Evidence on adoption, controls, workforce legitimacy, and AI literacy across public services and education
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.
5evidence records
4cross-source patterns
7topic lenses
Edition intelligence
Search the record over time
Enter one or more keywords to search the evidence record.
AI authority should expand by age, capability, and evidence
New York City's policy and Carnegie Mellon's controlled training study both reject undifferentiated access. NYC is pausing broad student-facing GenAI for younger learners while testing tightly supervised high-school uses; CMU found that a short intervention improved some AI competencies but not responsible-use knowledge or overall output analysis. The common lesson is to grant capability in stages and measure the exact skill or outcome expected at each stage.
What age, role, task, supervision, and evidence threshold must be met before this AI system can move from literacy or bounded practice into routine use?
AI literacy is measurable—and not a single competency
CMU's experiment separated model knowledge, responsible-use knowledge, prompting, output analysis, and self-efficacy, finding gains in only some dimensions. Public First's cross-country survey likewise shows that training, permission, approved access, and workflow embedding are distinct conditions. SLED training should be built and evaluated as a competency portfolio rather than counted as course completion.
Which observable knowledge, verification, prompting, judgment, and escalation behaviors must users demonstrate after training?
Legitimacy requires participation and contestability
Sydney's labor dispute shows that a published AI policy may not satisfy workers who want enforceable protections, while the Australian Information Commissioner found that the public often cannot tell whether automated decision-making is used or how to challenge it. NYC's coalition, limited pilots, disability exceptions, and planned public report point toward a stronger operating model: affected people help set boundaries and can see, question, and contest consequential use.
Who is affected by this deployment, what binding rights or review paths do they have, and what must the organization publish so they can exercise those rights?
Approved access must be paired with control evidence
Public First reports that weak organizational access drives personal-account and undisclosed AI use, but the OAIC audit shows that formal authorization alone does not produce transparency. A governed SLED platform therefore needs both a usable safe path and operational evidence: identity, approved tools, data boundaries, decision inventory, logging, review, and plain-language disclosure.
Can employees use a practical approved path, and can leaders prove which tools, data, decisions, and human controls are actually in operation?
New York City Public Schools and New York City Mayor's OfficeNew York City, United States
Largest U.S. school system pauses broad student-facing GenAI while running bounded high-school pilots
New York City announced a one-year moratorium for the 2026-27 school year on student-facing generative AI in grades 2-K through 8, affecting nearly 600,000 students. Companion chatbots are prohibited across all grades. High schools may run five bounded pilot types for no more than 50,000 students, all under trained-educator supervision, while all high-school students receive two 45-minute AI critical-thinking modules.
Standards or public-body guidanceEmergingK-12 public education
Read full analysis
What happened
New York City announced a one-year moratorium for the 2026-27 school year on student-facing generative AI in grades 2-K through 8, affecting nearly 600,000 students. Companion chatbots are prohibited across all grades. High schools may run five bounded pilot types for no more than 50,000 students, all under trained-educator supervision, while all high-school students receive two 45-minute AI critical-thinking modules.
Evidence read
The city's official announcement specifies the affected enrollment, duration, pilot ceiling, named tools, classroom time limits, supervision, and evaluation plan. It also permits teachers to use compliant AI for planning and operations and creates exceptions for assistive technology, multilingual learners, and career-readiness programs. This is a policy intervention, not evidence that the moratorium or pilots improve learning or safety.
Why it matters for SLED
This is the most consequential U.S. district-level reset yet from broad adoption toward age- and capability-specific governance. It distinguishes employee augmentation, AI literacy, assistive technology, supervised learning, and open-ended student-facing AI instead of applying one rule to every use.
Architecture implications
District AI gateways and device controls need age-, role-, course-, and accommodation-aware policy enforcement. Approved pilots should be isolated from unrestricted consumer systems, configured to preserve the student as primary thinker, instrumented for time and use limits, and connected to educator supervision and escalation. Companion behavior should be blocked independently of ordinary tutoring capability.
Governance implications
Use a portfolio approach: pause unbounded uses, preserve accessibility exceptions, provide universal literacy, and run limited pilots with predefined learning, safety, privacy, and user-feedback measures. Procurement reviews should cover vendor transparency, learning design, instructional impact, prior evidence, and continuous feedback rather than privacy and security alone.
