The Department of Defense’s Looming AI Winter
The Department of Defense is on a full-tilt sugar high about the potential for AI to secure America’s competitive edge over potential adversaries. AI does hold exciting possibilities. But an artificial AI winter looms for the department, potentially restraining it from joining the rest of the world in the embrace of an AI spring.
The department’s frenzy for AI is distracting it from underlying issues preventing operationalization of AI at scale. When these efforts fail to meet expectations, the sugar rush will collapse into despair. The resultant feedback loop will deprioritize and defund AI as a critical weapon system. This is known as an “AI winter,” and the Department of Defense has been here twice before. If it happens again, it won’t be because the technology wasn’t ready, but because the Department of Defense doesn’t know enough about AI, has allowed a bureaucracy to grow up between the people who will use AI and those developing it for them, and is trying to tack “AI-ready” components onto legacy systems on the cheap.
Previous AI winters arrived in the Department of Defense for their own peculiar reasons — immature technologies, overzealous regulation fixated on short-term results, and reality not living up to the hype cycle. This time, however, the core enabling technologies for AI are now widely available in the commercial sector. Computing power, cloud structures, data, and advanced software development are all readily available to anyone with the wherewithal to put them all together. The department’s looming AI winter will be unique, isolated, and of its own creation.
The Department of Defense is approaching the development of AI through the lens of traditional weapons development. But AI cannot be developed in the same manner as tanks and battleships — from the top down through exhaustive requirements lists, programs of record, and long development timelines before finally fielding the capability to the warfighter on the battlefield.
The department should invert this paradigm and partner with operational commanders — the leaders of those who will use or employ AI tools — from the beginning of the process, then iterate alongside real users to lead, create, and drive AI. If the department does not get this right, U.S. warfighters will be using badly designed and poorly integrated AI weapon systems, while near-peer AI-enabled adversaries seize competitive advantages.
The Seasons of AI
AI development is cyclical, bounding from periods of fanfare and funding to droughts of interest and investment. The former years are known as an “AI spring,” when rapid advancements in capability usher in funding and progress, leading to hype and buzz, which crash when the limits of technology or self-imposed restrictions lead to cynicism, fatigue, and a dearth of capital investment — an AI winter. This boom-and-bust cycle has occurred twice in the Department of Defense since World War II.
The First AI Cycle (1956-1973)
Coming out of World War II, a wave of excitement gathered around the potential of “thinking machines.” AI projects emerged from every corner of the Department of Defense — facial recognition, language translation, target detection. But these were primarily one-off toy projects. The core AI-enabling technologies weren’t yet mature, so AI applications were bespoke to particular uses. Without a concerted effort to build on the Department of Defense’s nascent research in technologies to enable AI at scale, these efforts were doomed to fail.
Just over a decade into the AI spring, the 1969 Mansfield Amendment prohibited the Department of Defense from funding any research without “a direct and apparent relationship to a specific military function.” The basic, fundamental research needed to mature core enabling technologies didn’t make the cut. The United Kingdom’s 1973 Lighthill Report highlighted the “pronounced feeling of disappointment” for AI research over the preceding 25 years.
The AI spring was essentially over by 1973. Research continued, but funding and excitement dried up. It would take a decade for the Department of Defense to take another shot at AI.
The Department of Defense’s Second AI Cycle (1983-1993)
In 1982, Japan launched a moonshot effort called the “Fifth Generation Project,” aimed at developing computers that could reason using AI. The race to achieve fifth-generation computing sent shockwaves throughout the Department of Defense. In October 1983, Defense Advanced Research Projects Agency launched the Strategic Computing Initiative, a 10-year effort to develop the core AI-enabling technologies to fuel future development. The AI spring was back in bloom.
Ideas for advanced AI gushed out of the military services: pilot’s associates to aid mission planning; battlefield management systems analyzing theater-level strategy; and autonomous land vehicles. In the midst of this excitement, the 1985 Packard Commission saddled AI advancement with programs of record not built with AI in mind. Strategic Computing Initiative projects were refocused to integrate with available technologies and systems for immediate, near-term impact instead of focusing on higher risk, basic research to develop the state of the art.
In 1993, the Strategic Computing Initiative closed up shop with little fanfare. AI again failed to live up to the hype, the Department of Defense reverted its focus to short-term initiatives, and another AI winter ensued.
The Four Horsemen of the Department of Defense’s Looming AI Winter
The four horsemen of the forthcoming AI apocalypse, so to speak, are this time around human problems without technological scapegoats. They are a lack of AI expertise, too much AI bureaucracy, the challenge of democratizing AI to the user level, and an old-fashioned approach to AI integration.
