Editor’s Note: This article was submitted in response to the call for ideas issued by the co-chairs of the National Security Commission on Artificial Intelligence, Eric Schmidt and Robert Work. It addresses the second part of the second question on AI expertise and skill sets for the national security workforce.
The race to harness artificial intelligence for military dominance is on — and China might win. Whoever wins the AI race will secure critical technological advantages that allow them to shape global politics. The United States brings considerable strengths — an unparalleled university system, a culture of innovation, and the only military that bestrides the globe — to this contest. It’s also constrained by shortcomings. Washington’s most serious problem isn’t a shortage of ideas. It’s a shortage of talent. And this shortage is large enough to threaten national security.
While the current administration has publicly recognized the need to invest in AI talent, a senior defense official admitted that “finding, borrowing, begging and creating talent is a really big challenge for all of us.” Institutions like the Joint Artificial Intelligence Center and university research labs are central to the Pentagon’s development strategy; however, challenges ranging from data collection to refining operational concepts place huge burdens on existing technical talent.
These demands could be reduced by integrating our junior military officers and enlisted personnel as partners in the development process. Hiring junior leaders as product managers would accelerate technology development and build new operational capabilities while integrating user feedback. This immediately expands the number of personnel contributing to AI development efforts and grooms the next generation of leaders for the challenges of multi-domain operations.
This year I worked with data scientists from the University of Southern California to test the thesis that military personnel could be integrated into the AI development pipeline as product managers. We did this through a forecasting tournament based on security issues on the Korean Peninsula. This tournament created an opportunity to simultaneously experiment with machine learning technologies and expand civilian-military collaboration. The results provided new behavioral insights for the University of Southern California’s research team and refined a method for expanding national security AI research using existing military personnel.
The Imitation Game Problem
Our AI experiments explored how to deal with the daily flood of data that is used to provide key decision-makers with predictive analysis and enhanced situational awareness. We chose this problem for our first round of experiments because the challenge is so common and is only getting worse, with 2.5 quintillion bytes of additional data each day. We termed this the “Imitation Game problem,” honoring the challenge that confronted British cryptographers cracking the Nazi enigma code, who began with more potential solutions each day then could be tried in multiple lifetimes.
Traditional methods for mitigating overwhelming data processing requirements, like assigning more personnel, cannot keep pace with this challenge. This is especially true given military recruitment shortages. The consequences of missing key information or processing it too late are stark, as evident in the findings of the 9/11 Commission Report.
Building the Team
The experiments to circumvent the imitation game problem began after I spoke with Fred Morstatter from the University of Southern California’s Synergistic Anticipation of Geopolitical Events lab. Unlike traditional machine learning models that use only quantitative data sets to train algorithms, USC’s lab combines human judgement with quantitative models so that the strengths of both can optimize predictive value. This hybrid model addresses the military’s traditional aversion to replacing human decision-making with technology, captured in the saying that “humans are more important than hardware.”
Our pilot pursued improving commander decision-making through greater situational awareness using tools that combined human judgement and machine learning models. This approach can scale to a variety of defense challenges, though our initial experiment used public facing questions that were immediately relevant to our organization. Those questions became the basis for the Korean Security forecasting tournament we hosted with the University of Southern California’s lab in the spring and summer of 2019, which served as our first research sprint exploring the following:
Will North Korea dismantle the Yongbyon facilities or make other major progress toward nuclear dismantlement before July 25, 2019?
Will the United Nations lift or suspend any economic sanctions on the DPRK before July 25, 2019?
Will a peace treaty be signed between the United Nations, China, and North Korea to officially end the Korean War by July 25, 2019?
Will Chairman Kim Jong-un conduct a third summit with President Donald Trump before July 25, 2019?
Will President Moon Jae-in and Chairman Kim meet in Seoul before July 25, 2019?
Will the United States and the DPRK normalize diplomatic relations by exchanging envoys before July 25, 2019?
