For more than a decade now, I have navigated four countries without really learning to find my own way in any of them. El Jadida to St Louis; St Louis to Brussels; Brussels to Zurich. In every one of those cities, the same small object got me from the airport to the flat, from the flat to the school gate, and back again: my phone, Google Maps open on it.

I do not know Zurich the way I know the town I grew up in; the backroads, the shortcuts, the alternative routes that are better at certain times of day, or the ones with a quirky stop en route. After five years in Zurich, I left knowing it reasonably well, but only because I had repeated the same handful of Google Maps routes often enough for them to stick.

That is not a confession about my sense of direction; it is a confession about what happens to a capability the moment its effort is handed to something else, for long enough. Wayfinding is cognitive load too, and I have been offloading mine, one turn-by-turn direction at a time, for ten years.

Every time we hand a task to AI, something feels easier, but more than this, we feel a sense of achievement; perhaps it’s the increased efficiency or the buy back of time that you get, but you get something in addition to not having to think. The juxtaposition of efficiency and offloading is exactly the problem. Optimisation at its best, is allowing your brain to focus on more important things than menial tasks. But those menial things stack; they layer on top of one another and before you know it; you are completely reliant on a piece of technology to get you from A to B, as I am.

Cognitive load, the demand a task places on working memory, is usually treated as pure cost: friction to be minimised. But half a century of learning science says part of that load is not waste at all. It is the place where understanding is actually built.

Cognitive load theory (Sweller, 1988) separates the demand a task makes into three kinds. Intrinsic load is the difficulty inherent in the material. Extraneous load is the effort wasted on poor design, clutter, and distraction. Germane load is the effort you spend building and refining mental models; the effortful processing that is learning. Good instruction has always meant cutting the extraneous so the germane can do its work.

AI is the most powerful extraneous-load remover ever built. It drafts, summarises, formats, and retrieves, clearing away exactly the friction good design tries to clear, and of course, optimising your time and workload. But it does not stop there. Left unchecked, it also removes the germane load: the discerning, the weighing, the deciding. And because nothing feels lost when it does, the answer still arrives, faster and cleaner, the erosion is silent. This is a practitioner framework, not a finding: a way to organise what cognitive load theory and the early evidence on AI and cognition, taken together, suggest about using these tools without hollowing out the capability they are meant to support.