Artificial Intelligence in Chemistry­ and Chemical Weapons ­Control: Fragmentation and the “Execution Gap”

Artificial intelligence is playing an increasingly significant role in chemical research.¹ However, recent developments raise questions about the sustainability of this trend toward widespread and long-term use in research, as well as its implications for the control of chemical weapons. While the 2025 Monitor primarily outlined the emerging possibilities offered by AI agents based on large language models (LLMs), we can now provide a more nuanced assessment of which of these tools are likely to become established in research practice. Although the number of AI-related chemistry publications is increasing,² rather than consolidating existing applications, many studies introduce or test new AI models, contributing to a fragmentation of AI tool use. The large number of models makes it more difficult to evaluate them systematically, develop appropriate safeguards to address their dual use potential, and integrate them reliably into research practice.

Fragmentation of AI Tools in Chemistry

The fragmentation of AI models is exemplified by the case of the AI chemistry agent ChemCrow³, whose advantages and capabilities were discussed in the 2025 CNTR Monitor.⁴ While ChemCrow is successful in terms of the number of citations it has received, it is significantly less so when measured by the number of scientific studies that actually use the tool in practice. An analysis of its more than 700 citations shows that a substantial proportion come from review articles and methodological papers that either develop analogous models, introduce improved machine learning techniques, or situate ChemCrow within the broader development of such tools. In contrast, the direct application of ChemCrow as a standard tool has thus far not been observed. This does not imply that the model is inapplicable; rather it highlights the dynamic nature of AI capability development. The rapid emergence of new, specialized, or more powerful models makes it difficult to permanently integrate individual tools into research workflows. This has resulted in an overwhelming proliferation of available tools, leading to the fragmentation of both AI tools and AI abilities. Another factor contributing to the fragmentation of AI applications in chemistry is that the most commonly used LLMs are incapable of performing chemistry-specific tasks. Although alternative AI models, such as graph neural networks, can address some of these tasks, they require task-specific training and can therefore not be used solve problems across different areas of chemistry.

The fragmentation highlights a fundamental problem for AI in chemistry: a growing gap between the methodological relevance of AI models—that is, their prominent role in the development and discussion of methodologies—and their widespread scientific application.

The AI Execution Gap in Chemical Weapons Control

While AI is already being used in certain areas of chemical research despite the challenges of fragmentation and rapid technological change, the execution gap is even more pronounced in chemical weapons (CW) arms control. Notwithstanding the considerable theoretical attention devoted to AI in arms control by the OPCW, its application in operational processes has so far remained limited.⁵ The reasons for this lie less in technical limitations than in the political sphere. The obstacles include geopolitical mistrust, dual-use concerns, the “black box” nature of AI-generated results, limited access to relevant data, and potential constraints on the mandates of international organizations. This gives rise to two problems for arms control. First, the risks posed by AI in the chemical sector, as described in last year’s Monitor, remain largely unregulated due to the high degree of fragmentation. Second, the opportunities that AI offers for strengthening arms control have yet to be fully exploited.

Meanwhile, the rapid pace of AI development is coming up against political structures whose inherent limitations constrain their ability to respond swiftly. There is a growing need for regulatory action across a broad range of AI models, not only LLMs, for which dedicated monitoring and governance frameworks remain lacking. These include:

• Graph neural networks. These can predict molecular properties, including toxicity, and can therefore be used to optimize therapeutic and potentially toxic molecules.⁶ However, the stability and suitability of the generated molecules for their intended purpose still need to be investigated under real-world conditions.

• Large language models. These democratize chemical and technical knowledge, making it easier for people without specialist expertise to synthesize toxic substances and understand methods of CW dissemination.⁷ That said, this risk is minimized by the need for laboratory infrastructure, limited access to the necessary materials, and a lack of practical safety knowledge.

• Retrosynthetic models. These propose new multistep synthetic pathways for any target molecule. In doing so, they do not distinguish between therapeutic and toxic substances.⁸ However, specialist knowledge is still required to identify feasible synthetic pathways from such model outputs.

• Agentic AI. This combines several other AI models in a single application and can be used in everyday language wire prompts, much like an LLM. Figuratively speaking, it acts as the conductor of an orchestra of AI models, coordinating them and making them accessible for people with less IT knowledge.

Many open-access models lack model-specific legal safeguards, while the built-in safety mechanisms that do exist are limited. Recent evaluation studies of large language models suggest that, although current safety measures can reduce harmful outputs, they cannot reliably prevent them and are often vulnerable to circumvention strategies.⁹

Within the spectrum of AI’s potential contributions to chemical weapons control, two particular risks stand out: inadequate oversight of fragmented and unregulated AI models, and an execution gap that leaves AI-enabled opportunities for weapons control underutilized. For example, the analysis of large datasets, such as trade data, open-source data, government declarations, and inspection reports, could be used to support verification processes by identifying unusual production patterns. In addition, automated analytical chemistry and forensic techniques could be used to provide rapid countermeasures following a chemical weapons attack, while AI could support mission planning.

Governance Implications and Actions

Artificial intelligence technologies pose significant challenges and risks for chemical weapons control. At the same time, it is important to emphasize that the CWC is one of the most robust international disarmament treaties, underpinned by widely accepted international norms. While AI could contribute to the development of chemical weapons and lower the barriers to access for non-experts, it is currently unlikely to pose an immediate threat in this domain. This, however, does not necessarily constitute a permanent guarantee. Consequently, the validity and effectiveness of the convention must be continually strengthened and developed to address the new challenges posed by emerging technologies, while also harnessing their potential benefits. The OPCW is already active in this regard, as evidenced by initiatives such as the Scientific Advisory Board’s AI Report, the Temporary Working Group (TWG) on AI, and the AI Research Challenge. Through these activities, the OPCW is promoting the use of AI in its own work. This includes AI-supported analyses of reports, declarations, and open-source information to identify trends or anomalies; enhancing the capabilities of chemical forensics; and using AI-based tools for training and exercises. The risks that AI poses to the prohibition of chemical weapons can be countered by systematically monitoring AI developments, providing technical information, and supporting capacity building. Therefore, the OPCW’s Temporary Working Group on AI recommends that the OPCW and states parties engage with the developer and research communities to highlight the dual-use risks of AI-enabled chemistry and promote responsible innovation. This includes discussing safeguards and access restrictions for particularly sensitive applications (e.g., toxic molecules, retrosynthetic planners, and autonomous laboratories), as well as disseminating best practice. The TWG also recommends recruiting data/AI experts with knowledge in chemistry to increase its internal capacity in this field. These measures have the potential to provide a robust response to new technologies, but their practical implementation is constrained by political obstacles. Overcoming these hurdles should be a key priority moving forward. For instance, the credibility of AI model results could be enhanced by deploying duplicated hardware and software systems across several states parties to verify the consistency of AI results. Such an approach would ultimately help uphold the robustness of the prohibition against chemical weapons.