive. Much of the evidence is qualitative, many referenced projects predate publication, legislation counts do not measure implementation, and the 12 city cases were partly selected for scale ambitions and news coverage.
A meta-analysis of 31 independent experimental, quasi-experimental, and correlational studies published from 2020 through 2026 found a moderate, statistically significant association between GenAI use and student creativity in art and design education, with an overall standardized effect of d = 0.61.
A meta-analysis of 31 independent experimental, quasi-experimental, and correlational studies published from 2020 through 2026 found a moderate, statistically significant association between GenAI use and student creativity in art and design education, with an overall standardized effect of d = 0.61.
The analysis found the largest effect among graduate students, followed by undergraduates and K-12 learners; the strongest creativity dimension was originality or innovation; and text-to-image tools produced larger effects than LLMs, GAN or deep-learning tools, and hybrid tools. The authors report no significant moderating effect from study type or cultural context and no significant publication bias in their analyses.
This is a more specific and defensible educational effectiveness claim than broad assertions that AI improves learning. It identifies a domain and outcome—creative production—where GenAI appears useful while showing that effects vary by learner level, creative dimension, and tool type.
Education platforms should support multimodal creation, artifact provenance, version history, age-appropriate access, and comparison of AI-assisted with independent work. Evaluation should capture the particular creative construct, not use generic engagement or completion metrics as a proxy.
Procurement and curriculum decisions should define the target creative outcome, comparison condition, allowed assistance, attribution expectations, accessibility requirements, and how teachers assess student contribution. Preserve non-AI practice where independent skill development is an instructional objective.
Protect student prompts and artwork, clarify intellectual-property and model-training rights, apply age-appropriate content controls, and ensure creative tools do not expose students to unsafe generation or public-by-default sharing.
The synthesis is limited to art and design education and combines experimental, quasi-experimental, and correlational evidence. The public abstract does not expose all heterogeneity and study-quality statistics, effects were smaller for K-12 students than graduate students, and a moderate average effect does not predict results for a particular curriculum or tool.
How to read this briefing
This edition prioritizes primary government material, public audits, independent research, and relevant public-sector association guidance available for theSeptember 1, 2026 run. Every surfaced item remains in the All view and keeps its original source.
Vendor, operator, and association claims are attributed and are not upgraded to independent evidence. Caveats identify self-reporting, bounded pilots, contested findings, and missing outcome measures.