The Fragile Edge of Innovation: Examining Policy Resilience in State-Level K-12 Artificial Intelligence Guidance

Factors Driving Policy Resilience and Fragility

To operationalize the concepts of policy resilience and fragility within the domain of AI education, this study posits three interacting factors that serve as primary drivers of systemic stability or collapse during reform cycles: Resource Sustainability, Coherence of Policy Signals, and Distributed Organizational Capacity. These factors determine whether a system absorbs new priorities (resilience) or whether the reform accelerates faster than the supporting infrastructure (fragility). These factors also provide insight on how AI education policies could potentially reproduce systemic inequalities and capacity gaps. We conceptualize these drivers in Table 1 to illustrate the theoretical distinction between policy resilience and policy fragility.

OPEN IN VIEWER

Table 1.

Drivers of Policy Resilience and Fragility.

Factor	Policy Resilience	Policy Fragility
Resource Sustainability	The shift from initial, often external, soft funding (grants) to integrated, predictable, and diverse funding sources.	Reliance on volatile or soft funding (i.e., time-limited grants) that supports implementation but does not cover the necessary recurring maintenance costs.
Coherence of Policy Signals	The degree to which new policy is integrated into existing standards, curriculum frameworks, and assessment systems rather than being treated as an add-on.	The introduction of new policies that compete for instructional time, teacher attention, and budget resources without eliminating or consolidating old policies.
Distributed Organizational Capacity	The widespread ability for all levels of the system, from state agencies to school leaders and educators, to understand and execute the policy effectively.	Capacity (knowledge, training, available teachers) is concentrated only at the state level or within a few lighthouse districts, leaving much of the system unprepared.

The interplay of these factors creates a dynamic system, where reforms are often initially fragile due to newness, but can become resilient through strategic institutionalization. Conversely, a previously stable program can become fragile due to external pressures or resource changes. Table 2 includes characteristic examples of systems that have moved from being policy fragile to policy resilient and vice versa.

OPEN IN VIEWER

Table 2.

Movement Between Policy Fragility and Resilience.

	Resource Sustainability	Coherence of Policy Signals	Examples of Distributed Organizational Capacity
Fragility → Resilience	Reliance on external champions and temporary funds → protected by standard operating procedures, PD, and recurring general funds.	The policy moves from being an elective or “add on” for some students to being an expectation for all students in different learning environments and courses.	Inconsistent or sporadic PD → ongoing PD opportunities One time funding to purchase technology → sustained funding for ongoing technology and support
Resilience → Fragility	External resource exhaustion: A non-sustainable element (e.g. funding, dedicated staff time) is unexpectedly removed or exhausted, causing the implementation to unravel.	A once-integrated AI framework becomes marginalized by new, competing legislative mandates or initiatives that divert instruction time and focus without sunsetting old requirements.	Required and/or ongoing PD for all teachers → Optional or occasional PD (PD doesn’t keep up with tech updates) Grant funds for PD and materials → No dedicated funding for ongoing support

The CAPE Framework has been widely employed in computing education research to analyze systemic conditions affecting equity, sustainability, and policy alignment, making it a complement to policy fragility and resilience. CAPE’s four domains (Capacity, Access, Participation, and Experience) serve as the empirical indicators that translate the abstract drivers of resilience into measurable policy components. As a policy design tool, CAPE has highlighted the impact of policy implementation and influence on capacity building efforts (e.g. funding, regulatory approaches, teacher capacity). To operationalize policy resilience/fragility framework in AI education, CAPE is used as a deductive codebook for content analysis.

We propose that the CAPE domains serve as indicators for evaluating AI policy resilience or fragility. In this framework, these are not parallel dimensions but are functionally linked: policy resilience is evidenced when gains in access and participation are matched by a corresponding strengthening of capacity and experience. Conversely, policy fragility occurs through domain fragmentation, where aspirations for modernization, specifically in access, advance more rapidly than the institutional capacity or organizational routines required to sustain them (Table 3). By using CAPE as a deductive framework, we can prioritize exactly which domain is lagging, thereby identifying the specific fracture point of a fragile policy.

