Delegating tasks to AI undermines learning-by-doing: teachers reshape curricula to elicit effort, lowering aggregate skill formation; better AI benefits high-ability learners but can slow learning among low-ability students when AI complements effort.

Main Finding

A teacher designing a curriculum must distort task difficulty when students can delegate work to AI. In a model where effort builds student skill but delegation to AI yields output without learning, the first-best curriculum (no delegation threat) sequences tasks from effort-intensive early tasks to skill-intensive later tasks. When delegation is attractive, the teacher makes earlier tasks more skill-intensive to deter delegation, producing an increasing-then-decreasing profile of effort intensity and ultimately less skill accumulation. If AI improvements raise both the productivity of effort and the reward to delegation (AI complements effort), better AI helps high-skill students learn faster but makes low-skill students learn slower.

Key Points

• Setup

  • Two agents: a myopic student and a forward-looking teacher who chooses a sequence of tasks (a curriculum).
  • Each task t has an "effort intensity" ηt ∈ (0,1]: higher η emphasizes effort over pure skill.
  • If the student works, he chooses effort xt ≥ 0, yielding output from a combination of effort and skill; effort is costly but builds skill. If he delegates, he exerts zero effort and obtains a delegation payoff u (no skill gain).
  • Skill evolves by learning-by-doing: st+1 = st + r · x∗(st, ηt) when the student works; otherwise skill does not increase.
  • Teacher chooses (ηt) to maximize total skill gain sT − s0 given the student's best responses.

• Student best responses (closed form)

  • Best-response effort: x*(s, η) = s · (π η / s)^{1/(2−η)} (π indexes productivity).
  • Payoff from working: u*(s, η) (given in the paper) increases in π and, for η < 1, in s.
  • When η = 1 (fully effort-intensive task), effort and payoff do not depend on skill.

• First-best (no delegation threat: u = 0)

  • Teacher maximizes effort on each task; there is a unique mapping ηFB(s) that maximizes x*(s,η).
  • ηFB(s) = 1 for low s (specifically for s ≤ πe), and declines with s for higher s. Thus optimal curriculum: earlier tasks more effort-intensive; later tasks more skill-intensive (Theorem 1).
  • Intuition: with skill-effort complementarity, later tasks exploit higher skill by emphasizing skill content; early tasks induce effort when skill is low by making effort payoff larger.

• Second-best (delegation payoff u > 0)

  • Incentive constraint: student works only if u*(st, ηt) ≥ u.
  • To prevent delegation, the teacher must make some early tasks more skill-intensive than in the first-best (i.e., lower η) so that good performance requires understanding rather than just delegating answers (Theorem 2).
  • Resulting curriculum: effort intensity is increasing initially (to build deterrent skill) then decreasing later — an increasing-then-decreasing profile.
  • Distortion lowers aggregate skill gain relative to the first-best. Higher initial skill or lower delegation payoff relaxes the incentive constraint and raises total skill gain (Theorem 3).

• AI quality extension

  • Introduce an AI-quality parameter that both raises the marginal product of effort (when used as a complement) and increases the delegation payoff u.
  • Comparative static (Theorem 4): improvements in AI quality can have heterogeneous effects — high-skill students become more willing to work and thus learn faster, whereas low-skill students are more tempted to delegate and learn more slowly. The optimal curriculum and outcomes are non-monotone in AI quality.

• Relation to evidence/literature

  • The model formalizes concerns from empirical work showing widespread AI use and associated skill declines when students delegate (Bastani et al., Chirikov, Kosmyna et al., etc.).
  • Connects to learning-by-doing, principal–agent incentive design in education, and recent theory on AI and skill formation.

Data & Methods

• Methodology: theoretical/analytical model. No empirical data used.
• Core techniques:
  • Continuous optimization for the student’s stage-by-stage best response (first-order conditions yield closed-form x*(s, η)).
  • Dynamic principal (teacher) problem with incentive constraints: teacher chooses nonlinear sequence (ηt) anticipating myopic student choices.
  • Characterization via lemmas and theorems: closed-form best-response, characterization of ηFB(s), and comparative statics under parameter changes (π, r, u, AI quality).
• Key assumptions and primitives:
  • Student is myopic (maximizes per-task payoff) but skill persists across tasks.
  • Effort cost is quadratic; output from working is a Cobb–type combination of effort and skill governed by η.
  • Delegation yields a fixed payoff u independent of skill and η.
  • Teacher cares only about maximizing student skill accumulation (value-added), not other outcomes (e.g., labor-market payoffs).
  • Model abstracts from stochastic AI errors, peer interactions, multi-agent classrooms, and external labor-market signals.
• Existence/uniqueness: the paper shows existence of solutions; uniqueness under regularity conditions discussed (some conditions referenced but not fully excerpted).

Implications for AI Economics

• Curriculum design must be endogenous to AI availability:
  • When delegation is available, simply assigning more or harder tasks can backfire. Teachers should design earlier assignments to require domain-specific understanding (skill intensity) so that delegation yields lower output relative to genuine student work.
  • Assessment reforms that require process evidence (worked steps, explanations) align with the model’s second-best prescription.
• Distributional effects and inequality:
  • AI quality improvements can widen skill gaps: high-skill learners benefit (complementarity boosts their productive effort), low-skill learners fall behind (stronger temptation to delegate).
  • Policy interventions (targeted supervision, scaffolded tasks, monitoring, process-based grading) may be needed to prevent AI from amplifying inequality in skill formation.
• Tradeoffs for institutions and firms:
  • Analogous tradeoffs likely arise in workplace training: letting employees use AI for tasks may increase short-run productivity but reduce learning; optimal task allocation or incentives may require deliberate distortion of work content.
• Research and policy recommendations
  • Empirical tests: measure task “effort intensity,” track AI usage, and estimate heterogeneous skill dynamics by baseline skill level; randomized assignments of task types (effort- vs. skill-intensive) while monitoring AI use would test the model’s predictions.
  • Design experiments that vary AI quality or access to study non-monotone effects on different skill groups.
  • Consider policies that make delegation less attractive (process-based assessments, graded oral exams, requiring partial solutions) and evaluate tradeoffs between productivity gains and human capital formation.
• Limitations and extensions
  • The model abstracts from multi-student interactions, stochastic AI errors, and incentives beyond skill accumulation (grades, labor-market signaling). Relaxing these could change optimal curriculum distortions.
  • Future work could model teachers who also care about grades or external signals, heterogeneous classrooms, or dynamic student preferences for long-run outcomes.

Summary takeaway: AI creates an incentive problem in education. Optimal curricula must anticipate delegation by shifting early tasks toward skills that cannot be cheaply delegated, but this corrective distortion reduces overall learning compared with a no-AI world and can affect low- and high-skill students differently as AI improves.