📈 economics

Curriculum design in the age of AI

This paper models how AI availability alters optimal curriculum design, showing that while AI can accelerate learning for high-skill students when it complements effort, it generally forces teachers to distort task sequences to incentivize student effort, ultimately resulting in reduced overall skill development.

Original authors: Benjamin Davies

Published 2026-07-22

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Benjamin Davies

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to get better at something, like playing a video game or solving a tricky puzzle. You know there are two ways to tackle a level: you can grind through it yourself, struggling with every obstacle, or you can use a "cheat code" that instantly solves the puzzle for you. In the world of economics and education, this is a classic trade-off. The "grind" is hard work, but it builds your brain muscles (skills) for the future. The "cheat code" is easy and gives you a perfect score right now, but it leaves your brain muscles weak. This paper, written by an economist named Benjamin Davies, asks a burning question for our time: If students have access to a super-smart cheat code (Artificial Intelligence), how should teachers design their homework to make sure students still get strong?

The paper uses a model where a student faces a sequence of tasks. They can either do the work (which costs effort but builds skill) or hand it over to an AI (which costs no effort but builds no skill). The teacher's job is to arrange the tasks in a specific order—a "curriculum"—to get the student to do as much work as possible. The paper explores what happens when the teacher tries to outsmart the student's temptation to use the AI. It turns out that the presence of AI forces teachers to change the game entirely, sometimes making early homework more focused on specific types of thinking to keep students engaged, and sometimes making AI actually help high-skill students learn faster while hurting low-skill ones.

The Cheat Code Dilemma

Think of a student's education as a long video game campaign. The "skills" are your character's strength and intelligence, which grow only when you fight the monsters yourself. The "effort" is the energy you spend swinging your sword. In this story, the AI is a magical wand that can defeat any monster instantly. If you use the wand, you get the treasure (the grade) without swinging a sword, but your character doesn't get stronger.

The teacher is the game designer. Their goal isn't just to give you a high score; it's to make sure your character becomes a legendary hero by the end of the game. The paper asks: How should the game designer arrange the levels so you choose to swing your sword instead of using the wand?

The "No-Cheat" World: A Smooth Climb

First, the paper looks at a world where the cheat code doesn't exist (or the student refuses to use it). In this "First-Best" scenario, the teacher designs a perfect path. The early levels are simple but require you to swing your sword a lot (high "effort intensity"). As you get stronger and your character levels up, the later levels become trickier. They require you to figure out how to swing the sword, not just swing it blindly (high "skill intensity").

The logic is simple: When you are weak, you need to practice the basics. When you are strong, you need to learn strategy. The paper proves that without AI, the best curriculum starts with heavy lifting and gradually shifts to brainy puzzles. This makes sense because your strength and your strategy work better together; the stronger you get, the more effective your strategy becomes.

The "Cheat Code" World: The Twist

Now, introduce the AI. Suddenly, the student has a choice: "Do I swing my sword, or do I just wave the wand?" If the wand gives a perfect score with zero effort, the student will wave it on every easy level. If the teacher keeps the "First-Best" plan (easy levels first), the student will just use the AI, learn nothing, and the teacher's plan fails.

To fix this, the teacher has to "distort" the curriculum. This is the paper's big discovery. To stop the student from waving the wand on the early levels, the teacher must make those early levels require more "brain power" (skill) in a specific way. They have to make the early tasks require the student to identify the right methods to solve the problem, rather than just applying a given method. This forces the student to understand the rules of the game to beat the level, which makes using the AI less effective because the AI cannot simply execute a pre-set routine; the student must first figure out what to do.

The paper shows that this creates a weird, U-shaped curve for the homework.

Early Levels: The teacher makes them require more strategic thinking (high skill intensity) to force the student to identify the right approach.
Middle Levels: As the student gets smarter, the teacher eases up on the complexity and lets the student do more "grinding" (effort).
Late Levels: Once the student is a pro, the teacher returns to the original plan: complex strategy puzzles that the student can solve on their own.