Security and privacy implications
Apply student-data minimization, training-data restrictions, identity and age controls, conversation retention limits, educator access boundaries, safety monitoring, and incident escalation. Exceptions for disabilities and multilingual learners need equivalent privacy protections and must not become a path to secondary use or disproportionate surveillance.
Limits of the evidence
The policy was announced before implementation and supplies no outcome data. The city's characterization of its review as exhaustive is not independently verified, the named pilots are not evaluated in the announcement, and a one-year moratorium could delay beneficial uses as well as risky ones. Accessibility exceptions require careful implementation to avoid unequal access or stigma.
What happened
New York City announced a one-year moratorium for the 2026-27 school year on student-facing generative AI in grades 2-K through 8, affecting nearly 600,000 students. Companion chatbots are prohibited across all grades. High schools may run five bounded pilot types for no more than 50,000 students, all under trained-educator supervision, while all high-school students receive two 45-minute AI critical-thinking modules.
Evidence read
The city's official announcement specifies the affected enrollment, duration, pilot ceiling, named tools, classroom time limits, supervision, and evaluation plan. It also permits teachers to use compliant AI for planning and operations and creates exceptions for assistive technology, multilingual learners, and career-readiness programs. This is a policy intervention, not evidence that the moratorium or pilots improve learning or safety.
Why it matters for SLED
This is the most consequential U.S. district-level reset yet from broad adoption toward age- and capability-specific governance. It distinguishes employee augmentation, AI literacy, assistive technology, supervised learning, and open-ended student-facing AI instead of applying one rule to every use.
Architecture implications
District AI gateways and device controls need age-, role-, course-, and accommodation-aware policy enforcement. Approved pilots should be isolated from unrestricted consumer systems, configured to preserve the student as primary thinker, instrumented for time and use limits, and connected to educator supervision and escalation. Companion behavior should be blocked independently of ordinary tutoring capability.
Governance implications
Use a portfolio approach: pause unbounded uses, preserve accessibility exceptions, provide universal literacy, and run limited pilots with predefined learning, safety, privacy, and user-feedback measures. Procurement reviews should cover vendor transparency, learning design, instructional impact, prior evidence, and continuous feedback rather than privacy and security alone.
Security and privacy implications
Apply student-data minimization, training-data restrictions, identity and age controls, conversation retention limits, educator access boundaries, safety monitoring, and incident escalation. Exceptions for disabilities and multilingual learners need equivalent privacy protections and must not become a path to secondary use or disproportionate surveillance.
Limits of the evidence
The policy was announced before implementation and supplies no outcome data. The city's characterization of its review as exhaustive is not independently verified, the named pilots are not evaluated in the announcement, and a one-year moratorium could delay beneficial uses as well as risky ones. Accessibility exceptions require careful implementation to avoid unequal access or stigma.
Carnegie Mellon University Eberly CenterPennsylvania, United States
Randomized university study finds 90-minute AI literacy modules improve some competencies but not critical analysis
A randomized study assigned 1,368 undergraduate and graduate students in 53 courses taught by 46 instructors to either no intervention or four self-paced modules totaling about 90 minutes. The modules significantly improved knowledge of how LLMs work, prompting skill, and self-efficacy beyond the control group, but did not significantly improve responsible-use knowledge or overall skill at analyzing AI output.
Academic researchMixedPublic-interest higher education and workforce preparation
Read full analysis
What happened
A randomized study assigned 1,368 undergraduate and graduate students in 53 courses taught by 46 instructors to either no intervention or four self-paced modules totaling about 90 minutes. The modules significantly improved knowledge of how LLMs work, prompting skill, and self-efficacy beyond the control group, but did not significantly improve responsible-use knowledge or overall skill at analyzing AI output.
Evidence read
Courses were randomly assigned; 610 students were in control and 758 in treatment. Pre/post measures covered knowledge, authentic prompting and output-evaluation tasks, and self-efficacy. A blinded subset of 174 students was scored by two independent raters. Benefits were reported across discipline, sex, race or ethnicity, class year, and first-generation status, with a pre-existing female self-efficacy gap closing after the intervention.
