*By Savvas A. Hatzichristofis

We would never let a stranger walk into a classroom, sit next to an eight-year-old child, and talk to him about history, health, religion, human relationships, or their fears—without knowing who they are, what they know, or what principles they follow.

And yet, that is exactly what we risk doing with artificial intelligence.

A child today might ask a system why his grandfather died. They might ask it to explain what happened in Cyprus in 1974. A student with reading difficulties might ask it to simplify a text. Another child might write to it in the Cypriot dialect and watch as the machine “corrects” the words they heard from their grandmother as if they were mistakes.

In all these cases, the response may be excellent. However, it may be incorrect, age-inappropriate, culturally alien, stereotypical, or simply different from what the teacher is trying to teach.

The Republic of Cyprus’s National Strategy on Artificial Intelligence, currently under consultation, starts from precisely this reality. It does not first ask, “Which chatbot should we put in schools?”. It asks something more difficult: who will set the rules governing how any artificial intelligence system will be made available to teachers and students?

A national strategy, not a software purchase

This distinction isn’t limited to education. It permeates the entire National Strategy.

Cyprus aims to become a reliable European digital hub in the Eastern Mediterranean by 2032. The Strategy links public administration, health, financial services, tourism, legal services, education, shipping, and entrepreneurship with a common goal: to leverage technology without relinquishing critical levels of control.

Its guiding principle is “sovereignty through partnerships.” Cyprus cannot and does not need to develop every artificial intelligence model on its own. It can collaborate with the world’s leading companies and research organizations. However, it must control data of national value, the rules, accountability, access, and the ability to switch suppliers when necessary.

In education, this philosophy translates into a unified national architecture.

The Cyprus AI Literacy Framework (CALF) defines what a child should know about AI based on their age. For younger children, the goal is to recognize simple patterns, understand rules, and use technology safely under supervision. During adolescence, the focus expands to include understanding data, recognizing biases, verification, data rights, and responsible experimentation.

The National AI Skills Observatory (NASO) will monitor whether this effort is effective and will provide real-world data to inform educational policy.

FutureAI CY extends AI education beyond the classroom to vocational and lifelong learning.

And somewhere between what we want children to learn and how we use AI to achieve that lies perhaps the most distinctive innovation of the Cypriot proposal: the Ethical Pedagogical Validation Layer (EPVL).

What, exactly, is the EPVL?

The EPVL is not just another chatbot. It is not a government algorithm that decides what is “right.” And it is certainly not a mechanism that replaces the teacher.

It is an open, controlled layer between the artificial intelligence model and the classroom.

The teacher starts with the learning objective. For example, the teacher asks a model to adapt a difficult text into material suitable for ten-year-olds. The response does not go directly to the students. It undergoes a review for age appropriateness, fidelity to the original meaning, alignment with the curriculum, cultural respect, accessibility, and data protection.

In the end, it’s not the machine that decides. It’s the teacher.

And this small difference changes almost everything.

Imagine a teacher who wants to explain war to elementary school children. A general model might provide her with a text that is historically accurate but full of images and details that are not age-appropriate. EPVL can highlight the problem and help adapt the material without distorting the history.

Consider a student with dyslexia. TN can break a page down into shorter sentences, change the structure, create an audio version, or provide a different presentation. EPVL can verify whether the “simplification” still teaches the same learning objective. Because inclusion doesn’t mean giving a child less knowledge. It means giving them a different path to the same knowledge.

Consider a child who asks about a difficult social issue. It’s not enough for the answer to be factually correct. It must be age-appropriate, not frightening, not stigmatizing, and, when necessary, lead the child back to a person.

Or consider an even simpler case. A teacher asks the students to create an illustration of a Cypriot village in the early 20th century. The image is beautiful, but the houses, traditional costumes, and objects belong to a vague “Mediterranean.” Without verification, a beautiful image can become a false historical memory in a matter of seconds.

EPVL does not promise that artificial intelligence will never make a mistake. It promises that the mistake will not automatically carry the authority of the school.

When the Same Story Changes Languages

This is particularly significant for a country like Cyprus.

A recently published study examined how well-known language models presented the events of Asia Minor in 1922 when the same questions were posed in different languages. It was not merely the translation that changed. The terminology, the attribution of responsibility, the causes highlighted, and the historiographical perspectives that were included or omitted all changed.

This particular study does not concern the Cyprus issue and does not prove that any model today systematically adopts the Turkish position. It does, however, show how easily language and the information environment can influence the history we hear.

So when a child asks, “What happened in Cyprus in 1974?”, words carry weight. “Invasion” is one thing; “intervention” is another. “Occupation” is one thing; “administration” is another.

We cannot allow our memory to depend solely on filters designed thousands of kilometers away, which change without consulting us and operate based on data we have no control over.

This does not mean we should create a “state-sanctioned truth” and place it above the algorithm. Quite the opposite. We need documented sources, a clear distinction between fact and claim, acknowledgment of uncertainty where it truly exists, and respect for the historical and legal context.

The same applies to our language and our traditions. The Cypriot dialect cannot be treated as “bad Greek.” A Cypriot custom cannot be reduced to a generic “Mediterranean tradition”. The technology introduced into schools must be able to recognize that behind every word, song, traditional costume, or place name lies a memory.

