Modelling lock-in of AI adoption through skill erosion

Renaud LambiottearXiv 2026 · arXiv

TL;DR

This paper presents a mathematical model showing how generative AI adoption can lead to a "lock-in" effect through the erosion of human skills. By treating skill as a hidden state that evolves based on an individual's history of AI use, the author demonstrates that a population can become irreversibly dependent on AI even if the technology's initial benefits disappear.

What Problem Does This Paper Address?

Generative AI offers a "productivity paradox": it boosts immediate performance but reduces the practice needed to maintain independent expertise. This creates a feedback loop—as skills erode, AI becomes relatively more attractive, further discouraging practice.

The author tackles the question of how this individual-level deskilling scales to a collective level. Specifically, why might AI adoption be "sticky" or hard to reverse? Conventional economic models often treat technology adoption as a choice based on current payoffs. This paper argues that individual memory—the lasting impact of past behavior on current skill—is the missing ingredient needed to understand why populations might stay "locked in" to AI reliance.

Methods

The author uses a coevolutionary game theory framework, where agent strategies (using AI vs. working independently) coevolve with an internal attribute (skill).

  • The Model: Agents choose between two behaviors: Human ($H$) or AI-assisted ($A$).
  • Human payoff: Directly proportional to their current skill $s$.
  • AI payoff: A fixed benefit $b$ minus a "crowding" cost $cx$ (representing loss of distinctiveness as more people use the same tools).
  • Skill Dynamics: Skill increases toward 1 with practice ($H$) and erodes toward 0 with reliance on AI ($A$).
  • Social Learning: Agents periodically revise their behavior by imitating successful peers or through random exploration.
  • Analytical Approach: The author derives exact stationary solutions for the skill distributions of both users and non-users. This allows for a formal analysis of tipping points (bistability) and the conditions under which adoption becomes irreversible (hysteresis).

Datasets

This is a theoretical modeling paper and does not use empirical datasets. The results are validated through agent-based simulations (up to $2 \times 10^4$ agents) to confirm the analytical derivations.

Key Insights and Ideas

  • Skill as a Switching Cost: Skill erosion acts as a history-dependent switching cost. Unlike a one-time fee, this cost grows the longer an individual uses AI, as their "independent" alternative becomes increasingly unviable.
  • Polarization of Expertise: Depending on the ratio of timescales ($\rho$)—how fast skills adapt vs. how often behaviors change—populations can develop very different skill distributions. In "persistent" regimes, the population polarizes into highly skilled experts and deeply deskilled AI-dependents.
Skill distributions for different timescale ratios. At high $\rho$, the population polarizes into experts and deskilled users.
  • Asymmetric Transitions: It is much easier for AI to invade a population of experts than for experts to reclaim an AI-dependent population. A "lone holdout" trying to return to independent work starts deskilled and earns very little, making them an unattractive role model for others to imitate until they have spent significant time practicing.
  • The "Lock-in" Trap: A temporary boost in AI benefit—such as a period of low introductory pricing or a temporary performance spike—can push a population past a tipping point. Once skills have eroded, the population remains stuck using AI even after the benefit returns to its original, lower level.

Key Results

  • Bistability and Hysteresis: The model identifies a "bistable" region where two stable states coexist: one with low adoption/high skill and one with high adoption/low skill.
The hysteresis loop shows that adoption jumps abruptly as AI benefit $b$ increases, but requires a much larger drop in $b$ to reverse.
  • The Lone Holdout Threshold: The paper derives an analytical threshold for when a population can return to independent work. This depends heavily on the persistence of behavior; if people switch back and forth too quickly, they never regain enough skill to "prove" the value of independent work to their peers.
  • Duration Matters: Simulations show that the duration of an AI "benefit boost" is critical. If a boost lasts roughly seven times longer than the timescale of skill adaptation, the population becomes permanently locked in.
Temporary boosts in AI benefit can lead to permanent lock-in if they last long enough for skills to erode significantly.

Limitations and Future Work

  • Simplistic Payoffs: The model assumes AI-assisted work doesn't benefit from human skill (pure substitution). The author notes that "complementary" AI (where experts get more out of the tool) narrows the bistable region and reduces lock-in risk.
  • Homogeneous Population: The model assumes all agents have the same learning and erosion rates. Real-world populations are heterogeneous.
  • Lack of Empirical Calibration: The model provides a theoretical framework but hasn't been tested against real-world data on skill erosion (e.g., in software engineering or writing).
  • Future Directions: The author suggests extending the model to social networks (where imitation is local) and incorporating endogenous AI quality (e.g., model collapse due to training on synthetic data).

Summary updated

Abstract

Generative AI can improve immediate performance but may erode the human skills required for future work. We study a model in which individuals tend to imitate successful peers, while their independent skills grow with practice and erode with reliance on AI. Because each agent's skill reflects their history, non-users and AI users develop different skill distributions, which we derive exactly. Skill erosion can produce bistability and hysteresis between a low-adoption, high-skill state and a high-adoption, low-skill state. When agents rarely change behaviour at random, the two transitions are asymmetric. A single AI user spreads through a population of experts only if AI-assisted work earns more than expert work, whatever the timescale of skill adaptation. A single non-user, by contrast, can lead an AI-dependent population back to independent work only if they regain enough skill before returning to AI, which becomes easier the more persistent behaviours are. As AI improves, adoption can jump abruptly, and a sufficiently long increase in its benefit, such as a period of low prices, can leave a population relying on AI after the benefit has returned to its initial value. Skill erosion thus creates a history-dependent switching cost that can make widespread AI adoption difficult to reverse.
BibTeX
@misc{lambiotte2026modelling,
  author = {Renaud Lambiotte},
  title = {{Modelling lock-in of AI adoption through skill erosion}},
  year = {2026},
  eprint = {2611.04886},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.04886},
}
Figure