Domain Expertise: The Education Challenge
To become an AI force, the Department of Defense should inculcate AI domain expertise at appropriate levels of leadership — with the warfighters who employ AI, not bureaucrats. Creating a curriculum for AI within enlisted and officer professional military education should be a priority. The curriculum should include a breakdown of the core components of an AI pipeline — data, computing, algorithm development, test/evaluation, and AI-enabled platforms.
Commanders and their staff should have AI Smartcards, a rubric for exercising domain expertise, at the ready. First introduced in February 2021, the smartcard is structured into six main categories corresponding to the core functionality comprising an AI system. The smartcard is a framework for curriculum, both readily adaptable for professional military education learning modules and unit-level adoption. Not simply a technical guide, the AI Smartcard also helps the unit to think through operationalizing AI and becoming “AI-ready.”
An indicator that the scales are being rebalanced will be warfighters peppering AI subject matter experts with questions that get to the heart of the smartcard — is real AI being delivered? When slick briefings full of drive-by tech-splanations of AI capability evolve into a user asking penetrating questions to get the heart of AI performance, then the expertise obstacle will have been overcome.
Bureaucracy: The Management Challenge
Bureaucracy can be a beautiful thing — structure, talent, mission, and resources coming together in a cohesive, logical manner to deliver mission-critical capabilities. With that said, Department of Defense bureaucracies have a way of calcifying around programs and people, not actual capabilities. The consequence of the AI bureaucracy — which consists of several one-off groups organized around the task of bringing AI to the department, but which are stovepiped away from the people they’re actually trying to buy AI for — is an unnecessary buffer between users and developers. This convoluted system confuses what should be a direct feedback loop to ensure that capabilities are acutely focused on mission-critical requirements.
The department should shrink centralized groups like the Joint AI Center and the Army Futures Command AI Task Force and instead send their authorities, decision-making power, and resources down to operational units. AI centers of excellence should be creating processes, policies, and resourcing to facilitate a constant feedback loop between user and developer, with no-one in between to garble the message. They should exist as idea-centric organizations at the service level that cut across the warfighting functions, providing the tools, lessons learned, resources, and expertise to help commanders to operationalize AI. The bureaucracy should connect intent, constraints and restraints, and resources — and then get out of the way.
The Department of Defense’s managing philosophy with respect to AI bureaucracy should mirror the computer programming concept called the “self-deleting executable.” It is a string of code designed to allow a program to delete itself. Instead of thinking in terms of hiring hundreds of pseudo-experts, think in terms of creating an organization and then setting an egg timer for it to expand and contract, responsive to a direct user-developer feedback loop. The AI bureaucracy only survives as a platform to connect users with development talent, contracting, and computing power. It should shrink, not expand, over time.
Democratization: The Team Room Challenge
The Department of Defense can’t manage its way out of the looming AI winter, but instead should focus as much as possible on democratizing AI to let AI flowers bloom across warfighting formations.
For those bold enough to enter the team room — where planning turns to action — and ask battlefield operators, “What is one thing you can do better today than yesterday because of AI?” the responses are not encouraging. The Department of Defense should get AI to the warfighters on the ground who actually employ these capabilities. AI only works for the military when it’s unleashed against a real problem defined by warfighters, not bureaucrats. Project Maven, the department’s pioneering AI development program, employs a “field to learn” approach that ensures users drive capability development from the field, not the Pentagon.
Despite what the department’s leaders have been saying about AI over the past few years, how much excitement is expressed regarding AI’s potential impact, and how much money is being thrown at program after program, AI still hasn’t made it into the team room to empower actual warfighters.
How to fix this? Embed AI developers down to warfighting components to ensure that use cases are driven by users and firmly understood by developers. This will guarantee a tight feedback loop to ensure that AI capabilities are hyper-relevant to real, viable needs and that development doesn’t stop when an obscure, ill-conceived requirements list developed at a separate command is deemed “complete.”
Democratizing AI to the team room will help to ensure that the Department of Defense takes a sober, thoughtful approach to AI development. Instead of thinking about the next AI moonshot, the department should start thinking about cases where AI will actually increase the warfighting advantage. If the Department of Defense’s AI capabilities fail this time it will not be because of flaws with the actual technology, but because the use cases were defined in the halls of the Pentagon rather than in the team room by the actual servicemembers.
