What types of thinking should you NOT outsource.

Calculators didn't make everyone innumerate. GPS navigation made driving easier. Both arrived with a small tax: a measurable atrophy of the skill the tool replaced. When researchers programmed calculators to give undergraduates wrong answers, the students didn't spot it - even when the answers were absurd1. People who relied heavily on GPS over their lifetimes had measurably worse spatial memory; in a three-year follow-up, heavier GPS use predicted a steeper hippocampal-dependent decline2. Looking things up online does the same thing to memory: when we expect future access, we encode where to find an answer rather than what the answer is3.

These costs were small enough to ignore. The thing on the calculator's screen was usually right. The road home was usually the road home. The web, despite its tax on memory, made people more capable in most ways that mattered. Cognitive offloading has been a worthwhile bargain for fifty years. AI changes the terms of that bargain4.

In a recent set of preregistered experiments at Wharton, Steven Shaw and Gideon Nave gave volunteers cognitive-reflection problems and let them consult an AI assistant whose answers had been silently seeded with errors. When the AI was right, people using it outperformed an unaided control by about twenty-five points. When the AI was wrong, the same people underperformed by about fifteen5. The authors call this cognitive surrender. People stopped checking.

The dependence would matter less if assisted work at least trained the skill. It does not. In a BCG field experiment, consultants used ChatGPT on data-science tasks beyond their skillset and approached expert performance. Tested later without the tool, they did no better than peers who never had it6. The authors call GenAI an exoskeleton: it carried the task and taught nothing.

EEG studies of multi-session essay writing find that ChatGPT users show the weakest brain connectivity, the lowest engagement, and - when later asked to write without the tool - remember little of their own prior text7. The web at least made you encode where the answer lived8. A model that hands you the answer in conversation makes even that footprint optional3.

The literature on cognitive offloading is older than the chatbot. It separates into four mechanisms - four different things that go wrong when you let a tool think for you. Any given AI use case combines several. Naming them helps decide which skills are at risk and which aren't.

The Shortlist taxonomy organises 75 thinking skills across five domains - cognitive, social, executive, operational, creative. Each skill is scored on two axes:

• Training importance - the empirically-supported contribution of this skill to decision quality23, durable learning27, leadership21, and self-regulated performance28. A 5 means the skill is a top-tier predictor of outcomes in the empirical literature; a 3 means it's contextually important; a 1 means it's easily compensated or domain-specific.
• Atrophy risk - how prone the skill is to be weakened, misled, or abandoned through AI-assisted work, weighted by the strength of the evidence behind the score.

Crossing the two thresholds (importance ≥ 4, atrophy ≥ 4) gives four bands. The Defend band - high importance, high atrophy - is where deliberate practice has the highest expected value.

Calculators were a good trade. So is unit conversion. So is route-finding in unfamiliar cities, most of the time. People have offloaded since the first abacus. The choice is which cognitive operations to keep resident, and why.

The lanes below sort each skill into one of three bands by signed offloadability: don't offload (documented harm or skills the model can't replace), with oversight (offload first drafts and rote work - verify the result), and offload freely (AI matches a competent practitioner).

Skills that AI replaces competently share a property: they are well-defined operations on well-defined inputs. Arithmetic. Translation. Lexical search. Boilerplate drafting. Skills that AI cannot replace - even when it produces output that looks like the replacement - share a different property. They require choosing under uncertainty between irreversible options with conflicting values. That is the definition of strategy21.

Frank Knight, a century ago, distinguished risk (measurable probabilities) from uncertainty (the unmeasurable kind, where the categories themselves are in motion). Statistical machines have traction in the first and no traction in the second22. Kahneman and Klein, decades later, named the boundary in modern terms: reliable expert intuition requires high- validity environments and ample feedback. In the low-validity environments where most strategic choices live - what business to enter, who to hire, when to relent in a negotiation - confident pattern-matching is overconfident pattern-matching23.

Carissa Véliz argues that predictions about people are speech acts, not descriptions. They bend the world toward themselves. An AI that predicts a bleak financial future for someone can push that person into economic disadvantage by denying them credit. A hiring AI achieves "accuracy" in part by creating the reality it claims to predict - and self-fulfilling prophecies produce no error signal24. The ethical question of strategic choice - should we predict this at all - sits at exactly the place where the model has nothing to say25.