Modern export-control systems are streaked with a brush of irony: The very civilian agencies that determine governmental approaches to exporting emerging technologies do not leverage these same technologies to perform key tasks. Rightfully, experts raise alarm at states and non-state actors that weaponize artificial intelligence (AI), yet they pay relatively little attention to how the government itself can use AI to shore up nonproliferation of sensitive technologies.
For years, both the military and intelligence communities have sought applications for emerging technologies suitable for their organizations. The Central Intelligence Agency has at least 137 artificial intelligence projects in the works, and the Department of Defense recently announced that up to $2 billion would be invested in AI research over the next five years. However, the US national security apparatus reaches far beyond the intelligence community and the Pentagon. A key question in years to come will be: How can civilian agencies in the national-security space leverage artificial intelligence to fortify security interests, with far fewer resources than their heavyweight military and intelligence counterparts? One answer comes from South Korea.
Countering proliferation. In the United States, using big data and emerging technologies to counter proliferation is not entirely new. Classified US military and intelligence programs already use AI to scour satellite imagery and other data in search of missile sites and signs of launch preparations in places such as North Korea and China. To date, AI-driven systems appear to focus on countering the proliferation of weapons of mass destruction, rather than a more holistic approach to protecting sensitive technologies. For instance, Gen. Raymond Thomas, head of US Special Operations Command, spoke earlier this year about using AI and deep machine learning to track weapons of mass destruction.
Blockchain technology, which offers a decentralized ledger for secure, fast, transparent transactions, has also been discussed as a means to bolster nuclear supply-side controls. Applied to nonproliferation, blockchain could use “smart” contracts to transparently verify information for controlled goods. AI could also be used to adjudicate arms control violations, as individuals attending a series of RAND Corporation workshops in 2017 proposed.
While these areas certainly merit further exploration, there are more immediate, less-explored lessons offered by South Korea’s applications of AI in strategic trade controls. This would not mean revising traditional multilateral export-control regimes such as the Missile Technology Control Regime or Wassenaar Arrangement, but rather strengthening how participating member states, such as the United States, implement the controls in their national systems. Focusing on supply-side management, AI should be examined as a key tool for states seeking more efficient, consistent export-control management and enforcement.
Intelligent export-control databases. In 2009, South Korea won contracts to develop four reactor units that constituted the inaugural United Arab Emirates civilian nuclear program and to build Jordan’s first research reactor, resulting in an unprecedented scale of nuclear exports for which South Korean export-control agencies were then unequipped. While the country had previously exported reactor components, supplying full-scale plants for export dramatically widened the scope of the Korean Institute of Nuclear Nonproliferation and Control (KINAC), which supports export controls for nuclear items from South Korea. In this context, KINAC developed the Intelligent Export Control Review System (IXCRS), building upon the existing Nuclear Export Promotion Service database.
In 2012, the institute’s officials began actively developing an AI system to manage the increasing number of documents related to validated technology transfers. They created IXCRS, a form of capacity building, to address insufficient manpower specializing in nuclear energy export controls. IXCRS maintains three subsystems that mine text to determine whether documents contain strategic materials; determine the similarity between documents; and use image-processing analysis to identify strategic technical specifications. Whereas humans may either lack know-how or make subjective decisions in case-by-case classification and license applications, these AI systems allow for greater comparisons between documents to facilitate a more objective, consistent application of South Korean export controls. This system still has some flaws, including language barriers and balkanized subsystems, but a marker of its success and utility to date is that South Korea intends to extend the system beyond pure classification issues to more sophisticated licensing-related responsibilities.
Although the Intelligent Export Control Review System is specific to nuclear trade controls on industrial exports of reactor units and components, the lessons it offers are far broader. First, larger countries with robust export-control systems, such as the United States, could use AI to help ensure consistency and optimize classification and licensing processes in a way that responds to exporters’ complaints—such as the bureaucratic burden associated with long wait times and a lack of transparency in license decision-making. Intelligent databases benefit from a larger number of records on file, so the accuracy and effectiveness would be even greater for a US system than is currently the case for the Korean system. Second, smaller countries could view AI-based systems as a capacity-building measure to achieve robustness in licensing and enforcement decisions. Thus far, two forms of AI stand out in their relevance to smart export-control management systems: image recognition and retrieval, and text mining with natural language processing.
Text mining and image processing tools. Building on a growing nuclear trade-control experience, AI could be leveraged to reinforce nonproliferation of controlled dual-use technologies and conventional weapons. IXCRS uses text-mining algorithms in its document retrieval and document clustering systems, allowing for deeper analysis than simple keyword searches. Once the algorithms match text to underlying concepts, they can recall similar results from the troves of export-control documentation. For instance, instead of relating “unmanned aerial vehicles” to similar keywords such as “unmanned aerial systems,” text mining could analyze the patterns around key terms to ascertain information vital to accurate classification—such as the payload, components, or accessories.
Similarly, AI could help fortify the new “fingerprinting technique” developed by the US National Nuclear Security Administration. This technique bundles similar dual-use commodities together based on technical information in the shipping documents, such as weight, rather than relying on the classification or code number. Noting, for instance, that pressure sensors for car tires and for uranium enrichment share the same code number, automating part of the process by looking for other technical specifications that are better indicators than code numbers can help make the process more efficient. Taking this a step further, text-mined fingerprinting could identify other such examples to take out the “low-hanging fruit” and allow human reviewers to spend more time on the more-sophisticated cases.