OPEN IN VIEWER

Table 3.

Aligning Policy Resilience and Fragility with the CAPE Framework.

	Policy Resilience	Policy Fragility
Capacity	Coordinated resource support including educator professional development, standards and aligned curriculum, policy with commensurate financial backing	Unfunded policy mandates or one-time infusion of funds (time-limited grant), insufficient teacher professional development, curriculum misalignment
Access	Access is open and available for all students, without reliance on specific technologies or prerequisites.	New resources (tech, personnel) are rolled out only where infrastructure exists
Participation	All students participate in an aligned pathway	Early participants excel as funding is readily available, whereas marginalized groups or late joiners lose access as funding or priorities shift
Experience	High quality teachers and high-quality instructional materials that meet the needs of diverse learners	Inconsistent experiences that rely on educator motivation and self-preparation; experience depends on the resources of the individual school; curriculum demotivates rather than inspires students to pursue additional study.

Data Sources and Sampling

The dataset consisted of AI guidance documents from the thirty-five U.S. states that had publicly published resources addressing AI in education from January 2024 to December 2025 (Table 4). Official publication was determined by the document’s availability on an SEA website, SEA identification within the text, or its formal announcement through official state communication channels. A total of 994 pages of documents and four web pages were coded and analyzed as text, with state documents ranging from 2 to 81 pages. Web pages were treated as formally completed documents provided they were hosted on official state domains. This constituted a census of eligible documents and categorized as a purposive sampling strategy based on specific inclusion and the availability of official, authoritative policy texts (Mason, 2018). A document was included if it (a) was issued or endorsed by an SEA or equivalent body, defined as documents bearing the official agency seal, authored by state-appointed task forces, or explicitly linked as a primary resource on the SEA’s landing page, (b) explicitly referenced AI or generative AI in relation to teaching and learning, and (c) articulated expectations, recommendations, or parameters for school or district action. Documents that served as secondary support but were not formally sanctioned or branded by the SEA were excluded. The primary unit of analysis was the entire guidance document, while the coding unit was any sentence or paragraph that addressed a conceptual theme.

OPEN IN VIEWER

Table 4.

States Included in Data Set.

Alabama	Alaska	Arizona	California	Colorado
Connecticut	Delaware	Florida	Georgia	Hawaii
Idaho	Indiana	Kentucky	Louisiana	Maine
Massachusetts	Minnesota	Mississippi	Missouri	Montana
Nevada	New Hampshire	New Mexico	North Carolina	North Dakota
Ohio	Oklahoma	Oregon	Rhode Island	Utah
Virginia	Washington	West Virginia	Wisconsin	Wyoming

The strategic choice to analyze officially published, state-endorsed guidance documents placed priority on authoritative, publicly accessible sources that directly influenced local decision-making and communicated official policy positions. This focus on “as written” guidance documents ensured that the analysis captured the formal, articulated expectations and parameters set by the SEA. By limiting the sample to formally published documents, the methodology adhered to the widely accepted principle of analyzing texts that possess institutional weight and communicative function in the policy ecosystem (Childs & Russell, 2017; Russell et al., 2015). This approach is crucial to policy research, as it offers a robust foundation for understanding the intended scope and official direction of AI integration in public education as opposed to preliminary drafts or non-authoritative internal communications.

Analytical Approach

We posit that CAPE is an effective analytical tool for understanding state-level AI education through a policy resilience and policy fragility lens. Capacity, access, participation, and experience as the four dimensions within CAPE have been applied in examining cross-state CS policy adoption and computing initiatives (Marshall et al., 2025; Garvin et al., 2024) and analyses of systemic inequities in access and computing pathways (McGill et al., 2022). Though originally conceptualized for understanding equity in computing education, CAPE was expanded to encompass policy formation (Warner et al., 2022).

