Move along with "AI" rather than rotting in the same place, blaming. (UOG0623022)

Move along with "AI" rather than rotting in the same place, blaming.

The article "Comparing Code Explanations Created by Students and Large Language Models" by Juho Leinonen, Paul Denny , Stephen MacNeil , Sami Sarsa, Seth Bernstein , Joanne Kim, Andrew Tran, and Arto Hellas (2023) was presented at the ITiCSE 2023 Conference in Turku, Finland. It looks at how code explanations written by students compare with those made by a large language model (LLM)(GPT-3) . The main aim of the study was to find out if AI-generated explanations could be useful in teaching programming, helping students understand code better and improve their own explanation skills.

Main points which guided this research were,

• To what extent do code explanations generated by LLMs differ in accuracy, length and understandability?
• What aspects of code explanations do students value ?

The results show that the code explanations generated by LLMs and by students are equivalent in terms of ideal length, but the code explanations generated by LLMs are perceived as more accurate and easier to understand. Around 1,000 first-year programming students from the University of Auckland took part in the research. In one lab session, students wrote explanations for three C functions, and in another, they rated both student and LLM explanations. Then the results were analyzed to compare the responses.

The findings showed that LLM’s explanations were rated as more accurate and easier to understand. But their length was about the same as the ones written by students. From students’ feedback, the researchers found that, step-by-step explanations that describe both what the code does and how it works were most preferred. The study concluded that AI-generated explanations could be a useful resource in large programming classes.


I find many strengths in this research article, they have got 1000 students to participate in this experiment which makes the their experiment's sampling more reasonable and it leads to a more accurate and a reliable result. And this research is very timely, as we are in an era of artificial intelligence dominating the programming and information technology sectors, this research helps people to get some insights and realize that its better to move on with the technological advancements while embracing and adopting to them rather than blaming them and stay rotting at the same place. Another good point is that the study took place in real classrooms, so the findings reflect actual learning situations.

However there are also some slight downsides of this research. They have used 1st year programming undergraduates and only used the C programming language . Because of that I think the scope of this research might got a bit narrowed. which means the results might not apply to advanced learners or complex coding tasks. The study also depended on student opinions rather than measuring how much they actually learned or improved. Another point is that GPT-3 model is an older version. Newer models like GPT-4 or GPT-5 might give different results. There’s also a small risk that students could rely too much on AI explanations and not develop their own problem-solving skills.

If we were to do this type of research I would prefer that it would be good to test with more programming languages and more. Because it will result in expanding our research scope for more accurate and reliable sets of data.

I found this article very interesting because it shows how AI can help students learn programming more easily. It explains that LLMs can make coding concepts clearer, especially for beginners. The research felt believable since it used real classroom data and student opinions. I also liked how it reminded readers that while AI can be a great learning tool, students still need to think for themselves and practice explaining code in their own words.

In conclusion, the study shows strong evidence that large language models like GPT-3 can produce accurate and easy-to-understand code explanations for beginner programmers. The large number of participants and the classroom based environment made the findings more meaningful. Still, future studies should include more complex coding tasks, different student groups, and ways to measure actual learning progress. Overall, this research clearly shows that AI-generated explanations can be helpful in modern teaching and learning in programming.

Reference: "Comparing Code Explanations Created by Students and Large Language Models" by Juho Leinonen, Paul Denny , Stephen MacNeil , Sami Sarsa, Seth Bernstein , Joanne Kim, Andrew Tran, and Arto Hellas (2023)

Index : UOG0623022