Will North Korea conduct a nuclear or intercontinental ballistic missile test before July 25, 2019?
What Did We Learn?
Solutions Require User-based Feedback Loops
When we separate technologists and military users, those who understand the problem cannot shape technology solutions, and those shaping technology solutions do not understand the problem. While there is a critical need to develop ties between researchers and operators, junior military personnel are generally removed from capability development efforts. This disconnect is largely due to the Army’s preference for institutional approaches to capability development that favor large commands and senior leaders, discounting the potential contributions of junior leaders. This bias is evident in the lack of billets for junior officer and enlisted personnel in Army Futures Command, despite their preponderance in the force.
While our individual experiment is valuable, the real impact will come from scaling our experimental design across the military. That is because using junior leaders as product managers mitigates the disconnection challenge and creates immediate value to both parties. We found that bringing current operational problems to the academic team diversified research applications while generating capabilities with immediate military relevance. This method also increases the interactions between research organizations and military innovators in a way that other models cannot replicate, expanding the idea sourcing funnel and increasing the odds that experimentation will lead to decisive capabilities.
This approach mitigates the current shortage of uniform-wearing AI talent that is the source of frequent Pentagon complaint. Our experiment shows that intelligent junior leaders can contribute to multi-functional teams in a “product manager” role. Technology companies use product managers to maximize the outcome value of products; servicemembers in this role can maximize an experiment’s value and operational relevance. Military product managers achieve this by turning force generated requirements into defined capabilities, managing requirement backlog, and liaising between their commands and technology development teams.
Silicon Valley companies rely on “non-technical” product managers to complement highly specialized professionals, and adopting that practice allows currently unused military personnel to achieve similar impact. While our initial experiments demonstrated the feasibility of this approach with comparatively minimal training, a second step is to train servicemembers in basic tech innovation practices. Product management and data science training will allow servicemembers to effectively contribute to military product development and increase the capabilities of America’s future force. This training is immediately accessible using resources like data science boot camps or online courses, and could be readily expanded through existing institutional partnerships.
Bringing in non-technical contributors to the project was valuable. Over the course of the tournament, forecaster accuracy improved (a development that speaks to the ability to rapidly train intelligence analysts to use these tools) and the best forecasters had the highest degrees of interaction with the system, accelerating algorithm training. The result was a virtuous cycle where the growing number of human forecasts enhanced the models’ predictive value while increasing user familiarity. The result provided USC researchers greater insight into behavioral patterns and optimization strategies for using their technology to inform future development efforts.
The post-product manager talent surge could expand the use of academic partnership programs like Hacking 4 Defense (H4D), since servicemembers could serve as problem sponsors for cross-functional academic teams. These teams could conduct problem curation and prototype development for AI initiatives and access senior mentors from the technology community through organizations like the Defense Entrepreneurs Forum. These research teams could report insights and progress to service-level AI organizations, simultaneously improving partnerships across the civilian-military AI ecosystem, training servicemembers in critical innovation skills, and closing capability gaps. The knowledge generated by these cross-functional academic teams could then be used to guide acquisitions efforts, including Small Business Innovation Research grants, forming an agile AI integration ecosystem.
The U.S. military could implement this strategy by launching programs through the Joint Artificial Intelligence Center or service-specific AI centers like the Army Artificial Intelligence Task Force that train innovative thinkers as product managers and junior data scientists. These leaders could then return to their host commands and sponsor operational problems through experimental pilots during initial concept development. After the efforts gain momentum, servicemembers could be mentored by experienced product managers and data scientists from startup partners to mature these capabilities. This would immediately create a Department of Defense talent development pipeline to meet the present shortage, while expanding the vibrancy of America’s AI ecosystem to regain its comparative advantage.
AI is Only as Useful as the Questions You Ask and the Data You Offer
AI demands specificity in asking questions, determining resolution criteria, and selecting training data sets. While AI is praised for its power and precision, those traits come with costs that must be included in experimental design.
