Pratyush Maini - Proactive Provenance in Foundation Models Preparing for Two Regulatory Futures

Foundation models are trained on web-scale data without clear provenance tracking, creating fundamental challenges for copyright enforcement and accountability. In this talk, I will prepare us for two possible regulatory futures and show how technical mechanisms can address both. In the first future, unauthorized training becomes illegal, requiring us to prove whether models were trained on protected content. I'll present Dataset Inference, which tests whether models were trained on entire distributions rather than individual examples, and introduce STAMP (Spotting Training Artifacts through waterMarked Pairs), a proactive watermarking framework where content creators embed detectable patterns before training begins. STAMP provides statistical guarantees that survive the training process, enabling enforceable attribution. In the second future, "significant data use" fails the fair use test, requiring us to define what constitutes meaningful memorization. I'll present a compression-theoretic foundation where models memorize when they compress information about specific training examples within their parameters, introducing the Adversarial Compression Ratio (ACR) as a rigorous measure. Together, these frameworks shift us from reactive patching to proactive design: if we can build in attribution and measure memorization rigorously, we can architect models where reliability, deletion rights, and legal compliance are native rather than bolted on. Bio: Pratyush is a Ph.D. candidate in the Machine Learning Department at Carnegie Mellon University, and a founding member of DatologyAI. In his work, he has developed scalable and performant methods for improving the quality of data that we train machine learning models on. He has also developed methods that allow us to evaluate, locate, and mitigate the memorization of data points by neural networks. His works have been recognized through a best paper award nomination at NeurIPS, and multiple oral and spotlight talks at majo