How Would an AI Verification System Work in the Biological Weapons Convention?
The U.S. proposal to introduce an AI verification system for the Biological Weapons Convention (BWC) cannot create a verification regime that the convention has not yet agreed to, but it can strengthen trust among the treaty’s members by improving how they collect, compare, and interpret information within the treaty’s framework.
At the September 2025 UN General Assembly, President Trump announced his intention for the United States to lead an international effort to enforce the BWC by “pioneering an AI verification system that everyone can trust.” While the proposal is timely and relevant, it will likely face political and technical challenges that multilateral diplomatic efforts have experienced. However, there are important treaty-compatible uses of such a system worth testing.
Biotechnology is rapidly advancing with many promising applications, such as gene therapies, accelerated vaccine development, and advanced therapeutics. However, these can also pose risks, as the same tools and knowledge can increase bioweapon-relevant capabilities, including through dual-use research of concern (DURC): having legitimate scientific purposes, but with the potential to be accidentally or intentionally misused to create new and/or modified biological weapons. AI’s advancement and its increasing convergence with biotechnology can amplify these dual-use dynamics by accelerating biotech design and optimization and lowering technical barriers. This makes a convention like the BWC and its influence in regulating biotechnology much more critical, because it can clarify boundaries against how biotech weaponization shapes norms and expectations for how they are implemented nationally, and provide avenues for strengthening transparency, oversight, and cooperation around these emerging technologies.
However, the BWC faces implementation challenges. It lacks a standing verification regime, which is necessary for establishing an agreed process of data collection, information provision, analysis, and follow-up among members. The BWC has been unable to agree on a verification mechanism since its inception in the 1970s. The last major diplomatic effort to close this deficit failed in 2001. Discussions on verification have so far been hindered until the agreement to establish a new “Working Group” on strengthening the convention in 2022.
Where AI would fit with BWC workflows
Amid the lack of a formal inspection regime, AI could aid verification gaps by analyzing data required for verification at a scale and speed that would otherwise take longer if just done by human analysts. While information on the United States’ specific plans for the proposed AI verification system is still sparse, we should note that any credible AI approach should be anchored to current BWC procedures, and where States Parties (countries that have ratified or acceded to the BWC) agree.
One of the BWC’s existing mechanisms is its Confidence-Building Measures (CBMs), a transparency channel for the States Parties consisting of voluntary annual information exchanges. Other mechanisms include Article V consultations between States Parties, Article VI complaints to the UN Security Council, and, if bioweapon use is claimed, recourse to the UN Secretary General’s mechanism. AI can add practical value to these mechanisms if used as a tool that improves how humans clean, prioritize, and interpret information used in the consultations. Given these same functions, AI can also be used beyond improving CBMs or the convention’s existing mechanisms, such as in assisting drafting “declarations” of States Parties’ compliance–borrowing from the structured format of the Chemical Weapons Convention’s declarations–without making any new legal obligations (which can be contentious for some states). Crucially, AI should only be treated as a support tool, not as an automated arbiter or the main determiner of treaty compliance.
Off-site analysis
One factor that makes verification difficult is its reliance on “on-site” inspections. States Parties have been skeptical of allowing external actors into their dual-use labs and biodefense facilities because biological activities are often indistinguishable from legitimate research, involve proprietary technologies, and can be cleaned rapidly. Hence, one of the most immediate potentials of an AI verification is off-site analysis, i.e., verification activities conducted without the physical inspection of a facility or location within a state party’s territory.
AI systems can process CBMs and large volumes of public or commercially available information – or open-source intelligence (OSINT) – to make a more coherent picture of national activities relevant to the convention, without having to make on-site visits. As this data comes in different formats and languages, AI –through natural language processing (software letting computers read and compare human language)–can simplify many CBM submissions. It can also identify omissions or inconsistencies in reporting vis-à-vis the data, pointing to any need for follow-up. This is most valuable in reducing the administrative burden on states, which could improve both the timeliness and completeness of reporting (a common struggle faced by the convention).