And who oversees the overseer?

This raises a reasonable question. If we create a national level that controls what is taught in the classroom, who can assure us that tomorrow this level won’t be used for censorship?

The answer lies in the EPVL architecture’s traceability.

The Strategy provides for transparent audit trails. It records the model used, its version, the response generated, the changes made, and the teacher’s final decision.

This is very important.

If a text is rejected, the rejection must not disappear into an invisible filter. It must be possible to see who rejected it, when, for what reason, and based on what criteria.

If a historical statement is changed, the change must be visible.

If a model provided an answer and a person modified it, we must be able to see both.

Logs do not make censorship impossible. However, they make arbitrary, invisible censorship much more difficult, because they leave a trail that can be verified and challenged.

And this is very much in line with the logic of the European AI Act, which places particular emphasis on event logging, traceability, transparency, and genuine human oversight. In high-risk situations, a human must not only see what the system has done but also be able to override, reverse, or halt its decision.

In other words, the EPVL must protect the child both from the machine and from the arbitrariness of the protection mechanism itself.

Let’s not build the school inside a company

However, there is also a second risk, less striking but perhaps more long-term: vendor lock-in.

Greece chose to begin the organized introduction of generative AI into schools through the “OpenAI for Greece” partnership, initially in 20 high schools using ChatGPT Edu. This approach offers quick access to mature technology and a structured environment. However, it starts with a specific product.

The Cypriot proposal takes a different approach: first, we establish national guidelines, and then we allow the available models to operate within those guidelines.

This is important because the experience of other countries shows how easily a useful platform can turn into an infrastructure that is difficult to leave. Estonia, an international model for the digital state, now acknowledges in the 2026 Human Development Report that lock-in to technological ecosystems is already widespread in its education system.

When educators are trained on a specific product, records remain there, processes are organized around it, and years of teaching practice accumulate within that same ecosystem, making a switch to a different vendor much more expensive than the initial purchase.

EPVL aims to break this cycle.

Pedagogical rules, criteria, logs, approvals, and the history of educational practice must remain within the national infrastructure. If we use one model today and a safer, better, or more cost-effective one becomes available tomorrow, we must be able to switch to it without having to rebuild our teaching methods from scratch.

Collaboration with companies is essential. Dependence on a single company is not.

Cypriot infrastructure, international standards

That is why it is particularly important that EPVL is not being proposed as a private product that the government will simply purchase.

In the National Strategy, it is one of the three national governance tools for digital transformation in education. It is intended to be developed within a distinct technical and pedagogical framework, with national governance, funding, readiness checks, and a pedagogical and safety committee that will maintain the criteria, approve pilot implementations, and monitor accessibility issues.

So the goal is not for Cyprus to create its own ChatGPT.

The goal is much more substantial: to establish its own level of trust and control, which will be able to interoperate with different models.

And this national framework is not being designed in a vacuum. The Strategy provides for alignment with the EU AI Act and the GDPR, while explicitly drawing on the principles and frameworks of UNESCO and the OECD for human-centered AI, pedagogical and ethical validation, age appropriateness, protection of human autonomy, and skills development.

UNESCO itself warns that educational institutions cannot automatically consider general generative AI tools suitable for use by children. It calls for the protection of personal data, age-appropriate use, and human-centered processes of ethical and pedagogical validation.

This is precisely what Cyprus is striving to transform from a general principle into a national infrastructure.

The first country to do so, with all the responsibility that entails

The EPVL did not emerge as an idea to enrich a strategic document. It is grounded in scholarly literature and research on pedagogy, child development, human oversight, explainability, data protection, cultural sensitivity, and inclusion.

It has already been implemented in an initial pilot version and has undergone an initial evaluation in Cyprus. Early indications regarding its usability and pedagogical value are encouraging. This does not mean that we have already proven its effectiveness on a national scale. Larger-scale pilot implementations, independent evaluation, and refinement in collaboration with the educators themselves are needed.

Based on the international mapping that preceded this, Cyprus is the first country to explicitly incorporate into a national strategy a distinct, open, and vendor-neutral pedagogical level of validation between TN models and the classroom.

This first step is only valuable if it is put into practice.

Because in the end, the question is not whether a child will use artificial intelligence. They will use it.

The question is what kind of child we want to be facing it.

A child who believes every well-written answer, or a child who knows how to ask, “How do you know that?”

A child who will learn that their own dialect is “wrong,” or a child who will recognize the language of their home within the technology.

A child who is given a ready-made version of their history, or a child who learns to search for sources, compare, and evaluate.

Protecting future generations does not mean keeping artificial intelligence out of school. It means not leaving the child alone to face it.

Teachers must maintain their judgment. The state must set the rules. Schools must preserve their memory. And technology should be able to change without our principles changing along with it.

We are not asking the machine to love Cyprus. We are asking the Republic of Cyprus to ensure that, when the machine speaks to our children, there will always be a person who can ask it, “Why?”

*Savvas A. Hadjichristofis is a Professor of Artificial Intelligence, Vice Rector for Research and Innovation at Neapolis University Pafos, and one of the authors of the Republic of Cyprus’s National Strategy on Artificial Intelligence, which is currently under public consultationIntelligence.