🔬 physics

Students using GenAI lag behind in problem-solving competence: an agent-based study of classroom networks

Using an agent-based model to simulate high school physics classrooms, this study reveals that while Generative AI offers support, its use ultimately hinders the development of problem-solving competence and increases the proportion of students remaining in lower skill tiers by encouraging cognitive offloading and disrupting collective learning dynamics.

Original authors: Lorenzo Betti, Iacopo Caporossi, Carsten Källner, Karolina Levanaitė, Chenyu Li, Xuan-Chen Liu, Giulia Lorenzini, Vittoria Socci, Michele Re Fiorentin, Ilaria Stanzani, Marta Baratto

Published 2026-06-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lorenzo Betti, Iacopo Caporossi, Carsten Källner, Karolina Levanaitė, Chenyu Li, Xuan-Chen Liu, Giulia Lorenzini, Vittoria Socci, Michele Re Fiorentin, Ilaria Stanzani, Marta Baratto

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

The Big Picture: The "Cheat Code" Trap

Imagine a classroom of 30 high school students trying to learn how to solve complex physics puzzles. They have two main ways to tackle these puzzles:

The Hard Way: They think it through, maybe ask a friend for help, and figure it out themselves.
The "Magic Wand" Way: They use a super-smart AI (GenAI) that instantly gives them the answer.

This study asks a simple question: What happens to the whole class if everyone starts using the Magic Wand?

To find out, the researchers didn't just watch a real classroom (which is hard to control). Instead, they built a digital simulation—a video game of a classroom with 30 computer "agents" (students). They ran this game 1,000 times to see how the students' skills changed over time.

How the Simulation Worked

The researchers set up the game with three key rules:

The Network: Students are connected like a social network. Some are random friends, some are in tight-knit groups, and some are just loosely connected.
The Choice: At every step, a student decides: "Do I solve this myself (or with a friend), or do I ask the AI?"
The Consequence:
- If you solve it yourself, your brain gets a strong workout, and you get smarter.
- If you ask a friend, you learn a little bit from them.
- If you use the AI, you get the answer, but your brain doesn't get the full workout. You might learn a tiny bit of the facts, but you miss out on the "problem-solving muscle" training.

The Surprising Results

1. The "Average" Student Gets Worse

When the researchers turned on the AI for the whole class, the average problem-solving skill of the group went down.

The Analogy: Imagine a gym where everyone decides to stop lifting weights and instead just asks a machine to lift the weights for them. Even though the machine moves the weight, the people's muscles get weaker. The class as a whole became less capable of solving problems on their own.

2. The "Stuck in the Middle" Effect

Without AI, students naturally moved up the ladder. The ones who started with low skills caught up, and the whole class ended up in the "High" and "Very High" skill zones.

With AI: The students didn't climb as high. A significant chunk of the class got "stuck" in the middle or even the lower zones. They didn't fail, but they didn't grow as much as they could have.
The Analogy: Think of a race where everyone is running. Without AI, everyone runs at their own pace and gets faster over time. With AI, it's like some runners are on a moving walkway. They get to the finish line, but they never actually built the leg strength to run fast on their own.

3. The Class Split into Two Groups

This is the most important finding. When AI was introduced, the class didn't just get "a little worse." It split into two distinct groups:

Group A: Students who used AI less often. They kept getting smarter and ended up at the top.
Group B: Students who used AI a lot. They ended up in a "low skill" group, significantly behind the others.
The Analogy: Imagine a group of hikers. If everyone hikes the trail, they all get strong. But if some hikers take a helicopter (AI) to the top, they arrive at the destination, but they are weak and tired. The group splits: the strong hikers and the weak helicopter-riders. The helicopter riders didn't just stay the same; they actually fell behind the hikers in terms of fitness.

4. The "Shortcut" Habit

The study found that the more a student relied on the AI, the more likely they were to end up in that lower-skilled group.

The Analogy: It's like using a calculator for simple math. If you use it for everything, you eventually forget how to do basic addition. The students who used the AI as a "shortcut" stopped building their own problem-solving skills, and the AI actually became a barrier to their learning.

The Bottom Line

The paper concludes that while AI can give students the answer, it can stop them from building the skill to find the answer themselves.

When a whole class uses AI, it doesn't just help everyone get smarter; it actually creates a divide. It leaves a large portion of the class lagging behind in their ability to think critically and solve problems, while only a small group (those who didn't rely on the AI) continues to grow. The researchers suggest we need to look at how AI affects the whole group dynamic, not just how well one individual student does on a single test.

