New Framework Detects AI Copyright Infringement Across Modalities

This paper introduces the Dual-Branch Conditional Sensitivity (DCS) framework, a novel method for detecting copyright infringement in AI foundation models. Current models often reproduce or rely on copyrighted training data, but simply observing similar outputs isn't enough to prove infringement, as common styles or public domain concepts can also lead to similarities. DCS approaches infringement detection by treating it as a counterfactual conditional distribution shift. It assesses whether a model's behavior would measurably change if a specific protected target were either added to or removed from its training data. This is formalized using conditional differential privacy. The framework operates by creating "learning" and "unlearning" branches around the deployed model, linking their observed displacement to the hypothetical effect of retraining. It also defines a calibrated detection statistic to differentiate true memorization from general model instability. DCS has been instantiated and demonstrated across various model types, including linear regression, diffusion models, autoregressive language models, and multimodal models, showing its broad applicability.