Why it matters for SLED
SLED organizations often treat one training completion as evidence of AI readiness. This study shows that scalable instruction can work, but that verification, responsible use, and model knowledge are separable competencies requiring different learning designs and assessments.
Architecture implications
A learning platform can deliver reusable foundational modules at scale, but should also support authenticated completion, versioned content as models change, authentic practice, immediate feedback, accessibility, and role-specific follow-on exercises. Training telemetry should remain separate from permission to access sensitive data or consequential tools.
Governance implications
Define AI literacy as multiple testable capabilities, require demonstrated verification and responsible-use skills for higher-risk access, and refresh training when models, data rules, or workflows change. Procurement should not accept seat time or completion rates as proof that users can evaluate outputs or protect data.
Security and privacy implications
Training should use synthetic or approved data and explicitly test source checking, sensitive-data boundaries, escalation, and uncertainty. Demographic analysis can expose inequitable effects, but the data needed for it should be governed with minimization, access, and retention controls.
Limits of the evidence
The study occurred at one selective university with instructors who volunteered their courses, measured outcomes four days after access, and does not establish durable behavior change or safer real-world AI use. The output-analysis measure used a 174-student subset, and the intervention produced no detected gain in responsible-use knowledge or overall output analysis.
What happened
A randomized study assigned 1,368 undergraduate and graduate students in 53 courses taught by 46 instructors to either no intervention or four self-paced modules totaling about 90 minutes. The modules significantly improved knowledge of how LLMs work, prompting skill, and self-efficacy beyond the control group, but did not significantly improve responsible-use knowledge or overall skill at analyzing AI output.
Evidence read
Courses were randomly assigned; 610 students were in control and 758 in treatment. Pre/post measures covered knowledge, authentic prompting and output-evaluation tasks, and self-efficacy. A blinded subset of 174 students was scored by two independent raters. Benefits were reported across discipline, sex, race or ethnicity, class year, and first-generation status, with a pre-existing female self-efficacy gap closing after the intervention.
Why it matters for SLED
SLED organizations often treat one training completion as evidence of AI readiness. This study shows that scalable instruction can work, but that verification, responsible use, and model knowledge are separable competencies requiring different learning designs and assessments.
Architecture implications
A learning platform can deliver reusable foundational modules at scale, but should also support authenticated completion, versioned content as models change, authentic practice, immediate feedback, accessibility, and role-specific follow-on exercises. Training telemetry should remain separate from permission to access sensitive data or consequential tools.
Governance implications
Define AI literacy as multiple testable capabilities, require demonstrated verification and responsible-use skills for higher-risk access, and refresh training when models, data rules, or workflows change. Procurement should not accept seat time or completion rates as proof that users can evaluate outputs or protect data.
Security and privacy implications
Training should use synthetic or approved data and explicitly test source checking, sensitive-data boundaries, escalation, and uncertainty. Demographic analysis can expose inequitable effects, but the data needed for it should be governed with minimization, access, and retention controls.
Limits of the evidence
The study occurred at one selective university with instructors who volunteered their courses, measured outcomes four days after access, and does not establish durable behavior change or safer real-world AI use. The output-analysis measure used a 174-student subset, and the intervention produced no detected gain in responsible-use knowledge or overall output analysis.
Public First and Center for Data InnovationTen countries: Brazil, France, Germany, India, Japan, Saudi Arabia, Singapore, South Africa, United Kingdom, and United States
Ten-country survey links effective public-sector AI use to approved access, clear rules, training, and workflow embedding
A survey of 3,335 public servants across ten countries reports that 74% use AI, yet only 18% think government uses it very effectively. The study separates enthusiasm, education, enablement, empowerment, and workflow embedding and finds large associations between those conditions and confidence, advanced use, and reported benefits.
Independent researchMixedGovernment workforce
Read full analysis
What happened
A survey of 3,335 public servants across ten countries reports that 74% use AI, yet only 18% think government uses it very effectively. The study separates enthusiasm, education, enablement, empowerment, and workflow embedding and finds large associations between those conditions and confidence, advanced use, and reported benefits.
Evidence read
In low-enablement organizations, 64% of enthusiastic users reported personal-login use and 70% reported work use unknown to managers. Across countries, 50% cited data security or privacy as a barrier. In high-empowerment environments, 91% reported confidence compared with 45% in low-empowerment settings; in high-embedding environments, 58% of workers aged 55 or older reported saving more than an hour with AI versus 16% in low-embedding environments.
