Beyond verbal self‐explanations: Student annotations of a code‐tracing solution produced by <scp>ChatGPT</scp>

Abstract ChatGPT is a generative Artificial Intelligence (AI) that can produce a variety of outputs, including solutions to problems. Prior research shows that for students to learn from instructional content, they need to actively process the content. To date, existing research has focused on student explanations expressed in words (either spoken or written). Thus, less is known about other forms of expression, such as ones involving spatial elements (eg, flowcharts, drawings). Moreover, to the best of our knowledge, there is not yet work on how students annotate solutions produced by ChatGPT. The study was conducted in the context of a first‐year programming tutorial focused on loops and code tracing. Code tracing is a fundamental programming skill that involves simulating at a high level the actions a computer takes when it executes a program. The students annotated a printed‐out code trace produced by ChatGPT using the strategy of their choice. Our goal was to describe the visual and verbal strategies students used to annotate the ChatGPT trace as well as how strategies relate to annotation quality, and so we used an observational study design with a single condition. Annotation strategies ranged from words‐only strategies to visual representations like flowcharts. As annotation quality increased, the proportions of strategies used changed, suggesting that some strategies may facilitate the production of quality annotations. In particular, the proportion of words‐only and flowchart strategies increased as quality increased; in the top quality quartile, there was a similar proportion of each but with slightly more flowcharts. Practitioner notes What is already known about this topic When students study instructional materials, they need to actively and constructively interact with the materials in order to learn effectively. Much of the research showing this has examined only verbal student output. In addition to verbal strategies involving only words, strategies including visual elements are also beneficial. For instance, when students are asked to predict a program's output by simulating the steps the computer takes when executing the program, they use representations like tables and/or visual elements to organise their work. These strategies are positively associated with tracing performance. To date, research has focused on how students study instructional materials produced by humans, rather than Large Language Models. What this paper adds We provide insights into the annotation strategies novice programmers from non‐traditional computer science backgrounds use to annotate a ChatGPT solution showing a code trace of a computer program. We identified six strategies; while the words‐only strategy was the most common overall, students used a variety of annotation types, including ones with visual and spatial elements (eg, flowcharts, outlines, lists). As annotation quality increased, the proportions of strategies used changed, suggesting that some strategies may facilitate the production of quality annotations. In particular, the proportion of words‐only and flowchart strategies increased as quality increased; in the top quality quartile, there was a similar proportion of each (with slightly more flowcharts). We integrate several existing frameworks to propose a qualitative method for comparing annotation quality across annotation modalities (verbal, sketched). Implications for practice and/or policy We provide insight into one way instructors could use ChatGPT in a first‐year programming class, that is, use ChatGPT to produce code‐tracing solutions and scaffold student processing of these solutions through annotation activities. We provide evidence that students annotate ChatGPT solutions using a variety of strategies, including ones with visual elements; of the students who provided demographics, the vast majority reported no prior experience.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.