Total Recall: Memorization, Guardrails, and Output Infringement in AI Models

Almost every conversation about generative AI and copyright focuses on the same question: is training a model on copyrighted works fair use? Underneath that question is a set of technical ones that get less attention. What actually happens to a book, a script, or a song when a model is trained on it? What does it mean for a model to “memorize” a work, and how is that different from learning from it? Is the work still in there afterward, and could anyone tell?

Our panelists are two of the leading researchers on what language models memorize, and they work on the same problems copyright lawyers are now litigating. Can a rights holder show that its work went into a model, and how reliable are the tests for that? When a chatbot has been trained not to recite a book, is the book gone, or is it still in there? Guardrails can filter out word-for-word copying, but what about non-literal copying that still can constitute infringement? And once a work is in a model, can it be taken back out?

Please join us as Professors Niloofar Mireshghallah and Robin Jia explain the technology for copyright lawyers: how training and fine-tuning work, how researchers measure what a model has retained, how memorized content gets out, whether by accident or on purpose, and what all of that means for copying, infringing outputs, and remedies.

Niloofar Mireshghallah is an Assistant Professor at Carnegie Mellon University, jointly appointed in the Language Technologies Institute and the Department of Engineering and Public Policy, a core member of CyLab, and a Founding Member of Technical Staff at humans&. Her research lies at the intersection of privacy, natural language processing, and the societal implications of machine learning, studying how data shapes model behavior and how people can control the use and disclosure of their information. Previously, she was a Research Scientist at Meta FAIR and a postdoctoral scholar at the University of Washington. She received her Ph.D. in Computer Science from UC San Diego in 2023. Her honors include the NCWIT Collegiate Award, the Rising Star in Adversarial ML Award, and selection for Rising Stars in EECS.

Robin Jiais an Assistant Professor of Computer Science at the University of Southern California. His research focuses on scientifically understanding large language models, including elucidating their internal mechanisms, characterizing their tendencies towards misaligned behavior, and tracing both their mechanisms and behavior back to their training process. Robin received his Ph.D. in Computer Science from Stanford University, where he was advised by Percy Liang. His work has received best paper awards at ACL and EMNLP.

This event is eligible for 1.0 California MCLE credit.