Why it matters for SLED
The survey provides a cross-national operating-model lens for employee copilots and assistants. It suggests that access and governance are complements: withholding practical approved tools can push work into personal accounts, while clear permission, support, and embedded enterprise access can spread benefits beyond already-confident users.
Architecture implications
Provide an identity-bound enterprise AI layer integrated into ordinary workflows, with approved models, data classifications, support, logging, and clear escalation. Hybrid or cloud choices should be driven by data sensitivity and integration needs, but an official platform must be usable enough to compete with personal accounts.
Governance implications
Pair safe-harbor rules for low-risk tasks with stronger review for sensitive data and consequential actions. Treat onboarding, role-specific training, support channels, manager visibility, and workflow redesign as parts of deployment. Track public-service outcomes separately from confidence, adoption, or reported time savings.
Security and privacy implications
Reduce shadow AI through managed identities, enterprise contracts, DLP, approved connectors, auditable use, and practical guidance on what data can be shared. Monitor for work performed through personal accounts without treating surveillance of employees as a substitute for usable approved tools and trust.
Limits of the evidence
The index is based on self-reported cross-sectional survey data and shows association, not causation. Public First produced it for the Center for Data Innovation with Google sponsorship. Country samples, job roles, public-sector definitions, and cultural response patterns may differ, and perceived benefit or time saved is not independently measured mission impact.
What happened
A survey of 3,335 public servants across ten countries reports that 74% use AI, yet only 18% think government uses it very effectively. The study separates enthusiasm, education, enablement, empowerment, and workflow embedding and finds large associations between those conditions and confidence, advanced use, and reported benefits.
Evidence read
In low-enablement organizations, 64% of enthusiastic users reported personal-login use and 70% reported work use unknown to managers. Across countries, 50% cited data security or privacy as a barrier. In high-empowerment environments, 91% reported confidence compared with 45% in low-empowerment settings; in high-embedding environments, 58% of workers aged 55 or older reported saving more than an hour with AI versus 16% in low-embedding environments.
Why it matters for SLED
The survey provides a cross-national operating-model lens for employee copilots and assistants. It suggests that access and governance are complements: withholding practical approved tools can push work into personal accounts, while clear permission, support, and embedded enterprise access can spread benefits beyond already-confident users.
Architecture implications
Provide an identity-bound enterprise AI layer integrated into ordinary workflows, with approved models, data classifications, support, logging, and clear escalation. Hybrid or cloud choices should be driven by data sensitivity and integration needs, but an official platform must be usable enough to compete with personal accounts.
Governance implications
Pair safe-harbor rules for low-risk tasks with stronger review for sensitive data and consequential actions. Treat onboarding, role-specific training, support channels, manager visibility, and workflow redesign as parts of deployment. Track public-service outcomes separately from confidence, adoption, or reported time savings.
Security and privacy implications
Reduce shadow AI through managed identities, enterprise contracts, DLP, approved connectors, auditable use, and practical guidance on what data can be shared. Monitor for work performed through personal accounts without treating surveillance of employees as a substitute for usable approved tools and trust.
Limits of the evidence
The index is based on self-reported cross-sectional survey data and shows association, not causation. Public First produced it for the Center for Data Innovation with Google sponsorship. Country samples, job roles, public-sector definitions, and cultural response patterns may differ, and perceived benefit or time saved is not independently measured mission impact.
Office of the Australian Information CommissionerAustralia
Australian audit finds automated decision-making authority far more visible than actual government use
The Australian Information Commissioner reviewed the websites, AI transparency statements, and publication plans of 23 federal agencies authorized by law to use automated decision-making. Only four agencies, 17%, disclosed in their publication-scheme information that they used ADM in decisions affecting the public; nine more referenced ADM without confirming use, and ten did not mention it.
Government auditCautionaryFederal public administration and digital services
Read full analysis
What happened
The Australian Information Commissioner reviewed the websites, AI transparency statements, and publication plans of 23 federal agencies authorized by law to use automated decision-making. Only four agencies, 17%, disclosed in their publication-scheme information that they used ADM in decisions affecting the public; nine more referenced ADM without confirming use, and ten did not mention it.
