Two years ago our faculty had the argument every school is having. Half the room wanted the tools banned; half wanted them embraced. Both positions were about the tools, which was the mistake.

The real question is what we are assessing. If an assignment can be completed by a model in nine seconds, the assignment was measuring the wrong thing — and it probably was before the models existed.

So we changed the assessments first. More work is now done in front of us: defended in a seminar, built in a lab, presented to a panel that asks follow-up questions. Written work still matters enormously, but it now sits alongside a conversation about how it was made.

Then we taught the tools properly. Students learn what a language model is actually doing, where it is confidently wrong, and how to check it. They learn that a citation it produced may not exist. They keep a record of what they asked and what they changed.

The surprise was that the strongest students became the most sceptical. Once you understand the mechanism, the output stops looking like an authority and starts looking like a draft.

We also teach the ethics as a first-class subject rather than a footnote — who is in the training data, who is not, and who carries the cost when a system is wrong about a person.