NATO key to standardization---solves innovation
Andrea Gilli et al 20, Senior Researcher at the NATO Defense College where he works on issues related to technological change and military innovation. He has been visiting and postdoctoral fellow at Johns Hopkins University and Columbia University as well as Stanford University (where he remains an Affiliate) and Harvard University. Mauro Gilli is a Senior Researcher in Military Technology and International Security at the Center for Security Studies (CSS) at the Swiss Federal Institute of Technology, ETHZurich. Before joining CSS, he was a post-doctoral fellow at the Dickey Center, Dartmouth College. Ann-Sophie Leonard is a former Mercator Fellow on International Affairs, focusing on the intersection of international security and technology. She was a Visiting Fellow at the Research Division of the NATO Defense College in Rome from September to December 2019. Zoe Stanley-Lockman is an Associate Research Fellow in the Military Transformations Programme at the Institute of Defence and Strategic Studies at the S. Rajaratnam School of International Studies in Singapore. “NATO-Mation”: Strategies for Leading in the Age of Artificial Intelligence. NDC Research Paper No.15 – December 2020 //pipk
Over the past 150 years, the complexity of technology – whether military or commercial – has grown exponentially.311 The fact that increasingly capable technologies developed in different countries, in different years and by companies operating in different industries can work together and interact efficiently is, to a significant extent, a product of standards.312 Standards are, literally, documents defining technical specifications, requirements, guidelines or characteristics.313 For instance, all USB sticks enter any USB port and our laptops or smartphones can connect to any WiFi network: this is the product of standards. By the same token, standards are intensively used in the military domain to address similar compatibility issues, from the measurements of cartridges to communications. Standards deserve attention because they are also going to play a key role when it comes to AI, ML and BD.
There are, fundamentally, two types of standards. Technical standards refer to specificities, parameters and characteristics. Non-technical standards define procedures, processes and practices. To examine this issue in greater detail, it is useful to look more broadly at what standards do and what their merits are. First, standards not only connect technologies vertically (in the same field) and horizontally (across different fields), but also span different times of development. In this way, standards enable innovation. A related point is that common standards also expand the size of a market, as they eliminate nontariff barriers and thus contribute to guaranteeing economic efficiency and technological effectiveness.314 Second, standards define and protect intellectual property and thus make it possible to gather royalties: in this way, they also incentivize innovation.315 In the language of social science, standards are thus an institution that makes it possible to solve a collective action problem related to adjudication of rents in complex systems.316 Third, standards also make it possible to comply with specific requirements. For instance, technical standards may require certain levels of safety and security through the imposition of specific parameters. By the same token, non-technical standards may call for procedures and processes that comply with these or other demands – for instance, emphasizing transparency or calling for specific verification steps.317
The above discussion highlights how, despite their technical nature, standards have an important political, economic and ethical dimension. Standards, however, also have an important military dimension, as the case of NATO highlights. Historically, NATO has played a critical role in standardization. As Paul Beckley notes, “with nearly 6,000 documents in its database, NATO is the most prolific international military standardization organization in the world”.318 Furthermore, standards are the backbone of interoperability – which, in turn, is the lifeblood of a military alliance. NATO collective capacity is, however, more than the simple sum of its parts, or “having the same missiles or tanks”.319 The emergence of AI, ML and BD highlights this point. As this new set of technologies are integrated into the economy as well as in countries’ armed forces, standards will provide a fundamental contribution. NATO has a key role to play here, both in ensuring a swift and risk-free technological transition through standards and in infusing its ethical and moral commitments into the new technologies.320
At least four main issues are important for the Alliance. Standards and ethics. NATO cannot lead and shape ethical discussions about AI if is unable to translate its ethical commitments into technology. The propagation channel to achieve this goal is constituted by technical and non-technical standards. However, in recent years, as Paul Beckley warns, NATO’s once robust standardization programme has lost vitality: since the early 2000s, NATO has developed few new standards and, since 2010, most have been doctrinal rather than technical.321 Additionally, while AI technologies evolve quickly, the standardization process may often be lengthy, and given its consensus-based nature, it can require significant time in large organizations. Key ethical values such as explainability, traceability or transparency need to be translated into technical standards, in order to have real meaning and effects.322 Following previous discussions, given that ML blurs the line between research and development, a robust system of T&E/V&V is necessary, both at national and at Alliance level.323 Working together on these standards is crucial to ensure that Allies can plug together their capabilities and that the same parameters are adopted, both for combat effectiveness and for economic efficiency.
Approach to standards. AI does raise new challenges, but this does not mean that all new challenges require new solutions. Some standards, for instance, can be imported horizontally from nearby fields, like that of cyber security. This issue highlights the importance of the approach to standards, which NATO Allies may want to discuss: should new standards be pursued when it comes to AI, or should countries agree to use existing ones? Should they opt for stable standards, or try to update them regularly? These are important questions. For instance, the U.S. National Institute for Standards and Technology provides a list of features that the US should pursue for standardization – e.g., innovation-oriented, applicable across sectors and applications, human-centred.324 The Atlantic Alliance would benefit from a similar discussion as well as, eventually, trying to coordinate each Ally’s preferences, position and approaches towards standard multinational standard-setting forums, like Standard Development Organizations.
Standards and effectiveness. One of the main challenges of the interaction between algorithms and BD is their brittleness.325 Algorithms can be (easily) fooled. In the civilian economy, this may lead to controversial or even unacceptable outcomes, such as discrimination or bias.326 For instance, HR algorithms may discard some candidates because historical underrepresentation of their sex or ethnicity disadvantages them in the data on which the recommender algorithm is trained. In the military domain, this may lead to errors, like mistaking a military target for a civilian facility or the other way around, and overall combat ineffectiveness, or even defeat. The use of generative adversarial networks which intentionally exploit these vulnerabilities to undermine the opponent’s capabilities makes matters worse.327 To address these problems, AI and related technologies must comply with a set of principles that standards can translate and consolidate, including safety and reliability as well as accepted degrees of accuracy. Non-technical standards can also play an important role in defining procedures, processes and approaches in developing and fielding those systems, including about issues related to human control.328 However, standards should be also complemented by tools to enable the development of effective, reliable, robust and trustworthy AI technologies, in areas such as accountability and auditing, annotations and documentation, benchmarks, metrics, AI testbeds, methodologies and datasets in standardized formats.