A decision is a better unit of practice than a page
Most digital learning still asks people to read, then answer questions about what they read. That works for recall. It works less well for the things professionals actually struggle with: choosing under uncertainty, balancing goals that pull in different directions, and noticing consequences that arrive late. A well-built simulation turns those moments into something a learner can rehearse. You set a policy, adjust a staffing plan or pick a laboratory step, and the system responds.
The value is not the graphics. It is the loop: decide, observe, explain, decide again. Each pass gives the learner a reason to ask why, and gives the teacher a concrete moment to discuss.
Two layers: the engine and the explainer
When people talk about “AI-powered simulations” they often blur two very different components. We find it useful to name them separately.
- The simulation engine holds the rules: how a carbon tax affects emissions in the model, how long a queue grows when a shift is short. It should be documented, inspectable and the same for every learner in a given scenario.
- The explanation layer is optional. A language model could describe what changed, suggest a question to reflect on, or rephrase feedback at a different reading level. It must never silently change the rules or the score.
Keeping the layers apart matters for trust. If an explanation is wrong, the engine is still right, and a teacher can see the difference. Our public World Climate Simulator is an example of the first layer only: an educational strategy experience with its own rules. We do not describe it as having an AI engine, and an explanation layer for it would be new development work with partners.
What international guidance asks for
UNESCO’s guidance on generative AI in education and research is useful precisely because it is cautious. Its emphasis is on a human-centred approach, protection of learners’ data and privacy, and validation of tools by institutions before wide use. It does not promise that AI improves learning by default. For a simulation project, we read that as three practical commitments: teachers decide how and when the explanation layer is used, data collection is minimal and agreed in advance, and the institution tests the tool in its own context.
The engine must be right on its own. The explainer only earns its place if learners and teachers find it useful.
A design checklist we use
- Name the decision. Write one sentence describing the choice learners practise and why it is hard.
- Write the rules down first. Document variables, assumptions and known simplifications before any interface work.
- Decide what the AI may and may not say. For example: it can describe changes and pose questions, but cannot give grades or invent data.
- Give teachers a switch. The explanation layer can be turned off per group, and its output is visible to staff.
- Collect only what you will analyse. Agree logging, retention and consent with the institution before the pilot.
Proposing an evaluation, not promising an effect
We do not claim that simulations with AI explanations improve learning outcomes. That has to be shown, in each setting. A modest pilot design could look like this:
- Two comparable groups use the same simulation; one sees AI explanations, one sees fixed, teacher-written feedback.
- Teachers review a sample of AI explanations for accuracy and tone, and record every error they find.
- Learners complete a short transfer task outside the simulation, so that we measure reasoning rather than game skill.
- Results, including null or negative ones, are reported to the partner institution before any wider rollout.
If the explanation layer does not help, the simulation still stands on its own. That is the advantage of building it in two layers.
Where to go next
Our AI-Assisted Training and Smart Assessment notes describe the in-development components this article refers to. For the environmental side of the same method, read what climate policy simulations can and cannot tell us.
