Collaborative Research: SaTC 2.0: RES: Local Rules, Global Models: Jurisdiction- Aware Solutions for Copyright Vulnerabilities in LLMs

Large language models (LLMs) increasingly support writing, learning, search, tutoring, and creative work, but they may also introduce new copyright risks and vulnerabilities. These systems can reproduce protected books, lyrics, assessments, articles, or other proprietary materials, and the risk changes across countries, licenses, and uses. A mitigation that may be acceptable in one setting can be inappropriate in another. Current security solutions and safeguards often treat copyright compliance as a one-time fix, but real systems face changing laws, new copyrighted works, and adversarial user prompts. The project's novelties are a jurisdiction-aware and lifecycle-based approach that connects LLM model training, real-time use, and post-deployment refinement. The project's broader significance and importance are that it helps make generative artificial intelligence more secure, safer, more accountable, and more usable for publishers, educators, researchers, businesses, and the public. The project supports innovation by reducing copyright leakage while preserving the useful capabilities of generative systems. The project develops a technical framework for copyright-safe generative writing in LLMs. It creates structured representations of jurisdiction, license, and usage context, and uses these representations to guide model behavior through efficient adaptation and targeted unlearning. It designs legal risk-aware reward models so that systems prefer useful responses that also respect local copyright constraints. The project also builds inference-time safeguards that analyze internal model states before generation, monitor partial outputs during generation, localize risky spans, and regenerate safer alternatives while preserving the user's intent. In addition, it develops a feedback-guided data editing process that uses detected failures to revise, remove, add, or reweight training examples, allowing the system to adapt as copyright laws, content, and user behavior evolve. The work is evaluated in publishing and smart agriculture settings, where copyright and regulatory constraints directly affect real-world deployment. Expected outcomes include methods, datasets, evaluation tools, and prototype systems that help organizations deploy generative artificial intelligence with lower legal risk and stronger public trust. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.