Code Reading Instruction: Essential in the Age of Generative AI

Abstract

It has been established that while natural language reading and writing draw on similar knowledge representations and cognitive processes, they are not identical [4]. Moreover, it is understood that reading interventions lead to better writing performance. Code reading and code writing also draw on similar cognitive processes, even though they are distinct skills. Explicit code reading assessment and instruction has also been shown to help CS students to develop a better understanding of code [1]. While it is yet unclear what role generative AI plays in students' learning processes, it is undeniable that students need to be able to read and understand the code output by these tools [2]. Thus, explicitly teaching code reading skills is even more pressing in our current reality than it was ever before. This work aims at shedding light on the effect of code reading interventions on the use of generative AI for learning. In this lightning talk, I will briefly introduce six types of code reading exercises that can be implemented in large intro CS courses: tracing, reverse tracing, comparing (equivalency), input–output pairing, impractical input tracing [3], and bug finding and fixing. Possible ways to assess the effectiveness of these exercises on students' ability to critically evaluate generative AI output include: comparison of performance on final exams, collection of perception of LLMs, and collection of perception of self [5]. Advantages and disadvantages of different data collection designs will also be concisely presented. Materials at adrianapicoral.com/SIGCSE2026

Ähnliche Arbeiten