Building AI for Education Starts With Public Data Infrastructure

AI’s pace and promise are undeniable. But in public education, the limiting factor is no longer the model; it’s the data. 

If we want AI systems that are effective, auditable, and equitable for students, we must start with something more basic: open, interoperable, standardized data infrastructure that public agencies can govern and sustain. 

Interoperability before intelligence. Governance before generation. Context before computation. 

That’s not a slogan; it’s an ordering of work. 

Why interoperability comes before AI 

K-12 has lived through many “next big things” in edtech. Some have delivered real value, especially in large-scale data collection and adaptive learning. But across states and districts, student outcome progress and gaps remain stubborn, and leaders are rightly skeptical that a new wave of AI tools alone will change that. 

This is not simply a failure of technology. It reflects a basic fact: no algorithm can solve challenges rooted in resource inequities or social conditions. What AI can do is help educators, policymakers, and researchers reason more clearly with the information they already have. That only works if the data are accurate, standardized, and connected, and if the people closest to students remain in charge of how those data are used.

The tension: local control versus comparability 

Public education data live in a real tension: 

• We want local control and use: data that reflect local policies, programs, and community priorities.
• We also want comparability and learning across sites: the ability to evaluate interventions, understand equity patterns, and report on progress across districts or states. 

Today, we often fail at both sides: 

1. Vendor-specific structures mean that each system defines its own schema (students, courses, interventions, assessments), making technical alignment expensive and requiring ongoing rework whenever a vendor changes its design.
2. Location-specific meanings occur because each jurisdiction defines key concepts differently, so even when data columns line up, the underlying logic does not. 

What goes wrong when AI guesses 

When AI systems are trained or prompted on loosely assembled, de-contextualized data, three kinds of problems show up quickly: 

1. Misclassification and unfair labeling 

In one district, “chronically absent” may mean missing 10% of instructional days; in a neighboring county, 15%. An AI-driven early warning system that ignores those differences will flag students inconsistently: over-identifying in one district and under-identifying in another. 

Similarly, a new teacher’s evaluation scores might look “below average” simply because they work in a school serving higher-need students. Without contextual features such as student backgrounds, course load, and support programs, an AI model can misclassify the teacher as underperforming, entrenching bias instead of surfacing need. 

2. Broken chains of meaning 

Education data are not simply a pile of clicks or test scores. They are a living system of rosters, programs, accommodations, calendars, policies, and permissions. 

If an AI tool pulls assessment results without the associated accommodations or course enrollments, its inferences about “growth” or “mastery” may be wrong. If it sees discipline events without the local codebook, it may treat a dress-code violation and a safety incident as equivalent. 

In both cases, the issue is not model capacity; it is missing context. 

3. Loss of trust and democratic oversight 

Public education is accountable to communities. When AI systems produce recommendations that families or educators cannot trace back to understandable data and rules, trust erodes. 

If we cannot explain which data were used, how they were standardized, and what policies shaped them, then we cannot credibly claim that AI-assisted decisions are fair. That is a problem not just for technical adoption, but for democratic governance: it becomes harder for boards, legislators, and communities to oversee how public data are being used. 

Standards and governance: how AI gets grounded 

The good news is that the field has already built critical pieces of the foundation AI needs to have a positive impact in education.

Frameworks like Ed-Fi and the Common Education Data Standards (CEDS) provide shared semantics for core concepts (students, courses, assessments, programs) across systems and jurisdictions.  

In EA’s work, we treat these standards as a grammar for education data: 

• Ed-Fi structures operational data: who is enrolled where, which courses they take, which assessments they sit for, and how schools and districts are organized.
• CEDS provides a broader, cross-sector vocabulary that supports connections and comparability from early childhood to workforce.
• On top of that, open frameworks like Enable Data Union (EDU) organize data into an analytics-ready warehouse where definitions, timelines, and business rules are explicit and versioned. 

