Generative AI and Weapons of Mass Destruction: Will AI Lead to Proliferation?
Technical troubleshooting: An important but distinct area from knowledge retrieval relates to troubleshooting in which the user provides the model with data on a specific real-world process, such as the log files from a piece of manufacturing equipment. The LLM can interactively troubleshoot with the user to overcome whatever problem has been encountered. This pathway is not necessarily distinct from being able to ask for help online, except that it could also be applied to proprietary and domain specific equipment. Technical assistance is an area that is currently controlled from an export control perspective when the technical assistance relates to weapons of mass destruction. As such, there is a need also to examine the potential ability of LLMs to provide technical assistance from an export control perspective.
Mathematical, biological, simulation, and engineering design: Each CBRN domain includes many specific subfields and processes, and the staff working in CBRN programs often have scientific interdisciplinary backgrounds to account for this. In the nuclear domain, physicists and nuclear engineers are vital. But so too are chemists, mechanical engineers, electrical engineers and even geophysicists. For each domain, subdomain, and process there is often specialist software and technical data libraries that are leveraged in solving practical engineering and scientific problems. Some of this software and technical data is export controlled because of the proliferation concerns associated with it. It seems likely that LLMs will be capable of supplanting at least some of this specialist software and technical libraries in the future. One can envision an engineer interacting with an LLM to optimize a rocket design with the LLM also running mathematical models to inform the interaction, for example. This leads to questions around whether LLM output in such domains should be controlled or restricted in some way.
With regards to engineering design, it presently seems fanciful that an interaction with an LLM could result in the production of a real-world item. While through their interaction with models such as Dalle 3, LLMs can produce artwork that resembles real-world items, they cannot currently produce engineered designs. However, as noted below, in other areas, machine learning is being used to produce practical engineering designs with manufacturability in mind. Such generative design will only increase in capability and scope in the years ahead.
Manufacturing: The manufacturing sector has evolved towards digital manufacturing across which led to the emergence of concepts such as Manufacture as a Service. Presently it is possible to place orders for the manufacture of a wide variety of materials and components online. For example, there are many ‘on demand’ 3d printing services through which a user can upload their STL file, specify the material, and have the material shipped to their address. There are also some bio cloudlabs in which users can submit a genetic sequence to be synthesized. It is foreseeable that LLMs will be able to directly interact with such services in the future and through that manufacture real-world materials and components. This leads to question of how best to ensure LLMs cannot be used to inappropriately manufacture CBRN components and materials or other forms of weapons.
Stewart concludes:
This article is a first effort to address the question: will generative design lead to Weapons of Mass Destruction proliferation? Specifically, the article sought to identify specific areas where LLMs could contribute to proliferation identifying four pathways: knowledge retrieval, technical troubleshooting, mathematical, biological, or engineering design, and production of physical goods. In many ways, the article raises more questions than it answers. The nexus between generative design and CBRN proliferation is not yet understood.
