Toxicology has never been a discipline that rewards speed for its own sake. A flawed conclusion does not stay within a report; it affects product decisions, adversely influences regulatory judgment, and, in some cases, creates a public health risk that cannot easily be corrected.
At the same time, the environment in which toxicologists operate has changed significantly. The volume of scientific literature continues to expand. Evidence now comes from new and old animal studies, next-generation in vitro systems, computational models, and increasingly complex mechanistic data. Safety assessments intelligently integrate this vast body of information while remaining transparent, traceable, and scientifically defensible. However, the challenge of finding, navigating, selecting, and justifiable use of data is more difficult today due to the volume of data (or lack of it in some cases – more common with innovative substances), the dispersed nature of where the data lives, and the high expectation of regulators to have full confidence that nothing has been missed.
This complexity and constraining environment have meant that Artificial Intelligence has entered the work of risk assessment and is under considerable attention and equally significant confusion.
Introduction to the Deep Dive
Over the past two years, discussion of generative artificial intelligence in toxicology has moved from formal reticence to intriguing speculation to structured evaluation. Regulatory agencies have begun evaluating how AI can support Safety Assessment workflows, particularly in evidence management and predictive modelling. The U.S. Food and Drug Administration’s AI4TOX programme is one example. It examines applications ranging from predictive toxicology models to document analysis and digital pathology tools within a defined regulatory science framework.
European regulators are pursuing similar efforts. EFSA’s AI initiatives focus on improving evidence management and integrating data from emerging non-animal methodologies. These programmes signal that AI is being explored as a practical component of risk assessment infrastructure rather than as a substitute for scientific judgement.
Industry discussions point in the same direction. Chemical sector forums and consumer product companies are examining how AI might reduce the manual effort involved in literature surveillance, structured extraction of toxicological evidence, and data integration across safety dossiers.
Despite this activity, an essential question remains unresolved.
Where does artificial intelligence genuinely improve toxicology practice, and where does scientific judgement remain indispensable?
This article begins a short series examining that question. The goal of this discussion is not to promote or dismiss technology adoption but to help understand how AI can be integrated into toxicology workflows to strengthen, rather than weaken, scientific reasoning.
In the articles that follow, we will examine several dimensions of this topic:
- What do different forms of AI mean within toxicology workflows
- Where AI currently improves evidence discovery and structured data extraction
- Where current systems struggle with relevance assessment and uncertainty evaluation
- How regulators and industry are approaching validation and governance
- What a realistic AI-assisted toxicology workflow may look like in the coming years.
Moving Forward
Before examining these questions, it is useful to clarify a basic point. Artificial intelligence in toxicology does not refer to a single technology. Several distinct approaches are entering safety assessment workflows, each with different capabilities and limitations.
Understanding these differences is the first step toward identifying where AI provides genuine value and where expectations remain unrealistic.
