AI in education: from experimentation to governed programmes

OpenAI is advancing Education for Countries: for schools, universities and public bodies, the issue is not only adopting generative tools, but building governance, training and impact metrics.

The news: educational AI is moving into a scale phase

OpenAI has announced the next phase of “Education for Countries”, an initiative that, according to the company’s announcement, aims to expand the adoption of artificial intelligence in schools through partnerships, teacher training and dedicated tools. The signal is clear: generative AI is no longer being positioned only as a local experiment, but as part of broader educational programmes, potentially at national or institutional level.

For schools, universities, public bodies and educational foundations, the operational question is not “which model should we use?”, but “which accountability system must be in place before these tools enter the classroom?”. An AI programme in education involves minors, personal data, teaching content, assessment, inclusion, procurement, security and staff skills. It is therefore an organisational project before it is a technology project.

The main risk is treating AI as a software licence to be distributed. In practice, a serious rollout requires usage rules, integration with existing platforms, role-specific training and measurement criteria. Without these elements, even promising tools can create confusion, inconsistent use, vendor dependency, privacy issues or unrealistic expectations about educational outcomes.

What changes for schools, universities and public bodies

The first implication concerns governance. An institution introducing generative AI should define who approves the tools, which data can be entered, which activities are allowed, how incorrect or inappropriate outputs are handled, and which controls are required for underage students. These decisions cannot be left to individual teachers or isolated departmental initiatives.

The second implication is training. Explaining how to write prompts is not enough. Separate learning paths are needed: executives and IT managers must understand risks, costs, contracts and compliance; teachers and tutors must learn how to design educational activities with human supervision; students and families need clear guidance on correct use, limitations, source citation and responsibility.

The third implication is technical integration. AI tools should work with digital identity systems, LMS platforms, document repositories, school management systems and cloud environments already in use. Without integration, separate accounts, manual exports, shadow IT and audit difficulties increase. This is why modular architectures are worth evaluating: controlled environments, adequate logging, permission management, data segregation and the ability to change provider without blocking the programme.

Finally, there is the measurement issue. Claims about improved educational outcomes, when made, should be validated in the specific context. Initial indicators should be concrete: reduction of administrative time for teachers, perceived quality of materials, content accessibility, student participation, number of incidents or reports, and real adoption rates. Only later does it become reasonable to assess more complex effects on learning.

Mistakes to avoid in an educational AI rollout

The first mistake is starting from the platform rather than from use cases. A programme should select a few high-value, low-risk scenarios: support for preparing materials, document summarisation, controlled tutoring, exercise generation, language accessibility and administrative assistance. More sensitive scenarios, such as automated assessment or student profiling, require much stronger safeguards.

The second mistake is overlooking data management. Before rollout, institutions need information classification, rules for personal data and children’s data, privacy assessments where required, defined retention policies and clear contractual clauses on data use, model training, localisation and subcontractors.

The third mistake is confusing automation with delegation. In education, AI should remain a support tool: teachers retain pedagogical responsibility, context and judgement. This principle should be written into policies and translated into interfaces, workflows and training materials.

The fourth mistake is failing to plan a review cycle. Policies and tools must be updated based on feedback, incidents, new regulations, model changes and measured results. Static governance quickly becomes obsolete.

A practical roadmap to get started

A prudent approach can follow five steps.

  1. Map needs and constraints: teaching and administrative processes, existing systems, user categories, privacy and security requirements.
  1. Define a minimum policy: permitted uses, prohibited data, responsibilities, human supervision, error management and transparency towards students and families.
  1. Launch a controlled pilot: a limited number of use cases, a restricted group, a secure technical environment, simple metrics and operational support.
  1. Train by role: executives, IT teams, teachers, administrative staff and students have different needs; a single generic course is not enough.
  1. Decide whether to scale: compare benefits, risks, costs, incidents, real adoption and feedback. Only then should integrations, contracts and the support model be consolidated.

From Dedaloop’s perspective, the key point is to treat educational AI as a socio-technical infrastructure: software, data, security, training and governance must move together. Initiatives such as the one announced by OpenAI may accelerate sector interest, but adoption quality will depend on local choices: architecture, rules, skills and the ability to measure what actually happens.

Approfondimenti Dedaloop

Fonti