AI in Education: What Changes in 2026

Artificial intelligence in education has moved from conference topic to public policy in the United States. In April 2026, the U.S. Department of Education released a guidance document for K-12 education; in May, the National Education Policy Center approved a framework that classifies AI use by risk level. Anyone operating Moodle, EAD, or any learning platform needs to respond now—with engineering, not just talk.

TL;DR

• On April 8, 2026, the U.S. Department of Education released the guidance document "Artificial Intelligence in K-12 Education" and the self-paced course "AI in Teaching Practice," hosted on the department's learning platform, in partnership with UNESCO.
• On May 11, 2026, the National Education Policy Center approved a framework that divides AI applications into four risk levels—from low to excessive, the latter being prohibited.
• Research by the Bill & Melinda Gates Foundation with Equidade.Info: 84% of students and 79% of teachers have used AI, but only 32% of students received guidance on responsible use.
• Moodle has had a native AI subsystem since version 4.5; Moodle 5.2, released in April 2026, brought Gemini and Amazon Bedrock into the core.
• Without institutional policy, audit trail, and human oversight, adopting AI becomes a liability under CCPA and other state privacy laws.

From Experiment to Public Policy: What Changed in 2026

Until 2025, the conversation about AI in education revolved almost entirely around one phrase: "students are cheating with ChatGPT." In 2026, the focus shifted to governance.

Three milestones explain the shift. The first is the "Framework for Responsible Use and Development of Artificial Intelligence in Education," published in March 2026 by the Office of Educational Technology, part of the U.S. Department of Education. The document emphasizes five points: effective human oversight, transparency and explainability of systems, strict compliance with data protection laws, valuing the teaching profession, and open innovation ecosystems.

The second milestone was the webinar on April 8, 2026, when the Department of Education presented the guidance document "Artificial Intelligence in K-12 Education," organized around four pillars—critical understanding of AI, intentional pedagogical use, protection of rights and well-being, and professional development. It included ten guidelines for administrators, aligned with state standards, ongoing training, and compliance with CCPA and other state privacy laws, plus the course "AI in Teaching Practice: Ethical, Creative, and Pedagogical Use" on the department's learning platform, produced with UNESCO, with a minimum duration of 20 days and certification that counts toward career advancement in public schools.

The third milestone came from the National Education Policy Center. On May 11, 2026, the center approved a framework with guidelines for K-12 and higher education, which went to public comment—contributions were accepted until June 14—before a plenary vote and eventual adoption by the Department of Education.

None of these documents are self-executing regulations. But they all point in the same direction: those offering educational technology will need to prove how the system makes decisions, with what data, and under whose supervision.

The Numbers That Explain the Urgency

The regulation was not born from an abstract debate. It chased after usage that was already widespread: AI in American education arrived through the student, not the administrator.

The survey "Perceptions on Artificial Intelligence in Education," by the Bill & Melinda Gates Foundation in partnership with Equidade.Info, heard from 1,947 students, 240 teachers, and 156 administrators in 142 schools. The result is a snapshot of adoption without a manual.

Indicator	Number	Source
Students who have used an AI tool	84%	Bill & Melinda Gates Foundation / Equidade.Info
Teachers who have used AI	79%	Bill & Melinda Gates Foundation / Equidade.Info
Students who received guidance on responsible use	32%	Bill & Melinda Gates Foundation / Equidade.Info
Students whose school never discussed AI in class	73%	Bill & Melinda Gates Foundation / Equidade.Info
Students who talked with teachers about AI in activities	19%	National Center for Education Statistics (NCES)

A compilation published on August 11, 2026, on Student Day, with data from the American Council on Education, the Bill & Melinda Gates Foundation, Equidade.Info, and NCES, completes the picture: only 37% of schools allow AI use in educational activities, 21% of teachers have never used the technology, and yet 84% of teachers cite AI as support in planning and 76% use it to create teaching materials. Among students, 54% recognize the danger of unregulated use and 75% are aware of the potential for creating fake news.

Read these numbers as a product diagnosis, not as newspaper statistics. Demand exists, the technical competence of the end user is uneven, and the governance layer is missing. That is exactly where a well-built platform solves the problem—and a poorly built one amplifies it.

The Risk Classification Framework and What It Demands from Your Platform

The framework approved by the National Education Policy Center establishes four risk categories for AI in formal education. This is the part of the text with direct technical consequences, because each tier implies different controls in the software.

Risk Level	Examples Cited	What the Platform Needs
Low	organizing materials, accessibility features	usage logs and labeling of AI-generated content
Moderate	virtual tutors, automated feedback	explicit consent, interaction history, teacher review
High	automated grading, biometric proctoring	mandatory human oversight, audit trail, impact assessment (CCPA/state laws)
Excessive (prohibited)	emotional surveillance, social scoring, fully automated approval	do not implement

The framework also conditions implementation on teacher training and the progressive inclusion of AI content in the curriculum. Translating that into requirements: your platform needs a dashboard that shows who used which model, for what purpose, with what output, and who reviewed it. That is not a compliance ornament. It is the difference between defending a pedagogical decision in an audit and not being able to defend it.

What Changes for Private Institutions

Private EAD operators often think that Department of Education guidelines do not apply to them. They apply in two ways: through accreditation, once the framework is adopted, and through contracts—public RFPs and large corporate clients are already asking for an AI in education policy as a mandatory attachment. Having yours ready is a commercial advantage before it becomes an obligation.

The Case of California: When AI Becomes a Required Subject

California is the most advanced example of AI in public education in the United States. Through the "California AI Initiative," implemented starting in 2024, the state included AI as a required subject in 9th grade and throughout high school, now serving more than 130,000 students—the first state in the country to do so, an initiative that earned recognition from UNESCO.

