How skill loss could make AI adoption hard to reverse

If people lose practice while using AI, returning to independent work may become harder even after the tool’s initial advantage fades. A new theoretical study explores how that feedback could turn individual choices into collective lock-in—not as a prediction already established by data, but as a mechanism to investigate.

AI adoption is often framed as a choice about present performance: does using a tool produce a better result than working without it? Renaud Lambiotte’s preprint asks what happens when that choice also changes a person’s future ability to work independently. In the model, reliance can become self-reinforcing because the alternative grows less attractive as skills erode.

The study combines mathematical analysis with simulations. It has not been calibrated against real-world data, so its findings describe what can happen under specified assumptions—not how much skill current AI users actually lose.

A choice that changes the chooser

Each simulated individual either works independently or uses AI, and carries an independent skill level between zero and one. Independent practice gradually raises that skill towards one; AI use lets it decline towards zero. These are modelling assumptions, rather than measurements of learning or forgetting.

Independent work earns a payoff equal to the person’s skill. AI-assisted work receives a benefit reduced by a crowding penalty: as more people adopt the same tools, their outputs are assumed to become less distinctive. In the main model, AI-assisted payoff does not depend on the user’s own expertise.

Individuals periodically reconsider their behaviour. They tend to imitate peers with higher realised payoffs, while occasionally choosing randomly. Skill therefore records a personal history that current behaviour alone cannot reveal: two AI users may retain very different capabilities depending on how long they have relied on assistance.

Lambiotte derives exact stationary skill distributions for an infinite, well-mixed population under the main model’s linear imitation rule. These calculations identify possible stationary states and where branches merge; their stability is inferred from simulations.

Two possible outcomes under the same conditions

Two feedbacks compete. Skill erosion strengthens the relative appeal of AI, while crowding reduces its advantage. When erosion dominates sufficiently, the model can support both a low-adoption, high-skill state and a high-adoption, low-skill state at the same AI benefit.

This creates hysteresis: the outcome depends on the route taken. Raising AI’s benefit can trigger a sudden jump in adoption, but restoring the previous benefit need not restore independent work. The population’s capabilities have changed during the intervening period.

Simulations of temporary benefit increases illustrate this mechanism. Starting from low adoption, the study raises the AI benefit and later returns it to its original value. Short increases have little lasting effect, whereas sufficiently long increases can leave the population near high adoption. A temporary period of lower prices is offered as one possible interpretation of a benefit increase, not as a separately modelled pricing strategy.

“Lock-in” has a bounded meaning here. In these experiments, it is assessed after 400 model time units following the benefit’s restoration. With finite populations and continuing random exploration, fluctuations could eventually allow escape. The simulations do not establish permanent or inevitable dependence.

Why persistence can help people return

An important asymmetry appears when random exploration becomes rare. A lone AI user can spread through a population of fully skilled independent workers only when AI-assisted work pays more than expert work. This threshold does not depend on how quickly skills adapt relative to behavioural switching.

The reverse transition is different. Someone abandoning AI in a deeply deskilled population initially performs poorly. Independent work can spread only if such holdouts recover enough skill and attract enough imitators before returning to AI.

Behaviours that persist longer relative to skill adaptation give holdouts more opportunity to recover. This makes the return to independent work easier and narrows the range supporting two stable outcomes. Yet the same persistence also lets continuing AI users lose more skill, increasing polarisation between users and non-users.

Adoption figures miss part of the story

The model shows why averages can conceal vulnerability. In one comparison, populations are tuned to 40% adoption and an average skill of 0.6. Changing the relative timescales transforms the skill distribution: the predicted share below skill 0.2 ranges from effectively zero to about one-third, despite identical adoption and mean skill.

The conclusions remain conditional on simplified learning, imitation and interaction rules. Alternative imitation rules and unequal learning and erosion rates preserve the main qualitative effects in the reported checks, but change numerical outcomes. An extension in which AI rewards users’ expertise weakens the feedback towards dependence and narrows the bistable region.

The practical implication is a measurement question, not a demonstrated forecast: understanding dependence may require tracking what people can do without AI, alongside how often they use it.