Similarly, AI can aggregate different data relevant to the convention’s compliance: for instance, public laws and policies, budget lines, procurement and facility notices, scientific publications, patents, and public health reports. Simple anomaly-detection models (modeled after other disarmament treaties like the International Atomic Energy Agency) can then highlight changes in patterns. For example, AI systems can flag sharp shifts in procurement, which can be associated with working with highly infectious pathogens, or it can flag an abrupt decline in publications from a facility typically active in a relevant field. A human analyst could then review these flags. These tools would not aim to attribute non-compliance to States Parties, but instead generate “leads,” improve comparability across years and countries, and help states decide when it is worth asking clarifying questions through Article V.
Predictive analytics
AI predictive analytics (using AI to identify patterns and forecast emerging risks from diverse, large-scale data sources) can add a forward-looking perspective. By combining data like trade flows, funding trends, research trajectories, and other OSINT, it can clarify emerging biological risk patterns that may not be obvious in any single dataset. Used carefully, this can support early-stage, less contentious conversations among States Parties and reduce the likelihood of concerns that appear in highly politicized settings.
To make this system work at scale, a few modest technical changes would help. Adding a small number of computer-friendly fields to the CBM forms would make the reports easier for the AI to process without exposing sensitive science. Simple upgrades to the BWC’s online portal would also allow the treaty’s support unit to take in this new, clean data and display simple summaries for all member states to see.
Support for visits and investigations
Although the BWC has no routine inspections, AI would still be relevant if States Parties agree to visits or if an investigation proceeds under the UNSG’s mechanism. Further, AI-based satellite imagery can help teams plan inspections, digital audit techniques can map collaborations and procurement trails relevant to a specific concern (subject to appropriate due-process and privacy safeguards), and AI-supported microbial forensics could assist in epidemiological modeling, helping distinguish engineered versus naturally occurring events and identify plausible pathways of a pathogen’s spread. Aligned with this, the BWC’s Implementation Support Unit (the convention’s administrative body, or the ISU) or the States Parties can train inspectors and diplomats of the convention with a simulation based on the data generated.
Science & technology review, surveillance, and assistance
AI may facilitate a regular review of science and technology (S&T) relevant to the BWC, which can be done under its S&T mechanism designed to systematically review and integrate ongoing S&T advances. AI’s help in rapidly scanning public sources, like papers, patents, news, and budgets regarding advances in life sciences and convergent technologies, can give officials with non-technical backgrounds a baseline and clear summary of S&T trends to help separate misconceptions and emphasize evidence. This same AI-assisted analysis should aim to create a pathway for constructive diplomacy. Where AI-assisted analysis points to gaps in limited capacity rather than malicious intent, the appropriate response is not punishment but support through the BWC’s International Cooperation and Assistance Mechanism. This mechanism allows for training, tools, and funding that enable better compliance rather than penalize states struggling with resources.
Outside of the BWC: aiding in nucleic-acid synthesis screening and public health
There are different applications where AI can help strengthen the BWC outside of its own mechanisms. Apart from aiding CBMs, for instance, AI can also strengthen verification by analyzing biotech relevant to the convection, such as through global nucleic-acid synthesis screening: a process of reviewing commercial purchase orders and customers for synthetic DNA and RNA products to prevent them from being misused. Many modern biotechnology projects begin by ordering custom DNA or RNA. Traditional and current practices of synthesis screening rely on lists of regulated agents and toxins, which are unevenly distributed across providers and are unsuitable for detecting novel or engineered sequences. A standardized, AI-enabled screening approach and AI biology models (like Evo 2) could help screening companies assess the functional risk in novel sequences that do not match any entry on these lists. By learning signatures associated with pathogenicity or toxicity, such systems could flag concerning orders before synthesis, assign risk levels, and route the highest-risk cases for additional review or halt pending clarification.
Earlier government research demonstrated that machine-learning models can infer functional properties of DNA sequences from sequence data with useful accuracy. Updating and operationalizing these methods with AI could reduce the burden on verification by preventing misuse earlier on.
Note that this would not require the BWC to regulate industry directly. It instead benefits from policy partnerships that promote interoperable screening standards and shared evaluation benchmarks, such as the International Gene Synthesis Consortium (IGSC), a consortium of gene synthesis companies that voluntarily agree to apply a common screening protocol of their customers’ orders.
In a wider public health angle, AI could support disease surveillance, vaccine target identification, and efficacy forecasting to reduce the impact of potential biological attacks. By promoting public health resilience, AI can participate in a global system that lowers incentives for deterring misuse of biotechnology.