This study utilized a directed content analysis approach (Hsieh & Shannon, 2005) which guided the systematic transformation of raw text into theoretical constructs. In this approach, we used the four CAPE domains as a foundational codebook to organize state guidance: (a) Capacity: Definitions of AI literacy for preparing educators including references to teacher knowledge, PD, institutional readiness, available teachers, and system-level support for responsible AI integration; (b) Access: Descriptions of how AI learning opportunities are allocated, scheduled, staffed or resourced, and any considerations for addressing inequalities across districts or student populations; (c) Participation: Expectations regarding who participates in AI learning, mechanisms for monitoring engagement, and whether AI instruction is positioned as universal or elective; and (d) Experience: Attention to ethical, social, and developmental implications of AI; considerations of data privacy, bias, or transparency; and descriptions of pedagogical or instructional use cases.

Our use of deductive coding aligned with recognized standards for rigorous content analysis, with the goal of evaluating textual data against our conceptual framework (Saldaña, 2021). This methodological approach has been employed in cross-state document-based education policy studies, such as cross-state reviews of STEM and CS standards (CSTA and IACE, 2024; Ekiz-Kiran & Aydin-Gunbatar, 2021) and analyses of district AI or digital learning policies (Eutsler et al., 2025).

A multi-stage analytic strategy was used to strengthen credibility and dependability by ensuring both within-case depth and cross-case comparability. First, documents were reviewed holistically to identify relevant excerpts and classify them into CAPE-aligned categories. Second, cross-case pattern analysis was conducted to identify areas of convergence, divergence, and variation across states. The purpose of the analysis was descriptive rather than evaluative; no scoring, ranking, or normative judgements were applied to states’ guidance documents.

Through the research team’s reflective conversations, it became clear that the AI guidance documents revealed specific policy signals that guide implementation toward either exploring innovative educational opportunities or reinforcing existing legal protections for privacy, integrity, and procurement. After this observation, the research team created a mutually exclusive classification system based on the predominance of policy signals holistically within each document defining caution, opportunity, and balance, which were then applied during a third round of coding.

States were categorized as caution, a reactive posture, when the text focused primarily on risk mitigation, legal compliance, data privacy mandates, and the prohibition of misuse at the expense of pedagogical exploration. Conversely, states were categorized as opportunity when the guidance adopted a proactive posture, framing AI as a catalyst for student agency, instructional integration, and systemic capacity building. This framing often emerges during periods of rapid advancement, when policies or practices are viewed as a panacea to address societal issues or early periods of innovation. For states displaying proportional parity between these two poles, we applied the balanced category, representing a dual-priority approach that simultaneously establishes rigorous guardrails, such as data privacy and ethical oversight, alongside promotion of pedagogical innovation. By utilizing these distinct categories, we provide a critical lens for exploring the spectrum of policy resilience and fragility within AI education.

To mitigate bias and enhance confirmability, a second researcher independently coded a 60% subsample of documents. Codes were documented, cross-checked, and iteratively refined. Coding decisions and discrepancies were discussed through reflexive conversations. This inter-rater discussion informed the interactive refinement of the code definitions and ensured a shared understanding of code application across all documents. In instances of coding divergence, the team engaged in a systematic reconciliation process where disagreements served as a catalyst for refining code boundaries, thereby enhancing the internal consistency and conceptual validity of the analysis. This reflexive approach emphasized transparency and trustworthiness in interpretive analyses of the guidance documents.

Limitations

Several limitations should be considered when interpreting the findings of this study. First, the scope of inquiry is delimited to official state-level guidance. Consequently, the analysis does not capture the nuances of district-level policy development or the localized implementation of these frameworks within specific school contexts. Second, this study represents a snapshot of an exceptionally fluid policy environment, reflecting a specific moment in the rapid evolution of AI education policy rather than a stabilized institutional landscape. Third, the inherent variability in the scope, length, and specificity of the source documents, ranging from broad statements of principle to exhaustive technical handbooks, necessitates caution when drawing comparative inferences based solely on code frequency. This variation, while a finding in itself regarding policy fragmentation, impacts the direct comparability of emphasis across different state contexts. Finally, given the accelerated pace of technological and regulatory updates, some states may have revised their guidance during or immediately following the data collection window, meaning certain findings may not reflect the most current iterations of state-level policy.

