We Built Forklifts, Then We Built Gyms

Offloading is not new and mostly not bad

The alarm about cognitive outsourcing is at least as old as literacy. In the Phaedrus, Plato has Socrates complain that writing will “create forgetfulness in the learners’ souls,” producing people who appear knowledgeable while knowing nothing. He was, on the narrow point, correct — we did stop memorising epics — and on the broad point spectacularly wrong.

The pattern has repeated with maps, printing, arithmetic, calculators, and search. In each case a capacity was surrendered and something larger was purchased with the surrender. Cognitive scientists call the general phenomenon cognitive offloading, and the position that this is normal rather than pathological has strong support: Clark and Chalmers’ extended mind thesis argues that the notebook in your pocket is genuinely part of your cognitive system, not a crutch attached to it.

So the question is not whether to offload. You already do, constantly, and should. The question is which offloads are cheap and which are expensive.

The two tests that separate them

Most public argument about AI and the mind fails because it treats all offloading as equivalent — either all fine (calculators!) or all catastrophic (Plato!). The useful position is neither. It is that a specific, identifiable minority of cognitive skills are both self-verifying and generative, and those are the ones requiring deliberate maintenance. Everything else can go, with thanks.

Desirable difficulty

Robert Bjork’s research programme produced a finding that is genuinely counterintuitive and extremely well replicated: conditions that make learning feel harder and slower often produce better long-term retention and transfer. Spacing beats massing. Interleaved practice beats blocked practice. Retrieving from memory beats re-reading. Generating an answer beats being shown one.

He called these desirable difficulties, and the crucial part is the perceptual trap that comes with them: learners consistently rate the harder conditions as less effective while performing better on them later. Our sense of how well we are learning is systematically inverted.

Every AI interface is optimised to remove friction. That is what makes them good products. It also means they are optimised, without intending to be, against the exact conditions under which humans learn best. There is no villain here — just a misalignment between what a good interface does and what a growing mind needs.

What a regime looks like

Borrow the structure from physical training, because it transfers surprisingly well.

• Specificity. Train the capacity you want to keep. If you want to keep reasoning, do reasoning — not puzzle apps. The transfer literature on general “brain training” is discouraging; the literature on domain practice is not.
• Progressive overload. The difficulty has to increase or adaptation stops. Re-reading familiar material is the cognitive equivalent of lifting the same light weight forever.
• Recovery. Consolidation happens in the gaps — sleep, walks, the shower. Filling every interstitial moment with input is over-training.
• Consistency over intensity. Thirty minutes daily beats one heroic Sunday, for the same reasons it does in the gym.

Half right, and the half it gets right is important: effort has no intrinsic value, and cognitive-fitness talk slides easily into puritanism. But the laundry analogy relies on scrubbing being non-generative and self-verifying — clean clothes are clean regardless of who cleaned them. The claim here is narrow and testable: for the small set of skills that are both generative and required to evaluate their own output, the trade is not free, and the cost is paid later, invisibly, at exactly the moment you most need the capability. Everything outside that set — automate it without guilt.

Where cognitive fitness sits on the ladder