Supervised Practice Time and the AI-Code Workflow: Evidence from Finance Undergraduates at a Regional Chinese University

In the modern classroom of finance, a quiet revolution has taken place. For decades, learning to analyze markets meant mastering the syntax of computer code, a barrier that kept many students from reaching the deeper logic of economics. Today, a new kind of artificial intelligence can write that code instantly, turning plain English sentences into working programs. This shift has moved the challenge from writing the instructions to managing the process: asking the right questions, running the script, checking the results, and fixing the errors. The central question for educators is no longer whether students can type commands, but whether they can effectively direct this new digital assistant. If the machine can do the heavy lifting of writing, does the student still need to practice the craft, and if so, how?

A team of researchers set out to answer this by observing students at a regional university in China. They focused on a simple variable: supervised practice time. This is the time a student spends working on a computer with an instructor present in the room, rather than just watching a demonstration or working alone at home. The researchers wondered if having that extra time, guided by a teacher, would help students get better at commanding the AI. They recruited four intact classes of finance undergraduates, totaling 138 students, and split them into two groups. One group was allowed to bring their own laptops to every class session, allowing them to practice the cycle of prompting the AI, running the code, and checking the output while the teacher watched. The other group followed a standard schedule, where they only had access to computers during specific, scheduled lab sessions, spending the rest of the class time watching the instructor demonstrate the work on a screen.

The study unfolded over two semesters. In the first term, the students worked with a financial modeling program called MATLAB. They faced four hands-on tests where they had to solve real-world problems, such as calculating the price of a bond or analyzing stock market trends. The rules were strict: the code had to run without errors and produce a specific numerical answer. The results showed a clear divergence. At the very beginning, both groups performed similarly, with scores hovering near the bottom of the scale. However, as the weeks passed and the tasks became more complex, the group with more supervised practice time began to pull ahead. By the final test, the students who had practiced with their laptops in class scored roughly eleven points higher than those who had only watched. This gap was not just a statistical blip; it represented a shift in the entire class distribution. The students with more practice were far less likely to get stuck on the first step, and more likely to complete the full sequence of tasks.

The researchers then followed two of the classes into a second semester, where they tackled a different software program called Stata for corporate finance problems. One of these classes received an additional round of the "bring your own device" arrangement, while the other stuck to the standard schedule. Even though this was a new software tool and a new set of problems, the pattern repeated itself. The class with the extra supervised hours performed significantly better, particularly on the harder tasks that required chaining multiple steps together. The students with more practice were nearly twice as likely to complete the most difficult tasks compared to their peers who had less time to rehearse the workflow under supervision.

To understand why this happened, the researchers asked the students directly. The answers pointed to a human trait rather than a technical one. When asked what stopped them from learning the AI workflow, the vast majority of students, in both groups, cited a lack of self-motivation. They admitted that without the structure of a class and the presence of a teacher, they simply would not have practiced the cycle of prompting and checking on their own. The supervised hours acted as a substitute for that missing drive, forcing the practice to happen when it otherwise would not have. The study suggests that the value of these extra hours was not in the technology itself, but in the environment that made the students use it.

However, the study also revealed a new bottleneck that emerged once the students got the code running. The biggest difficulty they reported was no longer writing the prompt or fixing a syntax error; it was judging whether the answer the AI gave was actually correct. Once the code produced a result, the students struggled to verify if that result made sense in the real world. This is a critical hurdle, as artificial intelligence can produce fluent, convincing answers that are subtly wrong. The researchers noted that while the extra practice time helped students get the code to run, it did not fully solve the problem of knowing if the output was right. This verification step, they found, requires a depth of domain knowledge that most students did not possess and were unlikely to develop on their own after class.

The findings offer a clear, if nuanced, picture of how education adapts to artificial intelligence. The study did not find that simply giving students access to computers or AI tools was enough to improve their skills. Instead, it showed that the structure of the learning environment matters deeply. When students are given the chance to practice the full cycle of directing an AI under the watchful eye of an instructor, they learn to navigate the workflow more effectively. The advantage grows as tasks become harder, suggesting that the ability to manage these tools is a skill built through repetition and guidance. Yet, the research also leaves a question open: how do we teach students to spot the mistakes that even a perfect-looking code might hide? The study concludes that while supervised practice solves the problem of getting the machine to work, the harder challenge of knowing when the machine is right remains a task for future exploration.