Legal Paper Argues Post-Hoc AI Mitigation Cannot Cure Training Infringement

A new academic paper contends that post-hoc mitigation techniques like machine unlearning and inference-time guardrails cannot retroactively address legal liability from unauthorized data acquisition and training in generative AI systems. The research argues that compliance depends on data lineage rather than output filtering, since unauthorized copying constitutes a legally complete act and model weights function as fixed copies retaining expressive value from training data. The paper further explains that contract law, terms of service, and anti-free-riding principles can restrict data access and use independently of copyright defenses such as fair use or text and data mining exceptions. This analysis emerges as generative AI faces increasing legal challenges, with the machine learning community often relying on post-training mitigation methods to argue for regulatory compliance. The paper's central thesis maintains that the value derived from protected inputs during training creates legal exposure that cannot be eliminated through subsequent filtering mechanisms. This research contributes to ongoing debates about AI training practices and intellectual property rights in the context of rapidly evolving generative technologies.