These are acute challenges when AI confronts security arena complexity, as both problems and solutions are often ambiguous. We encountered this challenge as we iterated through crafting tournament questions to sufficient granularity to drive algorithm development. The danger of focusing too much on asking the questions the right way is failing to ask the right questions in the first place. Further, opportunity cost is incurred for every model launch, since pivoting to a second batch of questions often requires generating new data sets to train algorithms.
After crafting the right questions, our next hurdle was sourcing data sets for model training. This is difficult for security problems due to the limited number of existing data sets and event infrequency when trying to create one. For example, individual missile launches offer less robust data sets than commodity market data on sugar prices over the same period. A powerful strategy for overcoming this hurdle and developing more robust security algorithms is to generate proxy tabular data sets from currently underleveraged and unstructured data sources, i.e., “dark data.” Learning to deconstruct your operational environment into data sets allows for more rapid subsequent adaptation to environmental changes.
Our pilot accepted risk on optimizing questions and data sets by focusing on high value topics; even if more timely inquiries arose later, our effort was justified. Despite this hedge, we were confronted with surprises during the pilot. The DMZ visit between Chairman Kim, President Moon, and President Trump resolved several questions in spirit on June 30, but not according to the definition we wrote in April.
The pilot also allowed SAGE researchers to test how forecasters reason over different time horizons by deploying two identical sets of the approved questions, one with a resolution date of April 25, 2019 and the other using July 25, 2019. Preliminary findings indicate that the forecasters who engaged in both tended to have more conservative forecasts initially for the longer horizon questions, and more aggressive forecasts for the shorter ones. These observed predictive trends offer insights into underlying cognitive properties.
The Goal for AI Capabilities is Not More, but Better
Our goal in this pilot was to create valuable insights that could be integrated into operational rhythms of units across the Army. While the research crowd understands the value of AI systems, introducing this value to operational units required minimizing barriers to entry and reproducibility.
The goal of self-evident value creation led to aligning our research efforts to existing military tools designed to improve commander decision-making and awareness called priority information requirements. This critical information and signaling criteria allow leaders to understand when to use certain courses of action, allowing them to become proactive regarding decision-making. The benefit of building the AI experiments using priority information requirements was to ensure our model could scale since all Army units use these tools. This ubiquitous framework provides a natural focal point for incorporating and training other algorithms.
The next challenge is avoiding overloading existing digital infrastructure once tactical leaders understand the value of integrating these systems. An all too common, and toxic, paradigm in capability development is limitlessly expanding the tools assigned to commanders on the assumption that more is better. The result of this approach is adding yet another layer of technology on top of arcane digital infrastructure without considering existing systems. Users become overwhelmed by the number of systems they are expected to simultaneously manage, essentially nullifying the impact of new military technology.
Military product management is uniquely suited to prevent saturation of user cognitive bandwidth and optimize the value created while introducing new technologies. The goal should not be simply adding additional systems, but eliminating waste and simplifying tasks to increase organizational speed and agility. AI research efforts approached from this perspective benefit military leaders by creating data ecosystems that help units efficiently navigate complex operational environments.
The Next Iteration
Preserving an American-led international system requires achieving the technological superiority necessary for military dominance. A critical step in reaching that objective is closing the talent gap confronting America’s defense ecosystem by pivoting current strategy to include junior leaders. This pivot should integrate servicemembers as product managers and junior data scientists on cross-functional teams with academic institutions and tech sector volunteers, simultaneously mitigating manpower shortages and training our servicemembers to leverage these tools.
The United States has a history of making up for lost ground by combining the power of our private and public sector — from surpassing the Nazis with nuclear weapons to defeating the Soviets in the space race. It’s time to align tactical action with strategic priorities to ensure America wins the AI race. The United States can start today by bringing its tactical leaders into the fight for AI dominance.