I spent the last week of August in Switzerland, working with a school on what assessment looks like in the age of AI. The last time I was there was two years ago, with the same school, on my first on-site in this work.

That first visit, I was showing people how to open the tools: where the chat box was, what a prompt was, and why the first answer was rarely the one you wanted. When I came back two years later, many of those same teachers no longer needed help using the tools. The conversation had shifted from how do I use AI? to what does AI mean for the way I design assessment?

Different countries, different systems, different constraints since then, and the same five conversations have come up in all of them.

1. Start with the work, not the tool

What works is giving staff time to build around the work they are already doing. That might mean creating something in Google AI Studio, building a chatbot around an existing process, or adding AI into a system they already use. The important part is that they are not starting with a tool and asking, What could I do with this? They are starting with something in their own practice that they want to make better and then building toward it.

When the starting point is their own workflow, they make different decisions. They know what the tool needs to do, and what would actually make it useful in their context. Lead with the tool instead and you are asking them to invent a reason to use it.

2. Exploration time is not integration support

Which raises the question of where that time comes from, and here schools tend to fund the easier half. Exploration time is the sandbox: no deliverable, no share-out, permission to be bad at it for an hour. Teachers ask for this constantly and rarely get it in a form that is actually protected from other agenda items.

Integration support is what comes after, and it is harder to provide because it is specific. It is sitting with a Grade 4 teacher and her actual unit and working out where this belongs, where it does not, and what needs to change for it to actually fit the work. Without that, the enthusiasm from the sandbox does not last past October.

3. Teachers want to shape the tool

The same ownership question shows up in the tools themselves, and I underestimated how much until I heard it everywhere. The version I hear most often runs something like this: if I am going to put this in front of my class, I want to change how it talks to them. I want to edit the feedback it gives. I want to decide that it asks a second question before it explains anything. I do not just want to prompt it once and hope; I want to author how it works.

The tools that get used are the ones teachers can shape around their own practice. Anything they can only run as given is much more likely to get piloted politely and abandoned. Professional judgment is the whole job, and asking a teacher to set it aside and trust the defaults is asking for the wrong thing.

4. Decide what stays human

Judgment also decides the harder question, which is what we are protecting in the first place. So I ask teams what they are trying to amplify rather than simply what they can hand over. What makes this work valuable in the first place? Are we using AI to strengthen that, or are we accidentally designing around it?

Saving an hour of formatting resources is a good trade. Saving an hour by outsourcing the decisions about what students need next is a different transaction, and it should be recognized as one before it is made. Efficiency is useful, but it is not neutral. Someone still has to decide which parts of the work are worth making faster and which parts are worth staying close to.

5. Detection is not the answer

Assessment is where this conversation has moved the most since my first visit. MIT’s Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training published its report this month, and it is blunt about detection. The committee recommends against relying on detectors. They invite an arms race, can disproportionately flag the writing of non-native English speakers and neurodivergent students, and turn assessment into an adversarial exercise.

What they recommend instead is going back to what the assessment was for. Decide what students should know and be able to do, then ask whether the thing you are putting in front of them actually gives you evidence of that.

That question matters even more now. If AI can complete an assessment easily, the answer is not automatically to find a better way to police AI use. Sometimes it is worth asking what the assessment was measuring in the first place. Was it asking students to think, make decisions, defend an idea, or apply something in a new context? Or was it mostly asking them to produce a finished product that now tells us very little about how they got there?

So the question is not how do I stop them? It is is this assessment still worth doing, and if AI can complete it, what does that tell me about what I am actually asking students to do? Every good assessment conversation I have been part of eventually arrives there. The schools that get somewhere are the ones willing to stay with that question.

Two years later

What we can do with AI has gotten better, and so have the ways we can put it to work in a classroom. What has struck me more is that as the tools got stronger, educators did not get quieter. They are advocating harder for keeping the human judgment, relationships, and thinking that make the work matter in the first place.

That is what I would tell anyone still deciding how to approach this work: the important question is not how much AI to let in. It is what you want to preserve in teaching and learning, what you want AI to strengthen, and where professional judgment still needs to lead.

Two years in, that feels less like a question about technology and more like a question about what kind of practice we want to build. That makes me optimistic. The tools will keep changing, but educators are getting clearer about what they want them to strengthen and more confident about shaping the technology around the work that matters.