Learning Outcomes, Human–AI Trust, and Behavioral Patterns in GenAI-Assisted Pair Programming
This quasi-experimental study demonstrates that integrating generative AI into middle school pair programming significantly enhances students' programming knowledge and project performance while increasing their trust in AI, though it reveals a behavioral tendency to rely heavily on debugging assistance rather than proactive problem decomposition.
Original paper licensed under CC BY 4.0 (https://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
In classrooms around the world, a simple idea has long been used to teach the complex art of writing computer code: two students sit at one computer and work together. One person, called the driver, controls the keyboard and mouse to type the instructions, while the other, the navigator, watches the screen, spots mistakes, and suggests the next move. This method, known as pair programming, helps beginners learn faster and feel less alone when things go wrong. But as powerful new computer programs that can generate text and code appear in schools, a new question arises. What happens when you add a third partner to this team? Can an artificial intelligence act as a helpful guide without taking over the learning process, or does it risk making students too dependent on a machine that might not always be right?
A team of researchers in China recently set out to answer these questions by watching how middle school students learned to program with and without this new technology. They worked with sixty-nine eighth-grade students who had very little experience with coding. The researchers split the class into two groups. One group worked in the traditional way, with two students sharing a computer and no help from artificial intelligence. The other group worked in the same way, but with access to a smart computer program that could answer questions, explain errors, and write code snippets when asked. The students in both groups spent eight weeks building a small software project, switching roles every ten minutes so that everyone got a chance to drive the computer and navigate the conversation.
The results showed that the students who worked with the artificial intelligence support actually learned more. When tested at the end of the eight weeks, they knew more about programming concepts and built better final projects than the students who worked without the tool. However, the study also revealed a specific pattern in how these students used the technology. When the code failed to work, the students often asked the computer for an explanation of the error, and then immediately asked it to fix the problem. This created a cycle where they would get stuck, ask for help, get a solution, and then get stuck again. They asked fewer questions about how to break a big problem into smaller pieces or how the code actually worked, suggesting they were using the tool to jump straight to answers rather than to understand the logic behind them.
Perhaps most interesting was how the students' feelings toward the technology changed over time. At the start of the study, the students were unsure about trusting the computer. By the end, after weeks of working alongside it, their trust in the system had grown significantly. They had learned to rely on it, but the researchers noted that this trust was built through repeated interaction and verification. The study did not find that the artificial intelligence replaced the need for human partners; instead, it acted as a third resource that the two students had to manage together. The driver would type the prompts, while the navigator watched the screen and helped decide if the computer's suggestions made sense.
This research highlights a delicate balance in modern education. The artificial intelligence tool clearly helped the students produce better work and learn more facts, but it also encouraged a habit of asking for immediate fixes rather than deep thinking. The students did not simply accept everything the computer said; they worked together to interpret its answers. The study suggests that while these tools can boost learning outcomes, educators must be careful to guide students on how to use them. The goal is not just to let the computer solve the problem, but to ensure that the students remain the ones thinking through the solution, using the machine as a partner rather than a shortcut.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.