“Will I be replaced?” Assessing ChatGPT's effect on software development and programmer perceptions of AI tools

Abstract

ChatGPT is a language model with artificial intelligence (AI) capabilities that has found utility across various sectors. Given its impact, we conducted two empirical studies to assess the potential and limitations of ChatGPT and other AI tools in software development. In the first study, we evaluated ChatGPT 3.5′s effectiveness in generating code for 180 coding problems from LeetCode, an online coding interview preparation platform. Our findings suggest that ChatGPT 3.5 is more effective in solving easy and medium coding problems but less reliable for harder problems. Further, ChatGPT 3.5 is somewhat more effective at coding problems with higher popularity scores. In the second study, we administered a questionnaire (N = 99) to programmers to gain insights into their views on ChatGPT and other AI tools. Our findings indicate that programmers use AI tools for various tasks, such as generating boilerplate code, explaining complex code, and conducting research. AI tools also help programmers to become more productive by creating better-performing, shorter, and more readable code, among other benefits. However, AI tools can sometimes misunderstand requirements and generate erroneous code. While most programmers are not currently concerned about AI tools replacing them, they are apprehensive about what the future may hold. Our research has also revealed associations between AI tool usage, trust, perceived productivity, and job security threats caused by the tools.