Risk-based governance of AI-tools used in health-based exposure limit derivation

Artificial intelligence (AI) is increasingly applied to toxicological assessment and health-based exposure limit (HBEL) derivation, offering potential gains in literature retrieval efficiency, methodological consistency, but introducing novel risks such as hallucinated citations, misattributed findings, and fluent yet scientifically shallow narratives. This article examines AI-assisted HBEL derivation through established HBEL quality framework and clarifies its regulatory expectations. HBEL monographs function as scientific expert assessments that inform GMP risk-management, rather than GxP records, a key distinction that is central to establishing proportionate GMP validation requirements. Building on EMA Annex 11 and draft Annex 22, FDA guidance on AI-supported decision-making, and ICH M7 model qualification principles, the article proposes a tiered, risk-proportionate validation framework. Full computerized system validation is not inherently required for AI tools supporting HBEL derivation. Instead, governance should focus on data limitations, and data source verification. Crucially, experienced human toxicologist review and peer review remain the primary, non-negotiable quality controls, ensuring scientific accountability and materially reducing the validation burden on AI components.