Technical Summary: Students Using GenAI Lag Behind in Problem-Solving Competence

Problem Statement
The development of Problem-Solving Competence (PSC) is foundational for preparing resilient citizens, yet the collective impact of Generative Artificial Intelligence (GenAI) on this development within classroom environments remains underexplored. While existing literature extensively covers individual learning outcomes and attitudes toward GenAI, there is a lack of understanding regarding how the distribution of learning strategies—specifically the choice between using GenAI, collaborating with peers, or working individually—shapes the evolution of PSC across a learning community. A critical gap exists in determining whether GenAI acts as a scaffold for co-constructive learning or a mechanism for cognitive offloading that hinders deep learning, particularly when considering the dynamic interplay between individual choices and peer network structures.

Methodology
To address this, the authors developed an Agent-Based Model (ABM) to simulate the evolution of PSC in a high school physics classroom over 60 time steps (representing 60 exercises).

Agents and State Variables: The model consists of $N=30$ $N = 30$ student agents. Each agent $i$ $i$ is characterized by three state variables:
1. Problem-Solving Competence ( $C_i^t$ ): The ability to solve physics exercises, initialized using empirical distributions from Italian INVALSI assessment data ( $N(\mu=195.9, \sigma=1.28)$ ).
2. GenAI Dependence ( $\alpha_i^t$ ): A value in $[0,1]$ representing reliance on AI, initialized at 0.5.
3. Reliance on Peers ( $\tau_i^t$ ): A value in $[0,1]$ representing the tendency to seek peer help, initialized at 0.5.
Data and Calibration: The model utilizes a dataset of 60 physics exercises covering the Italian Scientific Lyceum curriculum. Each exercise is annotated with three components: mathematical knowledge, physical reasoning, and cognitive problem-solving skills (CPSS). Crucially, the model quantifies the "score gap" between human solutions (without AI) and GenAI solutions (using OpenAI's GPT-5.5). The data indicates that while GenAI can solve exercises, it yields lower gains in CPSS compared to human problem-solving.
Dynamics and Strategy Selection: At each time step, agents choose a strategy based on a sigmoid probability function influenced by exercise difficulty, current competence gain, AI dependence, and peer reliance.
- GenAI Use ( $S_i=1$ ): Agents assume the exercise is solved, but competence updates are restricted. They gain potential mathematical and physical knowledge but experience a potential loss or zero gain in CPSS, reflecting cognitive offloading.
- No GenAI ( $S_i=0$ ): Agents either work independently or collaborate with a peer from their network. Success probability is set to 0.54 (based on national data). Successful independent work yields full competence gains. Peer collaboration allows for knowledge transfer: if a peer is more competent, the agent's competence is updated to a random value between their own and the peer's; if less competent, no additional learning occurs.
Network Structures: The study simulates interactions across four distinct network topologies to test structural robustness:
1. Erdős–Rényi (ER) random networks.
2. Watts–Strogatz (WS) small-world networks.
3. Stochastic Block Models (SBM) representing community structures.
4. Modified SBMs incorporating competence-based homophily (where initial peer connections are formed based on similar competence levels).

Key Results
The simulation results, derived from 1,000 independent runs per condition, reveal significant negative impacts of GenAI access on collective competence dynamics:

Reduced Average Competence: Across all network topologies, classrooms with access to GenAI consistently demonstrated a slower rate of skill acquisition and a lower final average PSC compared to classrooms without GenAI.
Stalled Upward Mobility: In the absence of GenAI, students progressively moved toward higher competence tiers, with the distribution concentrating in the "High" and "Very High" categories. Conversely, GenAI access caused a stagnation where a significant portion of students remained in "Mid," "Low," and "Very Low" tiers.
Emergence of Bimodality: The introduction of GenAI transformed the final competence distribution from a relatively homogeneous unimodal state into a bimodal distribution. Gaussian mixture modeling showed that GenAI access increased the size of a low-competence cluster from approximately 1–4% (in no-AI scenarios) to 8–12% of the population.
Stratification by Usage Frequency: Analysis of usage frequency revealed that students with the highest frequency of GenAI usage (75–100% of exercises) were disproportionately concentrated in the emerging low-competence peak. This suggests that heavy reliance on GenAI actively suppresses the mastery of problem-solving skills.
Interaction with Homophily: In networks with competence-based homophily (modified SBM), GenAI access appeared to amplify stratification, reducing the likelihood of students moving beyond their initial competence neighborhoods.

Significance and Claims
The paper claims that the educational impact of GenAI cannot be assessed solely through individual learning outcomes or immediate task completion. Instead, the study demonstrates that GenAI availability fundamentally reshapes the collective dynamics of competence development within learning communities.

The authors argue that while GenAI may facilitate the completion of tasks, its use as a shortcut can lead to cognitive offloading that diminishes the development of autonomous problem-solving skills. This effect is not uniform; it creates a stratification where a subset of students lags behind, resulting in a polarized classroom where a significant minority fails to achieve high-level competence. The study concludes that the risks of GenAI extend beyond individual critical thinking deficits to include the structural reorganization of classroom competence distributions, potentially leaving a larger share of students in lower competence tiers than would occur in traditional learning environments.