Human-Centered AI for Software Engineering: Requirements, Reflection, and Road Ahead

Abstract

Since its inception in the 2000s, AI for Software Engineering (AI4SE) has grown rapidly. AI in its different forms, e.g., data mining, information retrieval, machine learning, natural language processing, etc., has been demonstrated to be able to produce good results for automating many tasks, including specification mining, bug and vulnerability discovery, bug localization, duplicate bug report identification, failure detection, program repair, technical question answering, code search, and many more. AI4SE has much potential to improve software engineers’ productivity and software quality. Due to its potential, it is currently one of the most popular research areas in the software engineering field. To advance AI4SE, this keynote puts forward Human-Centered AI4SE. Without considering humans, it is easy for AI-powered tools to hinder rather than help humans in their job or introduce unwanted and unacceptable side effects. Human-centered AI4SE puts humans (i.e., software practitioners) at the forefront of the design of AI4SE tools, with the goal of amplifying and augmenting software practitioners’ capabilities. I will describe some requirements of human-centered AI4SE. Specifically, among others, the need to (i) listen to humans, (ii) learn from (and like) humans, and (iii) synergize with humans. For each requirement, I will present a reflection on the progress the AI4SE area has made over the years, including work done by our research group in Singapore. At the end of this talk, I will describe the road ahead for the above-mentioned requirements toward making AI4SE tools trustworthy, which is an essential attribute to allow them to be widely used by practitioners.