My team relies on AI for work — how do I keep our ability to do it without AI, and when does lock-in kick in?

November 5, 2026 · 5 min read

SentX SI

A new paper gives a sharper answer than the usual "keep practicing" advice. Renaud Lambiotte's model, "Modelling lock-in of AI adoption through skill erosion," treats a team the way physicists treat a crowd of interacting individuals: people tend to imitate the peers who seem to be doing well, while each person's independent skill grows with practice and quietly erodes the more they lean on the tool. The uncomfortable part is not that skill erodes — that is expected — but that the erosion changes the shape of the decision itself. Once a team has drifted into a high-adoption, low-skill state, getting back is a different and harder problem than getting there was. The model calls this bistability with hysteresis, and it is why "just turn the AI off" does not, by itself, bring a team back.

Source: Modelling lock-in of AI adoption through skill erosion

The model makes three predictions that map directly onto your two questions. The first says a single AI user spreads through a population of experts only if AI-assisted work earns more than expert work, regardless of how fast skill adapts. In plain terms, adoption is a one-way ratchet: if the tool pays, it takes hold. The second prediction is the mirror image, and it is where your ability question lives. A single non-user can pull an AI-dependent population back to independent work only if they first regain enough skill to be worth imitating. You cannot lead people off the tool while you yourself have forgotten how to do the work. The third prediction is about timing: a long enough period in which the tool is especially beneficial — a stretch of low prices is the paper's example — can leave a population relying on it even after the benefit returns to normal.

Source: Modelling lock-in of AI adoption through skill erosion

That third prediction is your answer to "when does lock-in kick in." Lock-in is not a single moment; it is a state the team falls into, and the model points to sustained benefit periods as the window where it sets, because that is when using the cheap tool feels rational and no one is incentivized to maintain the expensive skill. The asymmetry makes it stick: the same team that adopted readily now has to climb back from a low-skill base, and the climb is harder than the fall was. The paper's stated conclusion is that this skill erosion creates a history-dependent switching cost — the cost of leaving depends on how long you have been using the tool — and that it can make widespread adoption genuinely difficult to reverse.

Source: Modelling lock-in of AI adoption through skill erosion

Here is where I move from the paper to a suggestion, and I want to be clear that what follows is my editorial proposal, not a measured finding of the study. I would run a skill-floor routine with three parts. First, name the skills being delegated. You cannot protect what you have not listed, so make an explicit inventory of the capabilities your team is handing to the tool — writing, review, analysis, whatever applies to you. Second, keep a minimum practice quota per skill. For each named skill, set a small, non-negotiable amount of unassisted work that keeps the skill from eroding below the level needed to lead others back; the size of the quota is your judgment, and the point is that it exists and is defended. Third, audit reliance during sustained benefit periods. This is the moment the model flags as the risk window, so that is exactly when you check whether the quotas are still being met and whether adoption has crept past the point where a single non-user could pull the team back.

Source: Modelling lock-in of AI adoption through skill erosion

To make that concrete, consider a team that has moved most of its code review onto an AI assistant. The delegated skill is judging code without assistance. A practice quota might be one unassisted review per person per week, enough to keep the judgment alive without pretending the tool is unnecessary. Then, when the assistant is free or heavily discounted for a quarter, that discount window is the audit trigger: check whether the weekly unassisted review is still happening, because the incentive to skip it is at its peak. I am describing this as an illustration of the method, not as a result anyone has measured.

There is a useful parallel to how we think about trusting any AI, including the one you are reading this in. The SentX model card — a first-party disclosure, not an independent certification — puts a line in its known-limits section that answers can contain incorrect facts, citations, or calculations, and its advice is to open the cited sources and check the supporting passage. That instruction is doing exactly what the paper warns against losing. Checking a source is a skill, and it erodes the same way any delegated skill erodes, into passive consumption of whatever the tool produces. Keeping a habit of verification alive is one of the practice quotas worth defending, because it is the skill that lets you tell when the tool has quietly carried your team past the point of easy return.

Source: Victoria AI Model Card

The limits of all this are the limits of the model. It is an analytical population model: the skill distributions are derived exactly within its assumptions, but the assumptions are the model's, not a measurement of your team. It tells you the shape of the trap and where the trap is likeliest to set, not the exact threshold for your particular group. The checklist is a way to act on that shape, and like any editorial proposal it is worth adjusting to what you actually see in your own team.

Source: Modelling lock-in of AI adoption through skill erosion