Edge AI in Education. Who’s in Control?

Most schools already have everything they need.

Decades of carefully chosen textbooks. Lesson plans refined through years of teaching. Library collections built with real intention. Archives that hold institutional memory that no one has yet found a good way to surface. The knowledge is there, in filing cabinets, shared drives, and the accumulated expertise of educators who’ve spent careers figuring out what works.

And yet, AI is already in the classroom.

Not as infrastructure schools have chosen, but as tools students and teachers reach for anyway to explain concepts, summarise texts, generate ideas, or check understanding. The question is no longer whether AI belongs in education. It’s whether the AI already being used reflects the institution behind it, or something else entirely.

What’s harder to find is an AI that knows any of that. One that represents the values of the people who built it, protects the students it serves, and remains transparent enough for the community using it to genuinely trust it.

The gap is worth sitting with.

The Mismatch: General AI in Specific Institutions

The AI most students find today wasn’t designed for them.

It was built at scale, for general use, trained on data no school selected or reviewed, running on infrastructure no member of staff can inspect. That’s not an assumption; it’s the reality of how these systems are built.

General-purpose AI is, by design, general. It doesn’t know your curriculum. It doesn’t know your pedagogical commitments. It doesn’t know that your Year 9 class is still developing source evaluation skills, or that your literature department has spent years building a programme around specific authors and traditions.

More importantly, its limitations aren’t always visible. A model that confidently fills gaps without acknowledging uncertainty is pedagogically risky in ways that are easy to overlook until the habit is already formed. So the alternative isn’t simply use a different product. It has to do with the infrastructure those systems are built on.

And this is where the focus shifts: Who is responsible for the AI your students use? Who can see how it works? Who can feed the models, and with what data? Who can change it when it fails?
And just as importantly: Who is responsible for teaching students how to question it, use it, and understand its limits?

Data, Privacy, and the Students Who Can’t Consent

Any serious conversation about AI in education has to start here.

Students are not the general population. Many of them are minors. They interact with AI in contexts shaped by institutional authority, in classrooms, on school platforms, as part of assignments they’re required to complete. They rarely have a meaningful choice about whether to participate, and they almost never have full visibility into what happens to the data their interactions generate.

Frameworks like GDPR and FERPA exist precisely because this asymmetry matters. They establish obligations around data minimisation, purpose limitation, consent, and the right to access or delete information. These are not optional compliance exercises. They reflect something more fundamental: the right of the people most affected by a system to understand how it uses their data.

Most commercial AI platforms weren’t designed with these constraints at the core. They were designed for scale, and compliance was layered on afterwards. The result is a category of risks institutions don’t fully see: a vendor changes its data practices, a breach exposes student interactions, a terms-of-service update quietly expands how data can be used…

This is where infrastructure becomes relevant. Edge AI — or more precisely, systems that run on infrastructure the institution controls and where data doesn’t leave its environment — doesn’t remove every challenge immediately, but it changes the terms of the problem. When students' queries never reach external servers, a whole category of risk simply doesn’t exist. When institutions can audit what’s stored, where, and for how long, compliance becomes something they can verify rather than something they have to trust.

That distinction matters in terms of what schools can honestly say to students and families about how their data is handled.

From Tools to Governed AI Infrastructure

This is the shift that often gets overlooked.

Most conversations about AI in education focus on tools: which application to use, which assistant to adopt, which platform integrates best. But tools sit on top of infrastructure, and infrastructure determines what’s actually possible.

A self-hosted AI system, built on open-source models and deployed within infrastructure that the institution controls, changes the role a school can play. It moves the institution from user to operator.

And that shift is not just about control. It’s about visibility, too. It means building AI systems within your own domain. Training them on materials you trust. Defining how they respond and ensuring no data is sent to third-party services.

That’s what makes infrastructure meaningful. When an AI’s knowledge can be traced to identifiable sources, when its limitations can be observed, and when its behaviour can be inspected and adjusted over time, that’s when it becomes subject to the same kind of scrutiny that good education applies to everything else. And that includes teaching the educational community, students included, how it works. They’re no longer interacting with a black box, but engaging with a system whose boundaries are visible. Which means AI stops being something to simply use to become something that can be properly understood and governed.

Done well, this kind of infrastructure doesn’t limit what AI can do in education. It makes new kinds of learning possible, precisely because the system is constrained, inspectable, and open to challenge.

Governing AI in Education

When that shift happens — from using AI to operating it — another question emerges: What can we build with it?

And this is where AI infrastructure becomes practical. Empathy AI approaches this as a full stack, from privacy-first, self-hosted infrastructure to applications designed for new forms of knowledge discovery and learning.

Empathy AI’s Knowledge Base is one example of that. It changes how knowledge is accessed in educational institutions. Instead of navigating multiple systems, the community can interact with a conversational layer built on top of curated materials. Students and teachers ask questions in natural language, and the system responds using sources the institution has selected and can inspect.

The experience shifts in subtle but important ways: discovery becomes more fluid, but not less rigorous. Information can be surfaced quickly, while still pointing back to where it comes from. And because the underlying materials are shared, it creates a common ground between students and teachers.

But the implications can go further when this model is applied more creatively:

Take Hamlet in a literature class.

Traditionally, students approach the play at a distance, reading and analysing it, but not always fully engaging with it. The interaction is structured, but often indirect.

Now, imagine a different entry point where students directly interact with characters. They can ask: Why do you hesitate? Do you actually believe the ghost? Were you ever really honest with Ophelia? And instead of receiving a generic response, the system answers using Shakespeare’s text, drawing from the play, surfacing contradictions, and grounding interpretation in the material itself.

This is the approach behind Empathy AI’s Bring the Book to Life project, built on public-domain collections like Project Gutenberg.

The effect is immediate. Students engage more directly with the text, formulating questions even before the formal analysis begins. The play becomes something they interact with, not just something they’re asked to interpret. And the discussion that follows in class between students and teachers tends to deepen, because both are working through the material together, exploring ambiguities and testing interpretations in real time.

What makes this possible isn’t just the AI’s interface, but the system behind it, grounded in known, shared sources.

The Question Underneath All of It

At this point, the question becomes difficult to avoid. AI is already part of how students learn — and how teachers find new ways of teaching. So, the real question is: What kind of system are they learning from? One that’s external, opaque, and fixed? Or one that’s visible, adaptable, and shaped by the institution responsible for their education?

That choice doesn’t happen at the level of third-party tools; it happens at the level of edge AI infrastructure.

—--

Where to Begin

For most institutions, the challenge isn’t understanding the value of this approach. It’s making it tangible.

Infrastructure and AI development can feel abstract until you see it, interact with it, and understand what’s actually happening behind the scenes. That’s why hands-on exposure matters. For institutions looking to move from theory to practice, Empathy.AI runs a one-day workshop to build and explore self-hosted AI systems using their own curriculum materials. This workshop is designed for educators, administrators, and EdTech teams, not as end users, but as participants in how those systems are shaped for their own domains and needs.