Front Toxicol. 2026 ;8
1877128
Arianna Bassan,
Alexander Amberg,
Maria Letizia Barreca,
Paul M Bradley,
Frank Bringezu,
Joyce De Paula Souza,
Corinne Guicheney,
Elena Lo Piparo,
Simona Kovarich,
Andrey Massarsky,
Georgia Maundrill,
Lance Molnar,
Cristina Parenti,
Patricia Parris,
Manuela Pavan,
John F Reichard,
Ron Steigerwalt,
Olga Tcheremenskaia,
Kenneth Unice,
Christopher J Waine,
Geraldine Whelan,
Angela White,
Glenn J Myatt.
Artificial intelligence (AI), particularly in the form of large language models (LLMs), offers promising support for toxicological risk assessment (TRA), which integrates multiple lines of evidence to support health-based conclusions, including those used in regulatory decision-making. Effective use of LLMs in TRA requires a clear understanding of their strengths and limitations, especially in the context of regulatory expectations for transparency, reproducibility, and accountability. This work explores the emerging role of LLMs in TRA through three integrated components: an overview of LLMs in the context of TRA, an exemplar case study examining LLM-assisted derivation (mainly using ChatGPT) of the Permitted Daily Exposure (PDE) for acetyl tributyl citrate under the draft ICH Q3E guideline, and a workflow hypothesis for the integration of LLMs into TRA. Observations from the case study suggest that LLMs may support pattern recognition, targeted information retrieval, and rapid summarization of large bodies of text, facilitating literature triage, extraction of study details, and evidence summaries. At the same time, the case study highlights important limitations for LLMs. Observed variability in LLM outputs reflects inherent model stochastic behavior, model limitations, and the intrinsic complexity of TRA, where different experts may reasonably reach distinct, yet scientifically defensible conclusions based on the available evidence. The observations from the case study, interpreted in the context of the available literature, are used to inform a hypothesis that LLMs may be most effectively integrated into TRA through structured, modular, human-supervised workflows. This hypothesis reflects the view that the suitability of LLM-assisted TRA depends not only on LLM capabilities, but also on how LLMs are integrated into the assessment process to address both LLM variability and the interpretive nature of TRA. The proposed pragmatic workflow separates data search, information extraction, and integrated summarization while maintaining human-in-the-loop control over interpretation, expert judgment, and final conclusions.
Keywords: AI; LLM; PDE; human-in-the-loop; pragmatic workflow; toxicological risk assessment