Reviewing AI Code Is Not A Viable Argument (softwaremaxims.com)

In a critical examination of LLM Coding Assistants, Thomas Depierre argues that these AI tools are not yet viable for enhancing software development efficiency. He points out that, despite their popularity, LLMs frequently generate incorrect code, requiring developers to engage in thorough reviews—a process complicated by empirical findings that indicate effective code reviews can only handle about 400 lines of code per hour and should not exceed an hour’s duration. This raises questions about the practicality and productivity of integrating LLMs into development workflows, especially when review time consumes much of the potential gains from automated coding. Depierre highlights the limitations of the “just review everything” approach, noting that the output quality from LLMs could be lower than that from human coders, yet reviewers may feel falsely confident in their assessment. He calls for empirical studies to evaluate how well developers can review code generated by LLMs compared to human-written code, emphasizing the need for robust data to support the effectiveness of these AI tools. This discussion is crucial for the AI/ML community as it underscores the importance of scrutinizing the actual benefits and limitations of AI-powered coding before widespread adoption.