A review of

“Large Language Models in Teaching and Learning: Reflections on Implementing an AI Chatbot in Higher Education”

📄 Full paper: https://arxiv.org/pdf/2603.17773

The idea in one line

AI can make learning faster and more accessible.

It does not guarantee that people actually understand more.

This paper is valuable because it moves away from speculation. Instead of asking what AI could do in education, the authors implemented it inside a real course and observed what actually happens.

Who is behind this research

The study was conducted by

Fiammetta Caccavale , Carina L. Gargalo , Julian Kager , Magdalena Skowyra , Steen Larsen , Krist Gernaey , and Ulrich Krühne , all affiliated with the DTU – Technical University of Denmark (DTU), particularly within engineering and applied sciences.

What they actually did

They built a chatbot called ChatGMP, designed to simulate the role of a company in a learning exercise. Instead of students interacting with a human instructor acting as the company, they interacted with the AI.

The system used retrieval-augmented generation, meaning it could access course-specific material and ground its answers in the content students were supposed to learn.

To evaluate the impact, the researchers designed a controlled experiment where students experienced both types of interaction: human and AI. They collected data through surveys, performance analysis, and behavioral observation.

What they found

Students found the AI easier to interact with. It was always available, responded instantly, and removed the pressure that often comes with asking questions in front of a human. That alone increased engagement. Students were more willing to explore and ask questions.

But when the task required deeper reasoning, the picture changed. Students still leaned toward human interaction when they needed clarity, judgment, or confirmation that their understanding was correct.

The AI was helpful, but it was not fully trusted.

Another issue is accuracy. The paper highlights that AI can produce answers that sound correct but are not. In a learning environment, that creates risk because students may not have enough context to challenge those answers.

Over time, this leads to a subtle shift. Students rely more on the system, but may engage less with the underlying concepts.

The real insight

The most important takeaway is not about performance metrics. It is about behavior.

AI reduces friction. It makes it easier to move forward, easier to get answers, and easier to stay engaged. But learning is not just about moving faster. It requires effort, reflection, and sometimes confusion.

When AI removes too much friction, it can also reduce the depth of thinking. This creates a tradeoff. Efficiency goes up, but cognitive engagement can go down.

<>

The same dynamic is already visible inside companies. Employees using AI tools complete tasks faster and with less effort. But the risk is similar. If the system becomes the default source of answers, people may stop questioning or validating what they receive.

Over time, that affects decision quality.

The paper is essentially a small-scale version of what is happening at the organizational level.

What leaders should take from this

AI should not be evaluated only on speed or productivity. It should be evaluated on whether it improves the quality of thinking and decisions.

That requires designing systems where AI supports reasoning instead of replacing it. It also requires training people to question outputs, not just use them.

Without that layer, AI can create efficiency while quietly weakening understanding.

https://www.linkedin.com/embeds/publishingEmbed.html?articleId=7941650196589093350&li_theme=light

👉 Explore the Chief AI Officer (CAIO) Certification