CAREER: Metamorphic Debugging for Responsible AI-Software Development

With artificial intelligence (AI) increasingly embedded into critical software systems—from healthcare to defense applications—ensuring these AI-enabled software solutions (a.k.a. AI-Software) meet comprehensive responsibility standards has emerged as a pivotal challenge. Traditional software debugging methods, centered around identifying and rectifying defects, fall short in addressing the unique nature of AI-software. This project seeks to develop novel, principled debugging techniques specifically designed to detect and resolve responsibility - that is, issues related to transparency, accountability, and impartiality - before AI-software is deployed publicly. The project also includes education-related activities, where results from the research will contribute to the computing curriculum expansion. The project proposes the development and evaluation of a transformative debugging framework tailored to AI-software, which incorporates pre-trained deep neural networks and large language models during execution. At the core of this framework is the novel concept of metamorphic debugging, which explores the intersection of metamorphic testing - a method analyzing output variations caused by structured input changes - and relational verification across three dimensions: causality, information theory, and extreme value theory. Unlike conventional debugging, which inspects single program executions independently, metamorphic debugging simultaneously examines relationships across multiple program-runs to uncover nuanced requirement violations. The research will include creating methods for generating intuitive test cases sensitive to causal relationships, developing techniques for elucidating higher-order relations among multiple execution traces, and applying statistical approaches to rigorously assess worst-case scenarios for undetected defects. The framework's effectiveness will be routinely validated through integration into real-world AI-software, collaborations with industry partners, and surveys assessing student and developer perceptions of responsible AI-software. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.