As the end of the semester approaches and grading lies ahead, I look back on the past term. I drastically changed the way I teach the intermediate macroeconomics course I have led for the past 12 years.
The trigger was a sense of crisis. The previous semester's final exam average was 20 points lower than in typical years. Students who had done their homework and projects flawlessly collapsed when it came to the exam. Their submissions looked impressive, but were not truly their own. As they leaned on artificial intelligence (AI), genuine learning quietly slipped away. This was not my problem alone. When I asked fellow professors, I heard that in some classes the final exam average had fallen by as much as 20 points.
So I returned to the chalkboard instead of the slides I had been using. I banned laptops and the use of AI, and had students take notes by hand. I drew a hint from existing research in cognitive and brain science showing that using one's hands leads to deeper learning. I raised the level of teaching considerably. Rather than problems that could be solved by memorization with AI's help, I filled the course with material that could only be solved through understanding. I even brought into the undergraduate lecture the endogenous growth theory of Nobel laureate economist Paul Romer, which is usually covered in graduate school. For homework and projects, on the other hand, I encouraged active use of AI.
The result was unexpected. The exam was no easy one, yet the average exceeded 90 points. Nearly everything I had wanted students to gain from this course—the acquisition of knowledge, the internalization of methodology, and the establishment of a new framework of thinking—was achieved. Even more surprising than the exam were the projects. The assignment was to build a new growth theory model themselves or to simulate the process of economic growth using AI. One team brought in an existing growth model that explains the birth of the Industrial Revolution. They replaced the model's "representative household" with 100 AI agents and, on that basis, let the economic history of humanity unfold from 4000 BC to 2100 AD. The closer it came to modern times, the more AI clung solely to innovation and hardly had children, so production exploded but population growth halted. The result closely resembled the low birthrate of our era. Another team reproduced, in a labor market of 3,000 people, the process by which discrimination pushes talent into the wrong positions and erodes growth. After entrusting hiring decisions to AI, they measured AI's bias and even the negative impact of that bias on economic growth. Although they received help from AI, the students designed models beyond the undergraduate level on the basis of proper understanding, and completed simulations they had never attempted before. They were each accurately answering the core question this course posed—"What grows an economy over the long term?"—on top of new cases of their own choosing.
The secret was simple. Within class, helping students properly understand the basics even without AI, and outside class, having them push that understanding further together with AI—this approach proved effective. My fellow professors, too, worry about the decline in students' learning ability in the AI era and are each seeking their own solutions. Some increase the weight of in-class presentations and quizzes, or introduce one-on-one oral assessments that demand much of the instructor's time and effort. All are efforts to capture in assessment "the moments AI cannot replace."
This shift in learning in the AI era brings to mind Romer's insight. He saw the engine of growth as lying in "ideas." Ideas do not wear out from being used heavily; they combine with one another to give birth to new ideas. And the human capital that shapes them determines the direction of growth. AI is the most powerful amplifier of ideas humanity has ever held in its hands. But an amplifier only magnifies what already exists; it cannot create what is not there. When the foundation is empty, AI replaces the student; when the foundation is solid, AI multiplies the student many times over.
Learning in the AI era is, in the end, a foundation built with one's own mind. The deeper and firmer the framework of understanding and thinking one builds for oneself, the farther AI will take us. The weight of assessment, too, will shift toward capacities that only people can demonstrate on the spot—memory, spontaneous thinking, and communication. But what is clear is that solidly laying the groundwork to build a framework of thinking has become more important than ever. When the foundation we have built ourselves is firm, AI will raise its ceiling even higher. So how do we build that foundation firmly? This is the task that the education field must now solve together.