Integration: The Iceberg Challenge
The Department of Defense’s initial glimpse of AI led it to believe it could develop or buy AI on the cheap and integrate it with existing programs of record as a bolt-on. These assumptions were ill-conceived. For example, the Air Force experienced some initial success with integrating AI into the Distributed Common Ground System, but then realized the one-by-one “AI decision aids” are “single point demos” as opposed to AI at scale. AI worthy of near-peer competition is expensive, and should be integrated into systems built with AI in mind, from the ground up.
Since 9/11, the Department of Defense has been engaged in near-constant combat against violent extremist organizations across the globe. In this fight, warfighters achieved mission success almost in spite of the digital tools provided to them. Microsoft PowerPoint, not AI, could still win the day. Commanders and warfighters haven’t had the time or wherewithal to provide the feedback necessary to blow up the first-generation digital tools and software fielded by programs of record. Troops just compensated for the existing tool or purchased commercial off-the-shelf tools without going through the hassle of dealing with the bureaucracy.
The department’s development and acquisition machine — programs of record, program executive offices, requirements specialists, acquisition professionals — never got the message of how truly rotten and eroded the warfighter’s underlying digital foundation had become. Seeing the initial fielding of commercial AI to the warfighter, the same programs that have failed in the past now assume “we too can do this,” rolling out buzzwords like “AI-ready” or “AI-enabled” — and asking for more money to build their own AI on the cheap, anchoring their cost estimates and requirements on the same failed legacy programs. In this mass of confusion, it’s difficult to distinguish between “drive-by AI” and real, enduring capabilities. The Department of Defense is falling for the former. Drive-by AI is best characterized as one-off pilot projects without a robust AI pipeline — data, data labeling, computational power, algorithm development, test and evaluation — integrated on legacy tools that don’t allow users to continuously feed back existing and new use cases for the AI.
To address this challenge, defense policymakers should form a department-level, inter-service commission to review critical existing and legacy programs of record that will touch AI. Former Secretary of Defense Mark Esper, in remarks on the National Security Commission on AI, illuminated the scale required to seize the AI advantage — that in the long term, it will touch everything from existing maintenance to warfighting programs operating on legacy IT systems. The numbers around legacy IT systems in the U.S. government are shocking. The Government Accountability Office estimates that 80 percent of a $100 billion annual government-wide IT budget supports legacy tools, including a critical Department of Defense “maintenance system that supports wartime readiness, among other things.”
These programs should be evaluated with a critical eye to ensure that they adhere to best practices from the commercial sector, answer mission-critical needs, and can properly integrate with AI. If warfighters need it, then keep it, and start an egg timer to ensure the program is updated with AI in mind. If warfighters don’t need it, kill it. Secretary of Defense Lloyd Austin recognized this need in his initial guidance to the Department of Defense on March 4, 2021, telling the workforce, “Where necessary, we will divest of legacy systems and programs that no longer meet our security needs.” The battle to improve the department’s legacy tools could very well fall victim to the normal bureaucratic inertia or Congressional pushback. With that said, experience from the field tells us that these legacy platforms won’t survive integration with advanced AI capabilities and aren’t suited for a world of great-power competition.
Getting Out of Our Own Way
The Department of Defense is at a tipping point when it comes to AI. In 2025 we will either be hearing more platitudes from the department about how great AI could be or real substantive feedback from commanders on the battlefield about how great it actually is.
AI is commanders’ business — only with their perspective and leadership can the Department of Defense succeed in providing its front-line warfighters with the capabilities they need to remain technologically superior in future conflict. Only by involving commanders and those who will employ the AI in the face of the enemy can the department overcome the expertise gap and overabundance of bureaucracy that threaten to bring about another AI winter. The Department of Defense needs to get out of its own way and let warfighters lead the way to the AI spring.
The views expressed are those of the authors and do not reflect the official position of Duke University, the Department of the Army, or the Department of Defense.
Marc Losito is a first-year Master of Public Policy candidate at Duke University and an active-duty U.S. Army Warrant Officer, focusing on the intersection of technology and national security policy. Marc holds a degree from Norwich University and has served in the military and Special Operations Forces for 20 years specializing in counterterrorism, irregular warfare, and intelligence operations. You can find him on Twitter: @MarcAtDuke and on LinkedIn: www.linkedin.com/in/mlositompp22
John Anderson is a U.S. Army Reserve military officer focused on applying AI and machine learning to mission critical problems. He has served in the U.S. Army for nearly 20 years and previously worked in the financial industry as a senior research analyst at a value-oriented hedge fund and a mid-sized private equity firm. He holds degrees from the University of North Carolina at Chapel Hill and Columbia Business School. You can find him on LinkedIn: www.linkedin.com/in/john-anderson-a7a50738/
Image: Greg Vojtko