Evidence read
The regulator used defined website search procedures and external evidence to test what a member of the public could reasonably discover. It found examples in which agencies described what AI did not do while failing to state what automation did do. The report recommends disclosing statutory authority, actual use, decision types, examples, and governing policies.
Why it matters for SLED
The review is newly relevant as Australian public-sector workers press for stronger, enforceable automated-decision safeguards. It shows why a public AI inventory must cover rules engines, calculators, machine learning, and recommendations—not only systems labeled AI—and why affected people need plain-language notice and review rights.
Architecture implications
Maintain a decision-system register linked to statutory authority, service, inputs, business rules or model, human review, appeal path, vendor, and deployed version. Public disclosures should be generated from the same operational inventory used for access control, monitoring, change management, and incident response.
Governance implications
Require plain-language notice, examples, and review paths for systems that affect rights or interests. Inventory all automated decisions regardless of marketing label, assign accountable owners, and make change review and public disclosure part of deployment rather than an after-the-fact communications task.
Security and privacy implications
Record data provenance, accuracy checks, personal-information use, role access, retention, and the influence of each automated output on a final decision. Transparency should enable contestability without exposing security-sensitive details or creating new privacy risks.
Limits of the evidence
This January report is reused because September 2 workforce pressure made its findings newly relevant; it is not newly published evidence. The desktop review assessed public discoverability, not system accuracy, legality, fairness, or even confirmed use in every authorized agency. Authority to automate does not prove that an agency actually automated decisions.
What happened
The Australian Information Commissioner reviewed the websites, AI transparency statements, and publication plans of 23 federal agencies authorized by law to use automated decision-making. Only four agencies, 17%, disclosed in their publication-scheme information that they used ADM in decisions affecting the public; nine more referenced ADM without confirming use, and ten did not mention it.
Evidence read
The regulator used defined website search procedures and external evidence to test what a member of the public could reasonably discover. It found examples in which agencies described what AI did not do while failing to state what automation did do. The report recommends disclosing statutory authority, actual use, decision types, examples, and governing policies.
Why it matters for SLED
The review is newly relevant as Australian public-sector workers press for stronger, enforceable automated-decision safeguards. It shows why a public AI inventory must cover rules engines, calculators, machine learning, and recommendations—not only systems labeled AI—and why affected people need plain-language notice and review rights.
Architecture implications
Maintain a decision-system register linked to statutory authority, service, inputs, business rules or model, human review, appeal path, vendor, and deployed version. Public disclosures should be generated from the same operational inventory used for access control, monitoring, change management, and incident response.
Governance implications
Require plain-language notice, examples, and review paths for systems that affect rights or interests. Inventory all automated decisions regardless of marketing label, assign accountable owners, and make change review and public disclosure part of deployment rather than an after-the-fact communications task.
Security and privacy implications
Record data provenance, accuracy checks, personal-information use, role access, retention, and the influence of each automated output on a final decision. Transparency should enable contestability without exposing security-sensitive details or creating new privacy risks.
Limits of the evidence
This January report is reused because September 2 workforce pressure made its findings newly relevant; it is not newly published evidence. The desktop review assessed public discoverability, not system accuracy, legality, fairness, or even confirmed use in every authorized agency. Authority to automate does not prove that an agency actually automated decisions.
University of Sydney and National Tertiary Education UnionNew South Wales, Australia
AI safeguards become a bargaining issue as roughly 2,000 university staff strike
About 2,000 University of Sydney staff joined a 24-hour strike amid enterprise bargaining disputes involving AI protections, workload fairness, and job security. The union sought enforceable safeguards in the employment agreement; the university said it supported many objectives but preferred to govern AI through institutional policies and maintained that the strike was premature.
About 2,000 University of Sydney staff joined a 24-hour strike amid enterprise bargaining disputes involving AI protections, workload fairness, and job security. The union sought enforceable safeguards in the employment agreement; the university said it supported many objectives but preferred to govern AI through institutional policies and maintained that the strike was premature.
Evidence read
Independent reporting documented the strike, the positions of the union and university, and disruption to classes. A faculty internal survey reported zero agreement with a broad trust statement, while roughly 500 people were reported at campus entrances. The action involved several issues, so the evidence does not isolate AI as the sole cause.