The result? The student still learns, but they learn less than they would have if the AI didn't exist. The teacher had to twist the game so much to keep the student honest that the overall learning curve is flattened.

The "Magic Wand" Gets Better: High Skills vs. Low Skills

The paper takes it a step further. What if the AI gets even smarter? Specifically, what if the AI can be used to speed up the actual work (effort) once the student has figured out the strategy? The paper finds a surprising split in how this affects students.

The High-Skill Student: If a student is already pretty good at the game, a smarter AI actually helps them learn faster. Why? Because the AI can accelerate the effortful parts of the work (like proving a series of lemmas in a math problem) once the student has used their skill to identify the strategy. The AI acts to boost the productivity of their effort, allowing them to gain skill more rapidly.
The Low-Skill Student: If a student is struggling, a smarter AI makes them learn slower. The "super-wand" is just too tempting. They use it to skip the hard work entirely, and because they skip the work, they never build the skills they need to get better.

This suggests that as AI gets better, the gap between the "smart" students and the "struggling" students might get wider. The smart ones use the tool to fly; the struggling ones use the tool to float, never learning to fly themselves.

The Bottom Line

The paper doesn't say AI is bad or good; it says AI changes the rules of the game. If teachers keep designing homework the same way they did before, students will just use the AI to cheat, and learning will stop. To fight back, teachers might need to make early homework more about "showing your work" and understanding the "why," rather than just getting the right answer.

The model suggests that if you want students to learn, you can't just give them a list of tasks. You have to carefully design the difficulty so that using the "cheat code" feels like a bad deal. But even with the best design, the paper suggests that the presence of AI makes it harder for everyone to reach their full potential, and it might make the difference between a top student and a struggling student even more dramatic. It's a reminder that in the age of AI, the way we teach might need to change just as much as the way we learn.

Technical Summary: Curriculum Design in the Age of AI

Problem Statement
The paper addresses the challenge of designing educational curricula in an environment where students have access to Artificial Intelligence (AI) capable of delegating coursework. While AI can boost productivity, empirical evidence suggests it may inhibit learning by allowing students to "short-circuit the productive struggle" essential for skill acquisition. The core problem is a principal-agent dynamic: a forward-looking teacher (principal) must design a sequence of tasks (curriculum) to maximize a myopic student's (agent) skill development, knowing the student can choose between working (incurring effort costs but building skill) or delegating to AI (incurring no effort but building no skill). The teacher must balance task difficulty to incentivize effort while accounting for the student's temptation to delegate.

Methodology
The author develops a dynamic model of "learning-by-doing" with endogenous curriculum design.

Agents: A myopic student and a teacher.
Tasks: A sequence $t \in \{0, \dots, T-1\}$ .
Production: Working on a task combines effort ( $x$ ) and skill ( $s$ ) to produce output. Effort is costly; skill is costless. They are complementary inputs. Tasks are characterized by effort intensity ( $\eta$ ), where low $\eta$ implies high reliance on skill (identifying methods) and high $\eta$ implies high reliance on effort (applying given methods).
Delegation: The student can delegate a task to AI, receiving a fixed payoff $u$ with zero effort and zero skill gain.
Skill Accumulation: Skill increases proportionally to effort exerted ( $\Delta s = r x$ ) if the student works; otherwise, skill remains static.
Objective: The teacher chooses the sequence of effort intensities $(\eta_t)$ to maximize the student's total skill gain ( $s_T - s_0$ ), subject to the student's incentive compatibility (IC) constraint: the student works only if the payoff from working exceeds the delegation payoff.

The analysis proceeds by characterizing the student's best-response effort and payoff, then solving for the teacher's optimal curriculum under two scenarios: the "first-best" (no delegation threat) and the "second-best" (delegation threat). Finally, the model is extended to include an "AI quality" parameter ( $q$ ) that can complement effort (increasing the marginal product of effort) while also increasing the delegation payoff.