Despite these limitations, cross-state document analysis offers a critical first look at how states are interpreting federal AI priorities and transplanting them into educational policy. The approach provides a systematic, theoretically grounded foundation for understanding how early AI guidance positions educators, students, and systems within the rapidly shifting landscape of AI education.

Capacity

Analysis of state AI guidance documents revealed fragmented policy architecture, where capacity is conceptualized as a series of disconnected compliance tasks rather than a cohesive ecosystem of support. These varied signals distribute attention across risk management and pedagogical enablement, reflecting inconsistent orientations towards implementation. Thirteen state guidance documents adopted a primarily reactive stance, providing reminders on the legal expectations for protecting student privacy and procuring technology, functioning more akin to a risk management checklist. For example, Alabama provided a list of conditions to minimize risk such as complying with the National Institute of Standards and Technology (NSIT) AI Risk Management Framework, conducting an annual audit to ensure compliance, and maintaining a risk register that records AI systems’ relative levels of risk (low, medium, and high). In contrast, twelve proactive guidance documents emphasized ways to leverage the potential of AI to enhance the learning environment. Nevada, for example, celebrates the careful use of AI as “increasingly prevalent in students’ current education experience and in their future professional environments. . .The power of AI tools for education, community engagement and deeper learning will continue to drive innovation and policy” (Nevada Department of Education, 2025, p. 5).

Some documents explicitly suggested that districts foster distributed organizational capacity, suggesting forming a committee of a wide range of users. Others outlined specific, role-based expectations for educator groups. Virginia called out roles and responsibilities for SEAs, governing boards, higher education programs, and AI/technology directors in establishing clear policy. Additionally, West Virginia and Mississippi articulated how AI could meet the specific educational goals of students and educators. Often, documents provided guiding questions to prompt districts toward articulating their pedagogical values and intentions for AI use. Ten of the documents took a balanced approach, equally considering risk and benefit. While Alabama’s guidance emphasizes risk mitigation and compliance with national frameworks, the state also maintains that AI in K-12 schools

… will unlock everyone’s potential by catering to individual learning and teaching methods. It will allow students to think critically about AI applications and use them as an aid for research and problem-solving. It will empower our students so they can make significant contributions to society and be change-makers. . .[and] enable educators to cater to the unique learning preferences of each student and guide them to foster a sense of responsibility and empathy toward the ethical use of AI. (Alabama Department of Education & Early Development, 2024, p. 6)

Analysis revealed the underdevelopment of explicit pedagogical capacity and lack of clear examples for how to apply AI in the teaching and learning environments. States that approached their guidance with a lens of opportunity were more likely to include examples of how AI can enhance education by providing clear, situated examples or curricular models for effective AI application. Guidance from both Rhode Island and Washington suggested that students can use AI as a personalized tutor and feedback mechanism to deepen understanding rather than just producing final products. Arizona, Utah, and Wisconsin are examples of states that suggested school leaders leverage AI to streamline operations and analyze data for strategic decision-making. A specific application includes using AI to optimize master schedules, balancing complex variables such as teacher availability, room assignments, and student course requests much faster than manual methods. Administrators can also use AI-powered analytics to process large datasets, such as attendance records, enrollment figures, or academic performance data, to identify trends, forecast future staffing needs, or flag at-risk students for early intervention. The Florida K-12 AI Education Task Force (2025), for example, suggested using AI for administrative and operational support to “to optimize class schedules, bus routes, and facility usage” and that AI “Anticipates fluctuations in student numbers to guide hiring and budgeting” (n.p.).

Despite many states providing clear, situated examples or curricular models for effective AI application, we found that similar to other education policy efforts, there is an implementation gap that places the burden of educator training entirely on local districts, with little support from state authorities. To address the gap, nearly all documents were supported by links to resources from external organizations such as TeachAI and Digital Promise, or state websites that have training or resources.

