Collaborative research: satc 2.0: RES: achieving reliable, interpretable, and efficient at-scale machine unlearning on generative AI models

The National Science Foundation Division of Computer and Network Systems awarded The University of Tulsa $500,787 on October 1, 2026, under the Computer and Information Science and Engineering program (CFDA 47.070) to develop machine unlearning methods for generative AI models.

The project establishes formal criteria for removing data influences and harmful concepts from pretrained models, addressing regulatory compliance and ethical concerns around private user data, copyrighted material, and inappropriate content in systems like GPT and Stable Diffusion. Research focuses on foundational mechanisms of forgetting in diffusion models and large language models, informed by information theory and causal inference. The work introduces a novel task of dismantling high-level AI capabilities inspired by Alzheimer's disease research, offering an approach to steer future superintelligent systems. The project translates theoretical advances into practical unlearning algorithms deployable in real-world systems including e-commerce and finance, while monitoring whether deleted knowledge re-emerges over time.

Performance occurs in Tulsa, Oklahoma, over the period October 1, 2026 through September 30, 2029. This is assistance type Project Grant.

Abstract

Generative models like GPT and stable diffusion have significantly advanced the capabilities of artificial intelligence. Despite their impressive performance, these models can generate undesirable outputs, including private user data, copyrighted material, and inappropriate content. In response, machine unlearning approaches enable the removal of data influences and harmful concepts from pretrained models to comply with various regulations, including restrictions on nsfw content and the general data protection regulation. However, significant challenges remain in machine unlearning, particularly regarding reliability and applicability. The project's novelties are developing a foundational understanding of the mechanisms of forgetting and establishing reliable criteria for regulating unlearning quality. The project's broader significance and importance are that the developed solutions shed light on key ethical concerns about generative models while informing the design of better unlearning methods. The project also seeks to deploy practical unlearning algorithms in real-world systems, including e-commerce and finance. By establishing a formal and rigorous criterion for genuine forgetting and monitoring whether deleted knowledge RE-emerges over time, this project creates a shared foundation for machine unlearning rooted in information theory and causal inference. Simultaneously, the research investigates the underlying mechanisms of forgetting in diffusion models and large language models. Inspired by research on the progression of alzheimer's disease, the project introduces a novel task: dismantling high-level AI capabilities. This breakthrough approach offers a vital mechanism for steering future superintelligent systems. By bridging the GAP between theory and practice, the project translates research advances into real-world applications. The implementation addresses the inefficiencies of frequent model updates by investigating batch processing, managing interdependence, and exploring in-context unlearning. Operationally, the system integrates advanced backend unlearning modules with an intuitive frontend visualization dashboard. This dual setup allows users to easily initiate data erasure requests and transparently track unlearning effects, thereby alleviating privacy and safety concerns associated with generative models. 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.- subawards are not planned for this award.