An Autonomous Governance Framework for Generative AI: Real-Time PII Redaction and Compliance in LLM-Driven Data Pipelines

As Large Language Models (LLMs) are rapidly being implemented into high-stakes applications in order to provide immediate decision-making capabilities, a substantial gap in exists; this is because LLM’s probabilistic nature cannot be governed under existing deterministic frameworks that govern Personally Identifiable Information (PII) or Protected Health Information (PHI), such as the General Data Protection Regulation (GDPR) and/or the Health Insurance Portability and Accountability Act (HIPAA); therefore, an Autonomous AI Governance Framework (AAGF) has been developed to proactively inject real-time PII and PHI redactions at the high-throughput input/output interfaces of all LLM data pipelines in the form of a Context-Aware Output Sanitation Layer (OSL). The scalability and data residency compliance of the AAGF is based upon the development of multi-cloud serverless architectures utilizing Google Cloud Platform and Amazon Web Services to ensure compliance. As a result, the AAGF can serve as a determinate compliance gate that intervenes between an LLM’s output and its end-user(s) in order to sanitize LLMs’ output prior to delivery in order to mitigate potential data leakage risks. Furthermore, we have quantitatively evaluated both the detection efficacy of the combined use of structured metadata extraction from LLM function calls versus traditional techniques, as well as the "compliance tax" (i.e., latency overhead) associated with implementing the real-time governance functions of the AAGF, which validates our framework for implementation in any interactive and pervasive computing application where an unrepudiably auditable log is required for regulatory purposes.