Access

While nearly all state guidance documents explicitly called for “all students” to have access to AI learning opportunities, very few included an operational definition of what constitutes “access,” (such as how AI learning opportunities are scheduled, staffed and distributed) or included an implementation strategy to guarantee it. This rhetorical commitment mirrors past EdTech movements, where physical access outpaced instructional support. For example, while the pandemic accelerated device distribution, many districts lacked the training to leverage tools effectively to impact learning outcomes.

When access is discussed, the guidance fragments across three primary equity barriers: digital divide, structural bias in AI, and accommodation. Documents frequently acknowledged the existing digital divide, including lack of internet access or devices, with some states specifically citing disparities of access within rural communities as an area of equitable concern. For example, Colorado estimates nearly 65,000 households “lack consistent access to the internet for educational purposes” (Colorado Department of Education, 2024, p. 15) and Connecticut cautions that “learners without access to technology will fall farther behind in understanding and mastering the use of emerging tools” (Connecticut Commission for Educational Technology, n.d., n.p.).

The second barrier is structural bias in AI. The equity discussion in several states extended beyond technology access to address structural bias embedded within AI systems, emphasizing that equity in access must include caution regarding the algorithms themselves, illustrated by examples in Massachusetts, Connecticut, and Georgia. Finally, access as accommodation was discussed in a small portion of the guidance, including Rhode Island, Mississippi and Colorado, which recommend the use of AI features like scaffolding, speech-to-text, and personalized learning to reach existing instructional goals. This approach is in line with an EdTech’s approach to using tools to learn with AI as opposed to learning about AI (Ng et al., 2021).

Policy recommendations from state guidance documents for addressing access barriers were limited to descriptive tasks rather than prescriptive action. Several states (e.g. North Carolina, Alabama, and Ohio) suggested that districts complete a landscape analysis to better understand where AI is being used, with which tools, and in what ways to support developing policy. More states suggested that this audit focus on auditing access to devices, internet and AI tools to better ensure the digital divide isn’t exacerbated.

The findings within the Access dimension reveal a persistent tension between the rhetorical inclusivity and operational implementation. While state documents adopt the language of “for all,” the guidance remains largely confined to addressing the AI digital divide – the physical presence of devices and connectivity. By focusing on descriptive tasks like landscape audits and technology distribution (devices and/or software tools), states are reverting to an established EdTech playbook that prioritizes physical proximity over the specialized pedagogy required for meaningful intellectual access. This rhetorical-implementation gap signals a state of policy fragility, where the system is technically equipped with tools for adoption, but instructionally unprepared to sustain it equitably.

Participation

Within the CAPE Framework, participation is defined as which students are enrolled in or otherwise receiving CS instruction, typically as a stand-alone course. While the application of AI is frequently conceptualized as a cross-curricular integration across subject areas, the foundational understanding of its underlying mechanics remains rooted in basic CSEd. We adapt this dimension to examine how states define who is engaging with AI and under what conditions, acknowledging this tension between broad usage and specialized technical knowledge.

Most states’ guidance documents focused on acceptable use policies that governed when students may utilize AI and in what manner as it relates to academic integrity. This included disclosure guidance for districts, schools and/or teachers that explicitly defined the acceptable use of AI for assignments and outlined how AI-generated work without proper citations is considered plagiarism (See Missouri, Washington, and West Virginia for examples). Other states, like Utah and Washington, discussed prohibited uses of AI, including the over-reliance on AI that inhibits learning, and the misuse of AI for bullying/harassment (e.g. creating deepfake videos). The strongest guidance across states around misuse was adherence to existing policies and laws, such as data privacy. In particular, states strongly prioritized the protection of personally identifiable information (PII) or confidential data in accordance with state and federal laws, such as the Family Educational Rights and Privacy Act (FERPA) and the Children’s Online Privacy Protection Act (COPPA). State guidance documents consistently converged on the necessity of rigorous vetting protocols to ensure that AI tools have built-in controls to protect sensitive data. Finally, all states recognized the need to center human oversight and adjustment in AI usage. This included users (educators and students) being aware of potential algorithmic bias to make informed decisions about the use and validity of the output. States also called for evaluating AI’s impact on student learning; however, none offer any tools or guidance for how such assessments should be conducted.