Standards, however, are only half of the story. Governance is what keeps meaning intact over time: 

• Capturing how a state defines “chronic absence” or “career-ready” this year, and how that differs from last year.
• Documenting which interventions count as “tutoring” or “mentoring,” and under what conditions.
• Recording who is allowed to see which data, for which purposes, under which policies. 

AI systems that operate inside that governed, standards-aligned environment can be rigorously evaluated and improved. Systems built outside it cannot. 

Bring AI to the data, not the data to AI 

Most AI products today assume that data will be copied into a vendor environment, or collected directly by the vendor, such that models are trained and tuned with limited visibility for the public agencies that supplied the data. 

For public education, that model is often misaligned with both privacy law and governance expectations. It also makes it harder to compare tools: each vendor defines its own pipelines, features, and evaluation methods. 

EA’s stance, consistent with our interoperability and governance posture, is different:  

• Keep student data local and governed by districts, states, or their collaboratives.
• Standardize the operational layer using Ed-Fi and related standards, so vendors and research partners work from the same structures.
• Use an open, documented analytics model (such as EDU) so transformations and metrics are inspectable and reusable.
• Define a controlled “data surface” for AI: the specific tables, features, and context that models are allowed to use for a specific use case. 

This “bring AI to the data” approach makes evaluation and comparison tractable: 

• Agencies can run multiple AI tools against the same governed data surface.
• They can create documentation, testing processes, and equity reviews for their models that are tailored to their own goals and their own student populations.
• They can log access and decisions, meeting FERPA-aligned privacy expectations and emerging AI governance guidance. 

When interoperable infrastructure is in place, AI can do useful work 

Once systems share a consistent, governed data backbone, AI can support three areas that matter for public education and democracy: 

1. Governance and oversight 

With standardized, well-documented data: 

• Boards and state agencies can see where AI tools are drawing information from and how metrics are constructed.
• Privacy and security teams can audit access paths and enforce policy consistently across tools.
• Communities can ask, and answer, basic questions: Who is being impacted by this system? On what basis? With what safeguards? 

2. Measurement and continuous evidence 

Infrastructure built on Ed-Fi, CEDS, and EDU can act as evidence engines: shared pipelines that support repeated cycles of measurement, evaluation, and improvement, not one-off studies. 

In that environment, AI can help with: 

• Automating routine data quality checks and documentation.
• Generating first-draft analyses or visualizations that researchers and analysts refine.
• Supporting faster feedback loops on interventions by standardizing how “intent” (the theory of action) and “outcome” are recorded together.  

3. Transparency and public trust 

When the underlying infrastructure is open and standards-based, it becomes possible to: 

• Publish metrics, business rules, and even transformation code as public artifacts.
• Invite external researchers, watchdogs, and communities to inspect and challenge methods.
• Demonstrate that AI-assisted decisions are anchored in transparent systems, not opaque, proprietary black boxes. 

These are not fringe concerns; they are hallmarks of a healthy public education system and, more broadly, a healthy democracy. 

What education agencies and vendors can do now 

Different constituencies have different levers, and they can all move in the same direction. 

As a nonprofit, EA’s role is to help agencies and collaboratives implement this public-good infrastructure and to act with our partners as long-term stewards of open standards and shared transformation code, so that AI infrastructure doesn’t vanish when a contract ends or a grant expires. Ed-Fi-based operational layers, EDU-style analytics warehouses, and evaluation frameworks are open public goods that make AI tools testable and trustworthy without creating new data tollbooths. 

The bottom line 

AI will influence education. Whether that influence is helpful or harmful depends in no small part on the data infrastructure underneath it. 

The breakthrough we need will not come from the next model release. It will come from the systems that make those models trustworthy, contextual, and accountable: shared data surfaces, public AI evaluation methods, and governance structures that put agencies in control. 

• Educators can rely on them because the data are standardized and governed. 
• Researchers can evaluate them because the pipelines and metrics are transparent.
• And communities can trust them because oversight is built into the infrastructure, not bolted on afterward.

Interoperability before intelligence. Governance before generation. Context before computation. 

That is the path to responsible AI in K-12.