Ctrl+Z for GenAI: Undoing What Models Learn and Teaching them
Dr. Mayank Vasta

What if we could hit “Ctrl+Z” on a model, asking it to forget outdated, harmful, or unauthorized behaviors? And what if it truly grasped negation, reliably reasoning about what is absent, disallowed, or false across language, vision, and audio? This talk explores two complementary frontiers in generative AI: machine unlearning and multimodal negation. We will discuss practical techniques for unlearning, removing the influence of specific data or capabilities while preserving overall performance, to support privacy, compliance, and rapid iteration. We will also discuss how making models sensitive to negation improves both interpretation and generation: following instructions like “not blue,” describing scenes “without people,” or detecting “no speech present” in audio. Together, these directions point to AI systems that are easier to update, audit, and deploy, while meeting emerging regulatory and safety expectations. The session maps the problem space, recent methods and benchmarks, and open challenges in scaling unlearning and robust negation across modalities.