These findings reveal a significant shift in how states conceive student engagement with technology with respect to AI. Participation in the AI context is largely defined in state documents through a lens of behavioral regulation and risk mitigation. By focusing primarily on acceptable use policies and academic integrity, state guidance treats participation not as an instructional milestone to be achieved, but as a set of boundaries to be policed. This orientation suggests that states are currently more concerned with governing the manner of student engagement than ensuring the quality or depth of the learning experience itself. When participation is reduced to a list of prohibited uses or vetting protocols, it fails to build the structural capacity needed to foster the foundational understanding of AI mechanics.

Experience, Equity and Ethics

Experience, equity, and ethics were prominent throughout state AI guidance documents. Nearly all state guidance recognized the potential of AI to transform education. State documents marveled at the potential of AI to transform education through improved personalized and adaptive learning. Nevada suggested leveraging AI to enhance personalized competency-based learning which “allows students to advance upon demonstrating mastery of a concept or skill regardless of time, place, or pace, creating a more flexible and student-centered learning environment” (Nevada Department of Education, 2025, p. 24). West Virginia noted AI’s potential for instantaneous feedback and personalized support, particularly for bridging accessibility gaps in underserved regions. Similarly, California highlighted AI’s utility for students with disabilities, such as voice typing, while Oklahoma emphasized its role in providing specialized learning resources to rural districts.

State policies focused on how AI can make it easier to offer multilingual support, either for language development or accessibility of translation. For example, Rhode Island suggested that AI-enabled language support tools and translation can support the design of language instruction that is differentiated for students’ English language proficiency levels, which may help educators leverage a student’s home languages as an asset. This also connected to proposed policies encouraging the use of AI for personalized learning to allow deeper learning and classroom engagement. Colorado emphasized that AI could encourage students to “explore complex concepts and ask thoughtful questions” (Colorado Department of Education, 2024, p. 7), the personalized approach aligning with research-based practices for fostering a deeper understanding and retention of material. It can also automate more basic work in preparation for deep thinking. Louisiana explicitly linked AI use to federal civil rights laws (Title VI and EEOA), stating that AI technologies have the potential to ensure English learners have increased access to meaningful educational experiences. Finally, several states discussed the potential of AI to foster creativity. Colorado continues to describe AI aiding creativity by noting applications to writing, visual arts, and music composition which can “bring students’ imaginative ideas to life” (Colorado Department of Education, 2024, p. 10).

While the promise of AI to transform teaching and learning within schools exists, there are still learning experiences that risk being compromised by AI. All state documents noted that one of the most concerning areas was the inherent bias in AI systems. Algorithmic bias can lead to unfair or discriminatory outcomes that misclassify or inaccurately evaluate students from diverse backgrounds, penalize linguistic differences when automating grading, or unfairly track students through predictive analytics. Massachusetts cautioned that AI can “reproduce the very patterns of exclusion, underrepresentation, and limited access that public education seeks to dismantle” (Massachusetts Department of Elementary and Secondary Education, 2025, p. 24). The Arizona state guidance cited the scholarship of Punya Mishra who “highlighted a concern that most contemporary datasets are WEIRD, that is, they ‘disproportionately represent Western, Educated, Industrialized, Rich, and Democratic societies’” (Arizona Institute for Education and the Economy [AIEE] & Northern Arizona University [NAU], 2025, p. 18).

AI may also create potential mental health challenges. Rhode Island, Oregon, and others cited existing anti-bullying laws and a new potential for violations of these laws through deep fake videos. California also discussed the risk of student isolation particularly if AI replaces the human interaction between educator and student.

