ABSTRACT
Third-party fine-tuning has become a structural node in the general-purpose AI supply chain, yet prevailing liability frameworks continue to treat it as an ordinary act of use. The result is a systematic misalignment between legal responsibility and actual control. This Article argues that fine-tuning is better understood as an act of capability reconfiguration – an intervention that restructures a foundation model’s behavioral boundaries, risk profile, and deployment fit rather than merely invoking capabilities that already exist. Drawing on this recharacterization, it proposes a framework of capability reconfiguration liability built around five elements: control-capability as the basis of attribution; a dynamic duty of care that spans the full fine-tuning lifecycle; an independent data liability regime; an objective foreseeability standard for capability expansion; and mandatory risk disclosure coupled with record-retention obligations. The Article further identifies four characteristic risk structures generated by fine-tuning: bias compounding, capability spillover, deployment mismatch, and liability fragmentation. Comparative analysis shows that the EU AI Act has begun moving toward this position through its substantial modification doctrine, while China’s generative AI regulations provide a normative foundation that stops short of a targeted attribution mechanism. The proposed framework is intended to close this gap and bring liability structures into alignment with the tripartite development-fine-tuning deployment architecture that now defines the general-purpose AI industry.
Zhang, Leyi, Liability for Third-Party Fine-Tuning: Capability Reconfiguration and Risk Structures in General-Purpose AI (April 25, 2026).
Leave a Reply