The Cognitive Load You Now Have to Choose
AI removes the forced cognitive effort behind almost every task. What you do with the effort it frees up decides whether your mind gets sharper or softer.
A hard email used to cost you twenty minutes of drafting, cutting, and rereading. Now it costs eight seconds and one prompt. The email that comes out is fine. Better than fine, usually.
Something else came out of those twenty minutes that the eight seconds does not produce. Not the email. The version of you that gets slightly better at writing hard emails.
That difference has a name now, and you make the choice around it dozens of times a day without knowing you are making it.
1. Every task used to include a free rep
Every unaided task produces two things at once. The result you wanted, and a small, forced adjustment to the person who made it. Solve a hard problem and you get the answer, plus a mind slightly better calibrated to that category of problem. Write a difficult paragraph and you get the paragraph, plus a hand slightly more practiced at organizing a difficult thought.
That second thing is the rep. Nobody had to choose it. It arrived bundled with the task, the way exercise arrived bundled with a life that required walking somewhere, carrying something, or building something with your hands.
AI is the first tool that reliably separates the two. You can now get the output alone. Calculators did this for arithmetic. GPS did it for navigation. Spellcheck did it for spelling. Each of those unbundled one specific rep from one specific task. AI is different because it does this across writing, planning, analysis, decision-making, and synthesis simultaneously, for nearly anyone, on nearly any task, today.
2. Narrow tools took one rep. This one takes the category
Each narrow tool cost one specific rep and everyone accepted the trade, correctly. Nobody mourns mental long division. The trade always looked the same: give up a rep in a category small enough that losing it barely registered against everything else the mind still had to do unaided.
AI breaks that pattern because the category is not small anymore. It is most of the categories a knowledge worker used to build competence on by default, just by doing the job. That is a difference in kind, not degree, and it is why this now needs a name and a decision, where calculators never did.
3. Offloading is real, studied, and it changes what you remember
This is not a metaphor. A 2011 study published in the journal Science, run by the psychologists Betsy Sparrow, Jenny Liu, and Daniel Wegner, gave people trivia questions and then tested what they remembered. People who believed the answers would remain accessible later recalled the answers themselves worse, and recalled where to find them better. Facing a hard question also primed people to think of computers before anything else, a reflex the researchers measured directly.
The mechanism has a real name in cognitive science: transactive memory, the practice of storing information externally and remembering the address instead of the content. The internet had already made you a heavier user of transactive memory than any generation before you. AI does not introduce this effect. It extends it from facts into reasoning, which is a much larger and more consequential thing to store at an address instead of in your head.
4. The rep you skip is the rep you need at the worst possible time
In 1983, the researcher Lisanne Bainbridge published a paper called Ironies of Automation that has since been cited well over a thousand times. Her argument still holds exactly. The more reliable an automated system becomes, the less a human operator practices the manual version of that skill. The operator’s job shifts from doing the work to watching the system do it, which is a worse way to stay sharp than doing it yourself. Then, on the rare occasion the system fails, the human who has to step in is the least practiced they have ever been, at the exact moment the stakes are highest.
Bainbridge wrote this about industrial and aircraft automation, long before anyone typed a prompt. The shape of the problem transfers directly. A person who has answered every hard question by asking a model has not been failing at anything. They have been quietly moving down the same curve, and they will not find out where they landed on it until a moment when the model is wrong, unavailable, or simply not the tool in front of them, and being right depends on them alone.
“The rep you skip does not disappear. It waits for the one moment you cannot skip it, and hands you back exactly the skill level you left it at.
”
5. Struggle now is the trade for skill later, but only the right struggle
This is not a folk theory about hard work being virtuous. It is a specific, named finding in learning science. Robert Bjork, who has studied this for decades, calls it desirable difficulty: a practice condition that slows you down and feels worse right now, but produces learning that lasts and transfers to situations you have not seen before. Spacing your practice out, testing yourself instead of rereading, generating an answer instead of recognizing one. All of these make the moment feel harder and the result, later, measurably better.
Bjork’s own research includes the caveat that matters most here. A difficulty is only desirable if you can actually overcome it. Struggle that is too hard, or struggle unrelated to the skill you are trying to build, does not produce this effect. It just produces frustration. This is the honest limit of the whole argument, and it is worth sitting with before you go looking for load to add back.
6. Choosing what to keep
You cannot keep every rep. Nobody has the hours, and most tasks genuinely do not deserve the effort. The choice is not more effort everywhere. It is picking, on purpose, the small set of tasks where the rep is building a capability you are actively relying on, where feedback arrives fast enough to correct you, and where being wrong today is cheap. Everything outside that set, hand to the model without a second thought. Nothing is proven by drafting a routine email by hand.
Inside that set, do it slower, on purpose, before you check what the model would have said. Not forever. Long enough that the skill stays live instead of becoming an emergency you have not had yet.
What nobody has solved
There is no dashboard that tells you which of today’s tasks were the ones worth the friction. That judgment is still yours, made task by task, and you will get it wrong sometimes in both directions: doing too much by hand out of habit, or handing over the one thing you actually needed to keep. The honest position is that this is a skill on its own, choosing the load, and like every other skill it gets better with reps you cannot outsource.