All of the state guidance documents emphasized centering the human to mitigate these risks. Washington presented a foundational philosophy of “Human Inquiry → AI → Human Empowerment” which emphasizes that the use of AI always starts and ends with human reflection and insights (Washington Office of Superintendent of Public Instruction, 2024, n.p.). This perspective emphasized learning as an inherently social process which AI might enhance but cannot replace. Furthermore, human intervention was seen as necessary for managing AI’s inherent flaws, including awareness and a critical eye for potentially biased or erroneous output.

Measuring experience within the CS CAPE Framework has proven to be the most systematically challenging, resulting in limited empirical understanding (Dunton et al., 2022; Warner et al., 2022). Assessing an individual’s experience with CS is inherently resource-intensive and time consuming. While it remains to be seen whether measurement of AI experience will diverge significantly from that of CS, the potential of AI to transform learning introduces a greater sense of urgency. Consequently, more systematic methods for assessing student experience are required, such as evaluating progress on broad learning objectives like reading, writing, and mathematics proficiency.

The state guidance documents reveal a paradoxical conceptualization of equity, where AI is viewed simultaneously as a tool for radical inclusion and a catalyst for systemic exclusion. While personalized mastery and multilingual support offer promising avenues for accessibility, these aspirations are tempered by the persistent risk of algorithmic bias. Ultimately, the consistent call for a “human-in-the-loop” philosophy acts as the primary safeguard for policy resilience, anchoring the digital experience in human relationships to prevent student isolation. However, as the field struggles with the resource-intensive nature of qualitative assessment of students’ experiences in AI education, a measurement gap remains. Without systematic ways to evaluate student experience, the promise of human-centered AI risks remaining a rhetorical safeguard rather than a structural reality, leaving the actual quality of engagement unmonitored.

Policy Postures: Caution Vs Opportunity

The nature and purpose of states’ AI guidance, observed through the CAPE Framework, revealed a critical divergence in policy posture: caution and opportunity. As detailed in the methodology, these postures were developed as a mutually exclusive classification system based on the predominance of policy signals identified within each document. This distinction, rather than a simple presence or absence of policy, determined the underlying resilience or fragility of the resulting system. The landscape of state AI guidance was evenly split between the two primary policy postures: caution (n = 13) and opportunity (n = 13). Beyond this divide, a smaller group of states (n = 9) took a more balanced approach.

Guidance that leaned towards caution prioritized compliance, emphasizing restrictive ethical rules, stringent data privacy mandates, and warnings or prohibitions of misuse (as reflected in the reactive capacity and legalistic experience findings). For example, Arizona cautions that “GenAI relies on existing data, its output naturally raises questions about content ownership, copyright, and intellectual property (IP). Current policies and laws may be inadequate to address IP issues generated by a machine” (AIEE & NAU, 2025, p. 19).

While these safeguards provide basic stability, an overemphasis on caution induces policy fragility. By narrowly defining capacity as compliance, this reactive posture fails to cultivate the organizational depth needed for innovation, leaving the system brittle and incapable of adapting to future technological evolution. While caution is warranted, given that AI risks becoming another oversold and underused classroom technology (Cuban, 2021), its nature as an arrival technology requires balancing risk mitigation with proactive capacity-building for thoughtful integration. Crucially, even caution-dominant guidance often mandates educator PD, which serves as an inherent mechanism for policy resilience. By strengthening human capital alongside risk management, these requirements ensure the system can absorb new priorities, preventing a total fracture caused by educator unpreparedness.

Conversely, opportunity-focused guidance framed AI as a catalyst for redefining educational practices. By mandating professional development and prioritizing instructional integration and equitable access, this proactive posture fosters policy resilience. It enhances the system’s structural capacity to incorporate new priorities without compromising educational quality. The key distinction lies not in the presence of capacity-building or equitable access for all students (moving beyond rhetoric), but in the predominance of signals: whether the policy’s primary intent is to merely manage risk or to actively leverage AI for responsible, long-term transformation.