Implementation challenges
The asymmetry of data availability is a key challenge to implementing an AI verification system. AI systems perform best where information is abundant and comparable, but that is not always the case on biotechnology issues. Reporting quality would vary based on where there is more open, well-documented data. Consequently, the system may over-flag open, data-rich states (usually wealthier states) while under-monitoring low-capacity environments, creating bias. This is further complicated by the differences in public availability of data, for instance, with the CBMs – the ISU’s 2025 Annual Report notes, for example, that while 113 States Parties submitted CBMs, 74 of them (65%) kept their reports ‘restricted.’ Consequently, an AI verification system trained only on the public data would be effectively blind to the majority of the compliance landscape, disproportionately scrutinizing the minority of states that choose transparency.
A second challenge is building the system’s legitimacy among the States Parties. States Parties are unlikely to accept “black box” judgments of their compliance. AI outputs should be explicitly delivered as clear caveats, not verdicts. These should trigger concise pathways for resolving AI-raised concerns. This process can include notification, explicit reasoning, a written reply period, and if questions remain, recourse to the Article V consultations mentioned above. Structuring the procedure this way makes it predictable and fair.
A third challenge is political. The system’s design will have to navigate verification and tension with national sovereignty, in terms of their control over their data, facilities, and compliance decisions. States prize their sovereignty and privacy. There can be fears of having to be fully transparent due to negative misinterpretation and misuse of their data (e.g. espionage or a political attack). Hence, governance becomes a key question: who would operate and oversee this AI system, under what rules, and with what representation? Some governments will likely favor stronger prevention of biological risk, but resist arrangements appearing to be centralized control in a specific country (like the United States) or in a single company’s system.
Finally, there can be bad-faith use or access to the system. For example, bad actors (any nonstate actors, or even states, seeking to misuse biotech) can find ways to hurt the system. They can input false sources and misinformation for AI to use, or use any other tricks to make the system misbehave (e.g., jailbreaking). This behavior will require oversight and accountability, such as a low bar for flagging misbehavior, periodically “red-teaming” the system, and randomized system checks. These measures can make misuse harder without imposing intrusive measures that a state would be sensitive to.
However, as Carnegie scholars have noted since the President’s speech, these hurdles should not obscure diplomacy. While the U.S. proposal faces technical skepticism, it also brings a window of opportunity to engage a historically reluctant U.S. administration in multilateral disarmament, alleviate the deadlock on verification, and channel great power interest into practical multilateral support. The challenge then is pivoting from analyzing the political rhetoric to implementing a credible technical pilot that modernizes the convention for all States Parties without violating the treaty’s consensus-based framework.
Piloting a system
Given these constraints, the BWC can consider conducting a pilot test of an AI verification system, such as a one-year, voluntary project limited to off-site use. An AI system can flag leads of BWC misuse primarily from open-source information – already available reported CBMs and public data. This can test whether AI-assisted analysis improves clarity for States Parties. Operationally, the pilot could pair small CBM submission form changes described above with a modest portal update of the CBM process at the ISU and would commit to publishing policy-level summaries, not code or models. It would add a clear due-process response path for AI-flagged concerns and funding templates, translation, and small grants to keep the ability to submit CBMs more equal.
The pilot’s success could be measured by improved CBM completeness and consistency, the number and resolution of clarifying exchanges initiated from verification leads, the breadth of participation of States Parties, and user feedback on burden and fairness.
This is a practical path translating a high-level pledge to operational improvements. AI can make the BWC’s information flows more usable, help States Parties detect meaningful signals in the noise, and support faster, better-grounded consultations when concerns arise. It cannot, however, unilaterally create a verification system that the treaty does not authorize. If framed correctly – as an aid to transparency, monitoring, and assessment within existing procedures paired with measures that preserve equity and legitimacy – AI has promising applications for strengthening convention adherence and modestly improving bioweapon prevention and response.
The BWC has had its seventh Working Group session in December 2025, with only draft reports explicitly recommending establishing a new “Open-Ended Working Group (OEWG) on Compliance and Verification,” which can be a potential entry point for an AI pilot. The BWC has four meetings scheduled in 2026.
Views expressed are the author’s own and do not represent the views of GSSR, Georgetown University, or any other entity. Image Credit: Scientific American