The mind that outsources thinking

The quiet transfer

There is a moment, easy to miss, when a difficult problem no longer reaches the part of the mind that used to struggle with it. The cursor blinks. The model is ready. What used to be a few minutes of friction becomes a few seconds of acceptance.

That transfer feels like relief. Often it is. Relief is not the whole story.

Cognition is practice-dependent. The mind stays sharp where it still has to work. When drafting, structuring, and second-guessing migrate outward, the inner rehearsal that once built judgment begins to thin. We do not usually notice the thinning on the day it starts. We notice it later — when the model is offline, or wrong, or confidently incomplete, and the old path back into the problem has grown overgrown.

What is being traded

Tools that summarise, draft, and decide change what humans rehearse. Intermediate skill is especially exposed: the stage where people once struggled productively, made partial maps, and learned the shape of error. Fluent output can create a metacognitive illusion. Confidence rises faster than understanding.

Bjork and Bjork (2011) describe “desirable difficulties”: friction that feels inefficient and builds durable learning. Generative interfaces are good at deleting that friction. Whether they should is a design choice, not a law of progress.

This is not an argument against assistance. It is an argument about which difficulties remain human practice.

Evidence, held carefully

Workplace studies of generative tools are still early, industry-specific, and often optimised for short-term throughput rather than long-term skill. Universal productivity claims should be read as marketing until longitudinal measures of judgment exist.

Sparrow, Liu and Wegner (2011) showed, in a different technological generation, that when people expect to have later access to information they recall the information itself less well and the location of it better. The paper is a classic, not a complete theory of large language models. It is cited because it names a bargain we are still making: knowing that, knowing how, knowing where.

What careful practice looks like

Keep a private draft before you ask the machine. Delay the first suggestion until you have named the problem in your own words. Treat model output as a hypothesis, not a verdict. Design workflows so that verification is a first-class step, not an optional polish.

Close

Technology rarely steals skill in a single dramatic act. It rearranges what we rehearse. Protect the difficult practice that builds judgment. Use machines to clear the path — not to walk it for you.