Maybe the most important skill with AI is the ability to teach

When people talk about the skills you need to work with AI as a co-scientist, they often mention critical thinking and continuous learning. I would add a third that gets far less attention: the ability to teach.

Consider what the best teachers do. They diagnose where someone’s understanding breaks down. They ask questions that expose assumptions. They know their material well enough to catch a confident answer that turns out to be false. They withhold the answer at the right moment so the learner builds it themselves. And then they reflect on the learning with the learner after the fact. Robert Kegan, with whom I studied at Harvard, called this “meeting people at the growing edge of their understanding.”

Every one of those actions is what good collaboration with an AI requires. You scaffold the problem. You interrogate the output. You supply the context that the model lacks. You catch the fluent confabulation because you know the terrain well enough to see the seam. You then ask the AI to reflect on what was learned in the session to inform the future. Prompting, at its best, is a form of teaching.

It becomes a little dystopian when you have folks “teaching” an AI to replace them on the job. It’s also troublesome when the capacity to teach a subject and the capacity to evaluate a machine’s version of it draw on the same well. Domain mastery is what lets you tell a real insight from a plausible (or fake) one. We still must prioritize subject matter expertise alongside teaching skills.