Losing a skill you still need

The aviation precedent

Aviation automated earlier and studied the consequences harder than any other industry, and its findings transfer.

By the 2010s regulators were formally concerned that pilots flying highly automated aircraft were losing manual flying proficiency. A US Federal Aviation Administration working group reported in 2013 that pilots sometimes lacked sufficient manual handling and that automation dependency was a real risk, and recommended more hand-flying in line operations. The concern was specific: automation handles the routine perfectly and disengages in the unusual situation, which is precisely the situation requiring the skill that has not been practised.

That asymmetry — automation covers the easy cases and hands you the hard ones — is the general shape of the deskilling problem, and it applies to a radiologist, a translator, a junior lawyer and a programmer.

Three mechanisms, which need different responses

Skill decay. Abilities that are not exercised degrade. This is ordinary, well-established, and reversible with practice. It affects mainly the production skills: writing a first draft, deriving a result, composing a query.

Skill non-acquisition. More serious, and it affects new entrants. A junior who has never written the tedious version does not have a skill to decay — they never built it. The tedious work was the training. The paradox is direct: the tasks most worth automating are the tasks juniors learned on.

Judgement erosion. The slowest and hardest to see. Judgement is built from many exposures to cases and their outcomes. Someone who reviews model output builds a different, thinner set of exposures than someone who worked each case through, and they may not notice for years, because reviewing feels like doing.

The evidence outside aviation

Computer-aided detection in mammography. Studies found that readers' behaviour reorganised around the prompts, with sensitivity improving on marked regions and falling on unmarked ones. The aid did not simply add; it redistributed attention.

Endoscopy. A 2025 study in a large European endoscopy setting reported that adenoma detection rates when clinicians worked without AI assistance were lower after a period of routine AI use than before it — an observational finding, with the usual caveats, and the first direct measurement of deskilling from clinical AI use.

Navigation. Habitual satellite-navigation use is associated with poorer spatial memory and worse performance on unaided wayfinding. Modest effects, consistently found.

None of these says the tools are bad. They say the unaided capability is a separate quantity from the aided capability, and it moves independently.

Which skills to protect

You cannot maintain everything, so choose deliberately using two questions.

Will I need this when the tool is unavailable or wrong? Unavailable includes an outage, a client site with no network, an exam, a jurisdiction where the tool is not approved, and a moment when you must act now. Wrong includes every case in this course.

Is this skill the foundation of my judgement in this field? A doctor's examination skills, a lawyer's reading of a primary source, an engineer's feel for magnitudes, a writer's sense of structure. These are load-bearing, and their loss is not visible until something depends on them.

Skills that fail both tests can be let go without anxiety. Nobody needs to maintain their long division.

A protocol that works

Unaided repetitions, on a schedule. A fixed proportion of real work done without the tool. Ten per cent is enough to notice decline. Doing it on real work rather than exercises is what makes it survive a busy month.

Attempt first, then compare. Produce your version, then generate one, then diff them. This preserves the retrieval practice that builds skill and adds the feedback that accelerates it — it is the single best pattern in this lesson, and it is the same rule as in the studying lesson.

Periodic honest tests. Twice a year, do a representative task cold, timed. Compare with your memory of how it used to go. Unpleasant and informative.

For teams: protect the training path. If juniors never do the work that built senior judgement, the organisation has a five-year problem it cannot see this quarter. Some firms now deliberately reserve categories of work for people learning. That is a cost, taken on purpose, and it is cheaper than the alternative.

The line to hold

Use the tool for the work. Keep the capability that the work was building. Those are two different objectives, and only the first happens by itself.