Same Scrutiny, More Time: Eye Tracking Insights into Reviewing LLM-Labelled Code

Abstract

This paper uses eye tracking to study how developers review code that has been labelled or generated by large language models, finding that reviewers apply the same level of scrutiny but take more time compared to reviewing human-written code.

Mazen Mohamad

Researcher at RISE Research Institutes of Sweden

My research interests include Safety & Cybersecurity of Autonomous Systems, AI4Security, and Security4AI.