AI snake oil? A risk/benefit analysis for toxicology

Artificial intelligence (AI) is increasingly used to support predictive, mechanistic, and human-relevant toxicology at scale. However, its integration into regulatory science - particularly in drug development - remains uneven, because encouraging technical performance has not yet translated automatically into regulatory trust. Representative AI toxicology studies now span datasets from roughly 10³ chemicals to >3 × 10⁴ peptide or chemical records and report performance ranging from modest in prospective screening settings to strong on narrower, well-curated endpoints. This manuscript presents a critical analysis of the dual nature of AI in toxicology. We review the state of the art in AI-enabled applications, ranging from Green Toxicology and the Human Exposome to specific challenges in safety assessment for biologics and synthetic peptides. Particular attention is given to the gap between rapid model development and regulatory acceptance, highlighted by the challenges of model interpretability, dataset bias, insufficient external validation, and the assessment of complex endpoints like immunogenicity. To navigate these complexities, we discuss the next-generation "e-validation" framework and emphasize the TREAT principle - Trustworthiness, Reproducibility, Explainability, Applicability, and Transparency - as a foundation for building regulatory trust. We hypothesize that AI-based methods in toxicology can achieve regulatory acceptance when they satisfy the TREAT criteria and undergo continuous e-validation within a clearly defined context of use. This framework distinguishes credible AI applications from "snake oil" by establishing measurable criteria for trust-building, including dataset provenance, external validation, uncertainty characterization, and life-cycle monitoring. We argue that AI is neither a miracle cure nor a technological illusion, but a powerful evidence engine that can contribute to a more predictive and ethical toxicological science when it is rigorously validated.