ChatGPT is a useful starting point for a conversation about AI in education. It gives people something concrete to try: ask a question, draft an explanation, adapt a resource. The difficulty comes when that experience becomes the basis for an institutional strategy.
For a ministry, education publisher or NGO, the question is how those capabilities fit into the work of delivering education. Who approves a generated resource? What information can staff share with a provider? How will a translated explanation be checked? What happens when a funded pilot ends?
My view is that ChatGPT-class tools can make useful contributions across that work. But the programme around them needs as much attention as the technology. Without it, institutions risk expanding production faster than their capacity to check what they produce.
ChatGPT is an entry point, not the whole category
ChatGPT is one product within a much wider field, and generative AI is itself only part of AI. Keeping that distinction clear helps a leadership team avoid treating one familiar product as the default answer to every problem.
Even among conversational tools, the market is diversifying. Similarweb’s 2026 analysis estimates that ChatGPT remained the largest destination for generative-AI web traffic in May 2026, while Gemini and Claude had gained share. Web visits do not measure all AI use, particularly tools embedded in other services.
For education leaders, this matters because a programme should survive a change of provider. Staff need transferable judgement about sources, privacy and verification. Approved content and evaluation criteria should remain usable when the organisation changes its tools.
“ChatGPT training” can be a sensible introduction. It needs to lead somewhere beyond familiarity with one interface.
Adoption is already creating leadership decisions
RAND’s 2025 US research found that 53% of English language arts, mathematics and science teachers reported using AI for school. Yet only 45% of principals reported school or district AI policies or guidance. More than 80% of students said their teachers had not explicitly taught them how to use AI for schoolwork.
These figures describe US schools, not ministries or education systems worldwide. They nevertheless illustrate a problem leaders elsewhere should investigate: what are people already doing, and how much of that practice has institutional support?
An acceptable-use policy helps, but it cannot answer everything. Staff also need approved routes for using tools, time to learn, and someone to approach when an output or request is questionable.
That is what I mean by an operating model: the everyday arrangements that connect permission, responsibility, budgets and quality checks. A policy document is one part of it.
Where ChatGPT-class tools can help
I would begin with bounded tasks where an informed professional can check the result against an agreed source. The useful unit of planning is a complete piece of work, including review and delivery.
| Education task | Potential contribution | What the organisation must provide |
|---|---|---|
| Content production | Draft explanations, examples or resource variants | Approved source material, rights clearance and subject editing |
| Teacher support | Prepare lesson outlines and supporting materials | Curriculum context, classroom constraints and teacher review |
| Localisation | Produce translation or reading-level drafts | Language specialists and checks on meaning, terminology and curriculum fit |
| Assessment development | Suggest items, distractors and feedback | Subject expertise, assessment review and validation appropriate to the stakes |
| Reporting and analysis | Draft commentary on verified tables | Controlled data access, checked calculations and review of interpretations |
These are proposed applications, not promises that every tool will perform them adequately.
Consider a publisher preparing a teacher guide. A model might draft alternative explanations from approved material. An editor still needs to check whether the explanation is correct, appropriate for the grade, consistent with the curriculum and practical in the intended classroom. If the draft takes longer to repair than writing it directly, the workflow has not delivered a saving.
The same principle applies to analytics narratives. A model can help explain an approved table, but it should not invent missing values or turn a correlation into a causal explanation. The numbers need a reliable calculation process; the interpretation needs an accountable reviewer.
The OECD Digital Education Outlook 2026 identifies opportunities beyond classroom chat, including curriculum alignment, assessment-item development and classification of educational resources. Those uses deserve attention from leaders responsible for the wider education value chain.
Where a useful draft becomes a trust problem
Risk increases when provisional material begins to carry institutional authority. A generated explanation becomes an official resource. A translation goes to schools under a publisher’s name. A suggested assessment item enters a live examination.
At each handover, someone needs authority to reject the output. A vague instruction to keep a human involved is insufficient if that person lacks the expertise, time or access to sources needed to review it.
Learning brings a further complication. Completing a task successfully with AI does not necessarily mean acquiring the ability to do it independently.
In a high-school mathematics field experiment involving nearly a thousand students, researchers compared a GPT-4 interface resembling standard ChatGPT with a version designed to safeguard learning. Both improved performance during supported practice. When access was removed, the standard-interface group performed worse than students who had not used it. The safeguards in the tutor version largely mitigated that harm.
This is evidence from a particular setting and design, not a verdict on every current chatbot. Its significance is the distinction between assisted performance and learning. The OECD’s 2026 synthesis makes that distinction central to its assessment of educational generative AI.
For leaders, the implication is practical: evaluate what learners can do afterwards, not only the quality of the work they submit while using a tool.
Localisation includes how the resource reaches people
A programme intended for several countries cannot assume that translating its English materials completes the adaptation.
Curriculum terminology may differ. Examples may depend on unfamiliar institutions or resources. A lesson can be linguistically accurate and still make unrealistic assumptions about classroom equipment, lesson time or teacher support.
Access also needs more careful treatment than asking whether a mobile network exists. GSMA’s Mobile Economy Africa 2025 report identifies device affordability and digital skills as barriers to mobile-internet use even where coverage is available. Those continental findings do not describe every school, but they are a reason to test access assumptions locally.
In some settings, a connected content team producing reviewed resources for print or low-bandwidth distribution may be more useful than requiring every learner to use a chatbot. Elsewhere, direct interactive access may be appropriate. These are delivery choices to investigate with the people who will use the programme.
The decisions to make before expanding
Rather than commissioning a broad rollout, I would ask a leadership team to settle these decisions for one proposed use case.
Procurement and continuity. Specify the task and evaluate providers against it. Check contractual terms, access controls, support, portability and the full cost of delivery. Budget for expert review and maintenance, including what happens when grant funding or a pilot licence ends.
Data and content. Decide what may enter the tool, under which agreement, and who can access it. Learner information, unpublished assessment materials and licensed content need explicit consideration. A paid account alone is not a data-governance decision.
Assessment integrity. State what assistance is allowed and what evidence of learning is required. Give teachers and learners examples they can apply. Distinguish developing assessment materials with AI from allowing AI assistance in a learner’s response.
Professional development. Give staff repeated opportunities to inspect, adapt and reject outputs using realistic tasks. A demonstration establishes possibilities; sustained support builds the judgement to use them responsibly.
Localisation and evaluation. Name who checks language, curriculum fit and delivery conditions. Establish a baseline and assess the whole workflow: review time, quality, cost, reach and, where relevant, learning. Agree what would justify stopping as well as expanding.
This keeps the initial programme small enough to examine properly without treating governance as a reason to postpone all experimentation.
What education leaders should build around ChatGPT
My interest is in programmes and products that make useful capabilities dependable enough for their intended purpose. That requires decisions about content, professional judgement and delivery that a chatbot cannot make on an institution’s behalf.
The human-centred approach discussed in this Cambridge piece is a useful starting principle. It becomes meaningful when the programme specifies whose judgement matters, where they exercise it and what support they receive.
ChatGPT can help a team explore possibilities quickly. The next step is to choose a worthwhile education problem, design the work around it and test whether the result is better for the people it is intended to serve.
If you are planning an AI programme or education product, get in touch to discuss the use case, the delivery context and the decisions that need to be made before it grows.
AI assisted in the research synthesis of this article.