AI as an accelerator of chemical hazard assessment

Consumers and workers are exposed to a wide range of chemicals on a daily basis. In addition, new substances are constantly being introduced to the market. Assessing the potential health and environmental risks of all these substances is essential, but in practice, it is complex, time-consuming, and costly. Traditional toxicological methods fall short in this regard. Artificial intelligence (AI) offers new opportunities to perform chemical hazard assessment faster, more consistently, and in a knowledge-driven manner.

An important development in this field is the use of so-called large language models (LLMs). LLMs are playing an increasingly significant role in processing existing knowledge. Within chemical safety assessment, this means that vast amounts of scientific literature, reports, and regulatory documents are becoming accessible more quickly and in a more targeted manner. LLMs can identify relevant publications, extract key information, summarize it, and establish connections between studies. This helps experts gain a complete overview of the available evidence more quickly and reduces the administrative burden associated with literature research. However, there are significant risks involved, such as hallucinations, which necessitate further research into the use of LLMs.

In addition to the deployment of LLMs, there are in silico Models that make predictions about hazardous properties based on chemical structure and available data. Using machine learning, large datasets of known substances and toxicological effects are analyzed to recognize patterns. These models can provide early indications of properties such as carcinogenicity, reprotoxicity, or environmental harmfulness, even before extensive experimental tests have been conducted. In doing so, they support risk assessment at an early stage and contribute to reducing animal testing and turnaround time.

From endpoint to mechanism

AI applications also make it possible to look beyond just toxicological endpoints. For example, by extracting and summarizing large amounts of information about biological processes using so-called language processing models. In this way, AI systems can contribute to understanding how a substance acts at the molecular and cellular level and through which biological pathways this leads to harmful effects. This approach aligns with developments towards next generation risk assessment, in which understanding the underlying mechanisms is becoming increasingly important for a reliable and future-proof risk assessment.

Practical tools from TNO

Within TNO, these AI developments are being translated into concrete practical applications. For instance, the Substances Information System (SIS) supports professionals in quickly obtaining clear and up-to-date information on substances, including physico-chemical properties, regulations, and health risks. In addition, TNO is working on a substitution tool that can be used to search for substances with similar physico-chemical properties but potentially a lower health class. By systematically comparing substances based on risks and applicability, this tool supports informed choices for safer alternatives without compromising functionality.

AI as a complement to expertise

The use of AI changes the way chemical hazard assessment is performed, but does not replace human expertise. On the contrary: by automating routine tasks and accelerating complex analyses, space is created for experts to focus on interpretation, assessment, and decision-making. AI thus functions as a powerful tool within a broader assessment framework, contributing to a more efficient, transparent, and robust approach to chemical safety.

Want to know more? Bernice Scholten, PhD, toxicologist at TNO, will give the presentation during WoTS 2026. The use of AI in chemical hazard assessment. View it program and sign up.