The curriculum covers machine learning, algorithms, and AI ethics. Note the detail: it is not "a class on using ChatGPT." It is conceptual training. That difference matters for platform developers, because it changes the type of resources the district will request—labs, project tracks, rubric-based assessment—instead of a generic chatbot plugged into the homepage.

How AI in Education Comes to Moodle in Practice

If your learning environment is Moodle, the good news is that you do not need a workaround. Since Moodle 4.5, there is a native AI subsystem (core_ai), designed to standardize integration with external providers.

The architecture has two pieces: providers, which are the bridge to external AI, and actions, which are the instructions. The basic actions are three—generate_text, generate_image, and summarise_text. Moodle 5.0 gave administrators finer control over where the subsystem operates, and Moodle 5.2, released on April 20, 2026, brought Gemini and Amazon Bedrock into the core, along with improvements to the question bank and Report Builder.

Where AI Really Pays Off in the LMS

Not every feature is worth the token cost and review overhead. In our experience, AI in distance education pays off in four areas:

• Drafting questions from material already approved by the teacher, with mandatory review before publishing.
• Summarizing long content for student review, with a permanent link to the original source within the course.
• Detecting at-risk students by engagement patterns—historically the highest ROI feature in EAD, because it attacks dropout.
• Writing course descriptions and rubrics, an administrative task that consumes hours of coordination.

Where It Usually Goes Wrong

Automated grading without a teacher in the loop and proctoring based on behavior analysis fall, respectively, into high risk and into territory that the framework classifies as excessive when it involves emotional inference. If your roadmap includes these items, treat them as regulated projects, not as features.

CCPA, Human Oversight, and Common Pitfalls

Three errors appear frequently in AI in education projects, and all are avoidable at the architecture stage.

Before detailing them, a framing: applying AI in education is not the same as applying it in e-commerce. The data subject is often a minor, the system's decision affects the student's academic trajectory, and errors only appear months later, in assessment.

The first is sending minor data to a third-party API without a defined legal basis and without a data processing agreement. Student data is sensitive in practice, even if CCPA classifies it differently; treat it with the same rigor.

The second is not keeping logs. If the system generated a question, feedback, or an alert about dropout, you need to know which model responded, with what prompt, when, and who validated it. Without that, you cannot even correct a bias detected later.

The third is cognitive offloading. Researchers point out that students who use AI do better on immediate tests and worse on later retention. The product response is not to block: it is to design the flow so that AI asks for the student's attempt before delivering the answer—which changes the system prompt, not the usage policy.

When NOT to Use AI

In high-stakes summative assessment without human double-checking. In identifying students through behavioral biometrics. In any decision about approval, failure, or dismissal. And in classes whose teacher has not yet undergone training—the framework itself places teacher training as a condition for implementation, not as a later step.

A 90-Day Plan to Adapt Your Platform

This is the roadmap we use to put AI in education at an institution on an auditable track. It was designed for Moodle, but the logic applies to any LMS.

1. Weeks 1–2 — Inventory. List every point where AI already touches the environment, including third-party plugins and support integrations. There is almost always more than the administrator imagines.
2. Weeks 3–4 — Risk classification. Fit each use into one of the four tiers of the framework. Anything that falls into excessive risk is taken offline immediately.
3. Weeks 5–6 — Institutional policy. A short document, written with teachers and coordination, published in the environment itself. It should state what is allowed, what requires citation, and what is prohibited.
4. Weeks 7–9 — Technical layer. Configure the AI subsystem with a single provider and your own key, enable logs, add a visible label for AI-generated content, and create the audit report.
5. Weeks 10–12 — Training and pilot. Enroll the team in the department's course, run a pilot in two subjects, and measure: teacher time saved, student retention, and number of human reviews needed.

If the pilot does not produce numbers, it is not finished. Metrics without a baseline are opinion.

Mobile: Where Adoption Really Happens

An AI feature that only exists in the desktop browser reaches a minority of study time. Real adoption of AI in education happens on the phone, between commutes.

That is why we treat study assistant, summary, and deadline alerts as app features, not portal features. We have already detailed the comparison between the official app and a custom solution in Moodle Mobile App vs Custom Moodle App and the practical advantages in Custom Moodle App: 7 Advantages Over the Official App. For the retention effect that matters most in EAD, the path goes through well-calibrated push notifications—and, when the student studies without a stable connection, through offline features.

On the market tools side, it is worth following what the platform your network already uses is delivering: we have gathered recent changes in Google Classroom in 2026 and the landscape of autonomous agents in AI Agents for Businesses.

How We Handle This at Agathas Web

I have worked with Moodle for over a decade, am certified on the platform, and serve as CTO of IEJUR, where the EAD operation does not tolerate improvisation. In the projects I lead, the adoption of AI in education follows a simple rule: no model output reaches the student without passing through a teacher or an explicit rule defined by the coordination.

In practice, this means four deliverables: an AI subsystem configured with the institution's own provider and key—never shared; an audit report accessible to the coordination; a label for AI-generated content on every automatic output; and a custom app so the feature reaches where the student is. It is less glamorous than announcing "AI integrated" and much more defensible in an audit.

Conclusion: The Next Step Is the Inventory

2026 closed the phase where it was possible to treat AI in education as a tolerated experiment. There is a framework from the Department of Education, a guidance document, an official training course, and a framework from the National Education Policy Center with risk classification on its way to adoption. On the other side, 84% of students already use the technology and the majority have never received guidance.

The cheapest move you can make this week is not to hire anything: it is to open a spreadsheet and list where AI already touches your learning environment. That inventory usually reveals two or three uses that today no one can explain—and that is where you start.

If you want to discuss adapting your Moodle or your institution's app, the conversation is open.