Why it matters for SLED
This is direct evidence that workforce participation is becoming a deployment dependency in public higher education. Policies developed through consultation may still lack legitimacy when workers believe they are revocable, do not govern workload and role redesign, or cannot be enforced through employment arrangements.
Architecture implications
AI implementation plans should expose how systems change task allocation, staffing, monitoring, intellectual property, and performance measurement. Usage telemetry must be designed with clear purpose limits and worker access rules; it should not quietly become productivity surveillance or automated evaluation.
Governance implications
Engage unions, faculty governance, students, accessibility offices, HR, and academic leaders before deployment decisions harden. Decide which safeguards belong in durable agreements, which belong in adaptable policy, how disputes are resolved, and how productivity gains translate into workload, staffing, and service-quality commitments.
Security and privacy implications
Workforce AI contracts and policies should define whether employee prompts, teaching materials, research, or student interactions train models; who can inspect logs; how monitoring data is retained; and whether outputs influence performance, discipline, promotion, or redundancy decisions.
Limits of the evidence
The report covers an active labor dispute, not an adjudicated finding of unsafe AI use. AI was one of multiple bargaining and trust issues, attendance estimates were reported rather than independently audited, and the internal trust result came from one faculty and a broadly worded statement. No AI system performance or educational outcome was evaluated.
What happened
About 2,000 University of Sydney staff joined a 24-hour strike amid enterprise bargaining disputes involving AI protections, workload fairness, and job security. The union sought enforceable safeguards in the employment agreement; the university said it supported many objectives but preferred to govern AI through institutional policies and maintained that the strike was premature.
Evidence read
Independent reporting documented the strike, the positions of the union and university, and disruption to classes. A faculty internal survey reported zero agreement with a broad trust statement, while roughly 500 people were reported at campus entrances. The action involved several issues, so the evidence does not isolate AI as the sole cause.
Why it matters for SLED
This is direct evidence that workforce participation is becoming a deployment dependency in public higher education. Policies developed through consultation may still lack legitimacy when workers believe they are revocable, do not govern workload and role redesign, or cannot be enforced through employment arrangements.
Architecture implications
AI implementation plans should expose how systems change task allocation, staffing, monitoring, intellectual property, and performance measurement. Usage telemetry must be designed with clear purpose limits and worker access rules; it should not quietly become productivity surveillance or automated evaluation.
Governance implications
Engage unions, faculty governance, students, accessibility offices, HR, and academic leaders before deployment decisions harden. Decide which safeguards belong in durable agreements, which belong in adaptable policy, how disputes are resolved, and how productivity gains translate into workload, staffing, and service-quality commitments.
Security and privacy implications
Workforce AI contracts and policies should define whether employee prompts, teaching materials, research, or student interactions train models; who can inspect logs; how monitoring data is retained; and whether outputs influence performance, discipline, promotion, or redundancy decisions.
Limits of the evidence
The report covers an active labor dispute, not an adjudicated finding of unsafe AI use. AI was one of multiple bargaining and trust issues, attendance estimates were reported rather than independently audited, and the internal trust result came from one faculty and a broadly worded statement. No AI system performance or educational outcome was evaluated.
This edition prioritizes primary government material, public audits, independent research, and relevant public-sector association guidance available for theSeptember 2, 2026 run. Every surfaced item remains in the All view and keeps its original source.
Evidence classes
Government evaluation
A public body’s measured evaluation or documented pilot.
Government audit
An oversight review of performance, controls, or operations.
Academic research
Research produced through an academic institution or peer-reviewed venue.
Independent research
Research conducted outside the implementing organization.
Public-sector association guidance
Practitioner guidance or an association-supplied case; not independent outcome evidence.
Independent reporting
Independent reporting with attributable sources but without a formal evaluation design.
Standards or public-body guidance
Normative or advisory guidance from a standards body or public institution.
Vendor claim
A supplier-provided assertion that has not been upgraded to independent evidence.
Outcome labels
Effective
Evidence supports a useful result within the tested scope.
Mixed
Benefits and material limitations appear together.
Cautionary
The record surfaces failure, risk, or a control gap.
Emerging
A developing practice or claim without measured outcomes.
Claims discipline
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.