Key Contributions and Results

1. First-Best Curriculum (No Delegation Threat)
When the delegation payoff is zero ( $u=0$ ), the student always works. The teacher optimizes solely to maximize effort.

Result (Theorem 1): The optimal curriculum features tasks that are gradually less effort-intensive and more skill-intensive over time.
Mechanism: Early tasks are effort-intensive to incentivize effort when the student has low skill. As the student accumulates skill, the teacher shifts to skill-intensive tasks. Because effort and skill are complements, higher skill makes effort more productive; thus, the teacher reduces the effort intensity to capitalize on this complementarity, extracting maximal effort from the now-more-capable student.

2. Second-Best Curriculum (Delegation Threat)
When $u > 0$ , the student may delegate if the payoff from working is insufficient.

Result (Theorem 2): The teacher must distort the curriculum to satisfy incentive constraints. The optimal sequence exhibits increasing-then-decreasing effort intensities.
Mechanism: On early tasks, the student has low skill and low payoff from working, making delegation attractive. To prevent this, the teacher must make early tasks more skill-intensive than the first-best prescription (distorting $\eta$ downward) to raise the student's payoff from working. As the student gains skill, the delegation constraint relaxes, and the curriculum transitions to the first-best path of decreasing effort intensity.
Implication: The presence of AI leads to less overall skill development compared to the first-best case because the teacher must distort task design to keep the student engaged.

3. Impact of Initial Skill and Parameters

Result (Theorem 3): Total skill gain is increasing in initial skill ( $s_0$ ), productivity ( $\pi$ ), and the skill accumulation rate ( $r$ ), but decreasing in the delegation payoff ( $u$ ).
Mechanism: This confirms the "skill begets skill" dynamic. Higher initial skill reduces the temptation to delegate early on, allowing for a curriculum closer to the first-best, which further accelerates skill accumulation.

4. AI as a Complement to Effort
The model is extended to allow AI quality ( $q$ ) to increase the marginal product of effort (complementarity) while also increasing the delegation payoff.

Result (Theorem 4): The impact of improved AI quality is non-monotone and depends on the student's skill level.
- High-Skill Students: As $q$ increases, these students gain skill faster. Their working payoff rises faster than their delegation payoff, making them more willing to work. The teacher responds by increasing effort intensity to capitalize on their productivity.
- Low-Skill Students: As $q$ increases, these students gain skill slower. The increase in the delegation payoff dominates the increase in the marginal product of effort, making delegation more attractive. The teacher must distort the curriculum further (reducing effort intensity) to incentivize work, resulting in lower effort and slower skill accumulation.
Result (Section 5.3): The overall effect of AI quality on total skill gain depends on the length of the curriculum ( $T$ ) and the accumulation rate ( $r$ ). In short curricula, the negative effect on early (low-skill) tasks may dominate. In long curricula, the positive effect on later (high-skill) tasks may dominate, provided the student can accumulate enough skill to overcome the initial temptation to delegate.

Significance and Claims
The paper claims to provide a theoretical framework for understanding how educational design must adapt to AI. It argues that:

Curriculum Distortion is Necessary: To prevent AI-induced skill atrophy, teachers must intentionally design tasks that are more skill-intensive (requiring method identification) than they would be in an AI-free environment. This aligns with empirical suggestions to require students to "document their process" and "demonstrate understanding."
Divergent Impacts: Improvements in AI quality do not uniformly benefit or harm learning. They create a divergence where high-skill students accelerate their learning, while low-skill students may fall further behind due to increased temptation to delegate.
Dynamic Complementarities: The model reinforces the importance of early investment in skill. Because skill begets skill, preventing early delegation is critical for long-term development.
Broader Applicability: While focused on education, the model's logic regarding delegation and incentive design applies to other contexts, such as industrial offshoring (where governments may wish to incentivize domestic production to build human capital) or academic integrity (copying from peers).

The paper concludes modestly, noting that its assumptions (e.g., myopic students, single-student optimization) could be relaxed in future research to address more complex real-world scenarios, such as designing curricula for heterogeneous student populations.