SHF:Small:NSF-MUR: Automated and Explainable Provenance of AI-Generated Code

Generative artificial intelligence (AI) is rapidly transforming how software is created, enabling developers to quickly generate code, explore unfamiliar systems, and build new functionality. AI systems, however, can also produce code with unclear origins, raising important risks such as licensing and copyright violations, exposure of sensitive information, and insecure, incorrect, or unmaintainable software. This project addresses the need to make AI-generated code more transparent, trustworthy, and safe by enabling developers, AI engineers, and compliance practitioners to understand and evaluate where such code comes from and how it is produced. To address this need, the project will develop novel automated and explainable techniques for tracing the provenance of AI-generated code across complex AI pipelines, combined with new empirical knowledge about stakeholder provenance needs, practices, and challenges. The project will enable stakeholders to understand and evaluate the origins of AI-generated code, allowing them to make informed decisions about its use and integration. It will equip them with methods for producing AI outputs that are more transparent, auditable, and reliable, thereby supporting responsible adoption in practice and mitigating risks related to software licensing, security, and quality. The project will integrate its findings and techniques into computer science curricula, preparing students with foundational knowledge of AI provenance and the skills needed to responsibly develop and use AI-generated code. The project will advance the state of the art in explainable provenance for AI-generated code through two integrated components. The empirical component will conduct surveys and interviews with developers, model users, and compliance practitioners to construct taxonomies of provenance practices, challenges, and requirements. Guided by these insights, the technical component will develop automated techniques to establish traceability links between generated code and its contributing sources, including prompts, training data characteristics, external artifacts, and internal model components. The project will be developed through cross-cutting research on causal inference methods for model interpretability, local and global explanation techniques, model modularization and attribution analyses, and scalable code search and clone detection. The expected advances include new models and representations for provenance-aware AI/ML systems, as well as toolchain integrations that enable AI coding assistants to provide explainable provenance information, supporting licensing compliance, code quality assessment, and extensions to security and privacy analysis. 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. NSF Award ID: 2533367 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Oscar Javier Chaparro Arenas | Institution: College of William and Mary, WILLIAMSBURG, VA | Award Amount: $599,959 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2533367 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2533367.html