Arrival Technology: Balancing Opportunity and Caution

The integration of AI education is not a simple binary success or failure, but a complex tension between implementation urgency and system preparedness. This tension presents the crucial systemic stressor that tests the policy resilience of education systems responding to the urgency of AI in education while balancing the need to ensure systemic stability. Describing generative AI as an “arrival technology,” Smith et al. (2025) note that it entered classrooms through organic use rather than formal adoption. Consequently, the urgency for integration is driven by external innovation and workforce future-proofing (U. S. Department of Education, 2023; Pahlka, 2023) rather than pedagogical evidence. Unlike traditional initiatives supported by robust research and planning, this external pressure forces states towards rapid adoption. This haste risks compromising systemic stability and program coherence, creating an environment for policy fragility to take hold.

Policy resilience is less about slowing progress and more about aligning pace of implementation with preparation and consideration for educational value. This is particularly challenging in the case of AI where states have rapidly attempted to create meaningful and flexible guidance that supports policymaking for districts and schools (Eutsler et al., 2025). Systemic preparedness, a deliberate investment in educator capacity, technological infrastructure, and ethical frameworks, is essential to foster policy resilience. Intentional preparation ensures that the new AI priority can be absorbed without undermining core educational values or creating immediate implementation gaps. The policy landscape is thus defined by this struggle: the pressure to act swiftly versus the need to plan systematically.

A major finding is that the uneven integration of critical components, specifically teacher preparation, ethical frameworks, and explicit equity language, underscores a high risk of implementing AI education inequitably. This demonstrates a defining characteristic of policy fragility: compromising equity when adopting a new priority. Consistent with earlier CS and STEM reforms, rapid adoption may unintentionally favor well-resourced districts that have the pre-existing capacity, such as funding and PD structures, to absorb and implement policies while supporting teachers (Alzaharani, 2024; Karlin et al., 2018; Topper & Lancaster, 2013). Conversely, under-resourced districts with weaker infrastructure may experience AI as an imposed, techno-solutionist response to labor needs, rather than a supported innovation, which immediately compromises equitable access and widens existing educational gaps (Bulathwela et al., 2024). The structural unevenness confirms that the system’s current capacity is inadequate, making its commitment to equity fundamentally fragile in its commitment to equity. Without deliberate strategy focused on vulnerable districts, including mandated PD and participation metrics, AI policy risks reinforcing rather than disrupting educational inequity.

Policy Implications: Structural Integrity Over Speed

Successful AI integration requires prioritizing organizational reflection over rigid mandates. The observed fragmentation in policy signals, particularly the tension between legalistic caution posture and the innovative opportunity posture, underscores the risk of early policy fragility. To build resilience, policy should move beyond compliance checklists to help districts operationalize capacity through localized needs assessments and concrete definitions, shifting focus from vague PD recommendations (capacity) to articulating concrete strategies, timelines, and supports to teach with and about AI. Furthermore, AI guidance should shift from general “access for all” rhetoric toward mandatory mechanisms, such as resource reallocation and human interaction, that directly dismantle identified equity gaps and establish sustainable implementation. Our analysis identifies teacher capacity as a critical, structural equity issue; disparities in teachers’ preparation determine which students have sustained access to rigorous, conceptually grounded AI learning opportunities. To prevent a dual-system risk where only well-resourced districts provide critical AI education, reinforcing uneven policy adoption, school systems should make sustained investments in teacher capacity to ensure a “human in the loop” approach. Currently, AI guidance documents recommend PD aligned with EdTech application (immediate utility) over foundational computing education (conceptual coherence), leading to policy fragmentation: the misalignment of guidance, capacity-building, and implementation expectations across governance levels, resulting in inconsistent and inequitable educational experiences. If teacher preparation remains limited, the structural biases inherent in AI will remain poorly managed, reproducing the exclusion that policies are intended to dismantle. Ultimately, policymakers must recognize capacity as the primary driver of policy resilience and equitable student experience.