Hae Jin (Hayley) Song is a Fellow at the Berkman Klein Center at Harvard University and an AI Research Fellow at ThoughtWorks, where she works on the geometric foundations of AI interpretability and safety.
"[-Her research develops formal, geometry-based frameworks for understanding, analyzing, and steering the behavior of modern generative models. Hayley holds a Ph.D. in Computer Science from the University of Southern California. She earned her B.S. and M.Eng. in Electrical Engineering and Computer Science from MIT, with a minor in Mathematics. During her M.Eng., she specialized in artificial intelligence and worked under the joint guidance of Prof. Regina Barzilay and Dr. Julian Straub on computer vision problems in medical imaging, including non-rigid mammogram registration for breast cancer detection and 3D reconstruction of human arms for lymphedema screening. Across her academic career, Hayley has conducted research at MIT (CSAIL, Media Lab, McGovern Institute), INRIA (ILDA Lab), and USC (Information Sciences Institute, Visual Intelligence and Multimedia Analytics Laboratory (VIMAL), iLab, Knowledge Computing Lab). She has also had industry research internships at Apple, MathWorks, and a French robotics startup, Keecker. * Research Interests Geometric Foundations of Generative Models; AI Interpretability and Safety Hayley’s current research focuses on how complex, high-dimensional information-processing systems, particularly modern generative models and other AI models, behave, using tools from differential geometry, Riemannian manifolds, and latent-space topology.
Her goal is to develop principled and scalable methods to characterize, attribute, and control model behaviors by identifying geometric “fingerprints” of their internal representations and dynamics. Her work addresses foundational questions in the analysis and controllability of generative models, with applications to AI interpretability, model attribution, deepfake detection, bias and degeneration analysis, and model steering for safety and alignment. Her recent work has made significant contributions to these questions through research on fingerprinting generative models: In papers published at CVPR 2024 and ICCV 2025 (Highlight), she introduced a novel theoretical framework that represents model behavior on data manifolds and formalized the notions of “artifacts” and “fingerprints” of generative models in a geometric language. This framework enables effective attribution and systematic comparison of state-of-the-art generative models. More broadly, through this line of research, she aims to advance a generalized theory of generative models and their internal mechanisms, to help shape the safe and responsible integration of Generative AI into our society in ways that serve humanity.