The Geometry of Learning Under AI Delegation

1. Problem Statement

The paper addresses a critical tension in the era of AI-assisted learning: while AI tools provide immediate performance gains (accuracy, speed), their repeated use may degrade human skill over time due to reduced practice ("use-it-or-lose-it"). Existing literature is largely empirical or behavioral, lacking a mathematical framework to explain why short-term gains often lead to long-term skill atrophy.

The central research question is: How do repeated AI delegation decisions and evolving AI capabilities shape human skill acquisition over time, and under what conditions do short-term performance gains become irreversible long-term losses?

2. Methodology

The authors model the interaction between a human learner and an AI system as a coupled dynamical system.

Core Model Components

• State Variables:
  • $\theta(t) \in [0, 1]$: The learner's skill level (where 1 is optimal).
  • $p(t) \in [0, 1]$: The probability of delegating a task to AI.
• Parameters:
  • $\theta_a \in [0, 1]$: The effective skill of the AI.
  • $\Delta$: The decay rate of skill when not practicing.
  • $\kappa$: The adaptation rate of delegation.
• Objective Function: Both the learner and the AI update their states to minimize a shared instantaneous performance loss:
  $$\ell(\theta, p) = (1-p)(1-\theta)^2 + p(1-\theta_a)^2$$
  • If the learner delegates ($p=1$), loss is $(1-\theta_a)^2$.
  • If the learner practices ($p=0$), loss is $(1-\theta)^2$.

Dynamics (ODE Formulation)

The authors derive a system of Ordinary Differential Equations (ODEs) representing the continuous-time limit of the discrete stochastic updates. The system describes the joint evolution of skill and delegation:

1. Skill Update ($\dot{\theta}$):
   • Improves when practicing (driven by error $(1-\theta)^2$).
   • Decays when delegating (driven by decay rate $\Delta$).
     $$\dot{\theta} = \theta(1-\theta) \left[ (1-p)(1-\theta) - \Delta p \theta \right]$$
2. Delegation Update ($\dot{p}$):
   • Increases if AI performance is better than the learner's current skill ($\theta < \theta_a$).
   • Decreases if the learner outperforms the AI.
     $$\dot{p} = \kappa p(1-p) \left[ (1-\theta)^2 - (1-\theta_a)^2 \right]$$

Key Assumption: The system is locally aligned; both updates minimize the same instantaneous loss. There is no incentive misalignment or "gaming" of the system.

3. Key Contributions & Theoretical Results

A. Emergence of Multiple Equilibria (Theorem 3.1)

Unlike the "no-AI" baseline (which has a single stable equilibrium at $\theta=1$), the AI-assisted system admits two stable equilibria:

1. High-Skill Equilibrium: $(\theta=1, p=0)$. The learner masters the task and stops delegating.
2. Low-Skill Equilibrium: $(\theta=0, p=1)$. The learner relies entirely on AI, leading to skill atrophy.
3. Saddle Point: An interior unstable equilibrium $(\theta^\dagger, p^\dagger)$ where the learner's skill equals the AI's skill ($\theta^\dagger = \theta_a$).

B. Basin of Attraction and Irreversibility (Theorem 3.2)

The state space $[0,1]^2$ is partitioned into two distinct basins of attraction separated by the stable manifold of the saddle point (a curve $p = \psi(\theta)$).

• Path Dependence: If the initial state $(\theta_0, p_0)$ lies above the curve ($p_0 > \psi(\theta_0)$), the system converges to the low-skill equilibrium. If below, it converges to the high-skill equilibrium.
• Irreversibility: Once a learner crosses the basin boundary into the low-skill region, the feedback loop (low skill $\to$ high delegation $\to$ lower skill) makes recovery impossible without external intervention, even if the learner is rational and performance-driven.

C. The "Cognitive Debt" Phenomenon (Theorem 3.4)

The paper proves that AI assistance can strictly improve performance in the short run but induce persistent long-term performance loss relative to a no-AI baseline.

• Mechanism: A negative feedback loop exists where higher initial delegation reduces practice, lowering skill, which in turn increases future delegation.
• Crossing Time ($t_c$): There exists a finite time $t_c$ after which the AI-assisted learner's performance becomes strictly worse than the learner who practiced independently, despite the AI being highly capable.

D. Impact of AI Quality (Theorem 3.3)

Counter-intuitively, improving AI quality ($\theta_a$) increases the risk of skill degradation.

• As $\theta_a$ increases, the basin boundary $\psi(\theta)$ shifts downward.
• This expands the low-skill basin, meaning a wider range of initial conditions (including those with moderate skill) will lead to total reliance and atrophy.
• Higher AI capability extends the duration of short-term gains ($t_c$), masking the long-term damage for longer.

4. Robustness and Extensions

The authors demonstrate that the qualitative phase structure (multiple equilibria, basin boundaries, irreversibility) is robust to:

• Jagged AI: AI with variable performance (high accuracy on some tasks, errors on others) is equivalent to a deterministic AI with a specific "effective skill."
• Noisy/Asymmetric Updates: Even if learners misestimate AI skill or react asymmetrically to errors, the bifurcation structure remains.
• Alternative Loss Functions: Including penalties for AI usage (e.g., detection) can shift the basin boundary, encouraging practice.

5. Significance and Implications

• Mechanism Identification: The paper identifies stability dynamics (specifically the geometry of the basin boundary) as the root cause of skill degradation, rather than incentive misalignment or irrational behavior. Even perfectly rational agents optimizing for immediate performance fall into the low-skill trap.
• Policy Recommendations:
  • Timing Matters: Introducing AI after partial skill acquisition is safer than introducing it at the start.
  • Incentive Redesign: Systems should reward independent problem-solving or delay high-quality AI assistance to prevent learners from crossing the basin boundary.
  • Monitoring: Early-stage learners are most vulnerable; monitoring delegation levels is crucial to prevent irreversible skill loss.
• Theoretical Contribution: It bridges the gap between learning theory (skill acquisition) and control theory (dynamical systems), providing a rigorous mathematical explanation for the "cognitive debt" observed in empirical studies of AI in education and coding.

Conclusion

The paper concludes that AI assistance creates a structural trap: the very mechanism that makes AI useful (immediate performance optimization) drives the system toward a low-skill equilibrium. The "geometry of learning" suggests that without deliberate intervention to maintain practice, the long-term trajectory of human skill under AI delegation is one of degradation, regardless of the AI's capability.