"Use it or lose it" is one of those phrases that sounds like folk wisdom and turns out to have a careful research base behind it. The skill-decay literature is decades old, methodologically diverse, and surprisingly consistent. It does not say what the strongest readings imply — that disuse is catastrophic — but it does say what the weakest readings deny: that disuse is real, measurable, and follows a predictable pattern.
This post walks through the canonical evidence, the inverse case from the deliberate-practice tradition, and why the AI moment makes both newly relevant.
What the skill-decay literature actually shows
The canonical reference in the skill-decay literature is Arthur, Bennett, Stanush, and McNelly's 1998 meta-analysis in Human Performance, 11(1) — "Factors that influence skill decay and retention." The meta-analysis collected studies across military, industrial, medical, and educational training, and quantified how performance on a previously learned skill changes as a function of how long the skill has gone unused.
The findings, summarised in broad strokes:
- Skill decay is real and measurable. Performance on a previously learned skill declines with disuse, in a curve that varies with skill type but is consistently negative.
- The curve is sharper for cognitive skills than for motor skills. Procedures, factual knowledge, and decision rules tend to decay faster than physical actions like riding a bicycle. This is intuitively familiar — most people can still ride a bike years later, but few can still solve calculus problems they once solved easily.
- The amount of original practice matters. Heavily over-learned skills decay more slowly than skills that were only practised to initial proficiency. Practising to mastery is partly an investment in retention.
- Retrieval cues matter. Decay is reduced when the testing context resembles the original learning context, and exaggerated when it does not.
The literature is not about loss of capacity. It is about access. The skill is not deleted; the retrieval pathway becomes harder to use without recent practice. This is also the framework that explains why a few hours of refresher practice tends to recover most of what looks like a deep loss — the underlying competence is largely intact, just less accessible.
A note on the citation: the Arthur et al. 1998 paper appeared in Human Performance, 11(1). The exact title and journal name are sometimes garbled in secondary sources. If you cite it, cite the journal and volume accurately.
The inverse: deliberate practice
The other side of the same coin is the deliberate-practice tradition. The foundational paper is Ericsson, Krampe, and Tesch-Römer's 1993 paper in Psychological Review, "The Role of Deliberate Practice in the Acquisition of Expert Performance" — one of the most cited papers in cognitive psychology and the source of the cultural idea that expertise comes from sustained, focused, effortful practice over time.
Deliberate practice has four properties in Ericsson's formulation: it is goal-directed, it requires effort, it includes feedback, and it focuses on the specific weaknesses of current performance rather than coasting on what is already easy. Music, chess, sports, and surgery were the original domains studied. The pattern generalises to most skill domains where performance can be measured.
The 1993 framework has been refined since. Macnamara and Maitra's 2019 paper in Royal Society Open Science revisited the original Ericsson, Krampe, and Tesch-Römer study — a notable re-examination of the data and methods. The result is a more nuanced picture: deliberate practice matters substantially, more in some domains than others, and is one of several factors that contribute to expert performance.
The synthesis of the decay and deliberate-practice traditions is straightforward. Skills that receive regular, focused, effortful reps maintain and develop. Skills that do not receive those reps slowly become harder to access. This is true across cognitive and motor domains. It is true across age ranges. It is the empirical backbone of "use it or lose it."
Why AI accelerates the pattern
The skill-decay literature predates the AI conversation by decades. The relevance to AI is that AI changes the daily-rep count for a wide range of cognitive skills.
The clearest analogue case is the 2020 UCL GPS-and-spatial-memory study in Scientific Reports. Heavier lifetime GPS use predicted measurably worse spatial memory on independent navigation tasks. The mechanism is straightforward: GPS replaces the reps that built and maintained the underlying skill. The skill is not lost in a single trip; it is gradually less practised across years of cumulative tool use.
The same logic extends to AI. Each time AI drafts a sentence, the user's drafting skill receives one fewer rep. Each time AI summarises an article, the user's reading-and-synthesis skill receives one fewer rep. Each time AI suggests the next line of code, the programmer's recall of that pattern receives one fewer rep. In any single instance, the cost is invisible. Across thousands of instances over years, the cumulative pattern is exactly what the skill-decay literature would predict.
The 2024 Springer paper on AI overreliance in Smart Learning Environments ("The effects of over-reliance on AI dialogue systems on students' cognitive abilities") is one of the first careful empirical entries on the student-and-AI version of the question. The 2011 Sparrow Google-effect paper in Science is the analogue case for the search-engine version.
What is new in 2026 is not the skill-decay mechanism. The mechanism is forty years old. What is new is the breadth of cognitive work being mediated by tools that genuinely substitute for the user's reps. That is the structural reason "use it or lose it" is having a moment in the AI conversation.
What the literature does not say
It is worth flagging what the skill-decay framework does and does not support.
It does not say that disuse causes catastrophic loss. The decay curves in Arthur et al. are gradual. Most skills can be recovered with a fraction of the original practice time. The honest framing is "harder to access without recent practice," not "permanently gone."
It does not say that AI use causes broad cognitive decline. The literature is about specific skills, not general cognition. If a particular skill is no longer being practised, that specific skill becomes less accessible. Other skills are unaffected.
It also does not say that anyone should stop using AI tools. Like the rest of the offloading literature, the skill-decay framework is neutral about whether the trade is worth making. The trade is worth making in many cases. The practical question is whether the skills you actually want to keep are still in regular practice somewhere in your day.
What this means for daily practice
The practical synthesis of the decay and deliberate-practice traditions is simple. Skills you want to keep need reps. The reps need to be regular, effortful, and matched to the specific skill. They do not need to be long. A few minutes a day across a set of skills is, on the deliberate-practice evidence, more useful than a longer session once a week. The cumulative effect of small daily reps is what the literature most consistently supports.
That is structurally the Senwitt bet. A short, daily, mixed Set of unaided cognitive practice across writing, math, code, memory, reading, and reasoning — the six Senwitt Skills — keeps those specific skills in regular use. It does not replace any of the lifestyle factors that drive cognitive aging. It does not claim to prevent decline of anything in particular. It is the daily-rep habit applied to the thinking skills the AI environment is most likely to underfeed.
The research/skill-disuse page is the formal explainer for the framework, and the research/scope-of-evidence page is the careful version of what we will and will not claim about it.