In the meantime, image-processing techniques can be used to analyze the similarity between diagrams and images that contain technical specifications of the material to be exported. Image-retrieval systems that analyze similarities (or “distance measures”) between images could also strengthen licensing procedures. Architects of IXCRS have argued that AI is useful for comparing graphics across documents. Small distances between images from different records—for instance, detecting whether laser designs have been modified with different military specifications than previous versions—could help classify exports in a consistent manner. AI systems that are already used to catch design plagiarism and other forms of trademark infringement could be used to scrutinize graphs, blueprints, or other images attached to license applications.
Although the Intelligent Export Control Review System , the only available benchmark today, is just beginning to expand from classification to licensing systems, it’s easy to imagine AI-based compliance systems in the future. AI pattern-recognition algorithms could add value for export enforcement. For example, the Strategic Intelligence Division within the US Department of Commerce is responsible for determining “potential diversion risks when making licensing decisions.” AI, along with graph databases to store information with more natural processing of linkages between objects, could complement human intelligence and improve predictions on matters such as: which controlled items are susceptible to in-country transfers that eventually put items in the hands of parties of concern; whether linkages between parties on the US Consolidated Screening List can be found through analysis of open-source or classified data, to better understand those parties and their access to sensitive technologies; and when illicit procurement of export-controlled goods can be traced to a network rather than a mere individual. Using both image processing and text mining, applying smart solutions to existing databases could result in the discovery of new related nodes and stronger enforcement of export-control regulations.
Future AI applications for evolving supply-side controls. Applied to export-control databases, document clustering could identify patterns undetected by humans. Clustering algorithms analyze how groupings of data occur organically, rather than how patterns are set by humans. When AI was applied to the strategy game Go, for example, a computer won against humans using moves described as “alien.” Because machines evaluate information differently than humans do, computers can identify non-intuitive patterns in data—and thereby strengthen counter-proliferation measures for strategic technologies. Using intelligent clustering models would mean complementing human intuition with unsupervised algorithms to interpret unforeseen connections between distinct data points. Whilet he Intelligent Export Control Review System does this to compare documents with one another (to varying degrees of success) for consistency’s sake, future intelligent export-control systems may use clustering more strategically.
AI could also be useful for classifying emerging technologies in a way that helps the United States maintain its technological edge at the limits of the export-control regime. Recently, for example, the Chinese police began using long-range surveillance cameras originally developed at Duke University with funding from the Defense Advanced Research Projects Agency (DARPA), but not subsequently selected for use by the US government. A Duke university spokesman said the university received clearance from the State Department to export the technology, yet the legal export of these cameras still brings to light a key concern: that uncontrolled items in early research and technology phases or with low readiness levels may slip through the cracks. Such cases of retroactively realizing that potentially malignant end-users wield sensitive US technologies show how the pace of technological advancements can still outstrip the rate at which governments can anticipate the discovery of new, potentially sensitive applications of cutting-edge technologies.
This incident was not an outlier: Research suggests that technologies developed at universities and private institutions, even with government grants or funding, often become classified at later stages once military applications become more evident. Consistent with trends of finding predictive applications for AI, algorithms may be better than humans at anticipating future uses for emerging technologies—such as preliminary research on synthetic biology in the agricultural sector, where sensitive applications may be illuminated only when they are in the public domain or have been transferred abroad to potentially adversarial actors. While far from a silver bullet, AI could analyze the troves of published research and publicly funded projects with low technology-readiness levels to better predict which items should be controlled in the future.
Implementing AI. While the rapid evolution and dissemination of technology today poses daunting challenges to the future of arms and export controls, technology may also be a partial remedy. Fighting fire with fire, technologies capable of tracking changes in real time—faster than any human—may offer a way forward if appropriately implemented.
The US government faces seemingly chronic difficulties keeping pace with commercially driven technological change, but the protection of national security interests—such as preventing the proliferation of sensitive technologies—could be prioritized. An artificially intelligent export-control management system would require a shift in recruitment to meet a greater need for data specialists, and implementation would require dedicating resources to training and, perhaps more importantly, cultural changes to build greater trust in “smart” systems.
Implementation would also require steps to ensure that this trust does not become unconditional. While automating tasks previously carried out by humans may be culturally challenging at first, the larger fear is what happens afterward. Subsequent dependence on AI may be dangerous precisely because of the “alien” reasoning that computers undertake and which they cannot explain to human handlers. Here the IXCRS precedent is again instructive, because the South Koreans have made efforts to demonstrate that the machine learning in that system is meant to supplement, rather than supplant, human decision-making. Particularly in the realm of national security, determining accountability would have to be a prerequisite to implementing any large-scale usage of AI.
Rather than injecting too much autonomy into the process, system architects could, for instance, define key steps in the sequencing of AI-driven tasks that require a human handler to sign off, thereby ensuring that the process remains auditable and that humans remain accountable. These key steps would differ depending on the export-control-related task involved—ranging from classification issues, to licensing decisions, to violation monitoring and adjudication. A side benefit of such procedures could also be curbing uninhibited trust in machines and allowing for gradual course-correction if the information or prescriptions from machines strike export-control experts as illogical or wrong.
Keeping humans in the loop would increase confidence in the validity of AI-driven conclusions. While the temptation to have AI control itself might eventually overwhelm human capacities to keep pace with rapid technological evolution, early steps should be gradual and not overly ambitious. Instead they could stick to modest goals of making export-control processes more efficient, more consistent, and safer.
The Bulletin elevates expert voices above the noise. But as an independent, nonprofit media organization, our operations depend on the support of readers like you. Help us continue to deliver quality journalism that holds leaders accountable. Your support of our work at any level is important. In return, we promise our coverage will be understandable, influential, vigilant, solution-oriented, and fair-minded. Together we can make a difference.