Department of Defense Adoption of Generative Artificial Intelligence

---

Executive Summary

The United States has opened a critical technological advantage over the rest of the world with the recent advances in generative artificial intelligence (GenAI). This moment, which may not last, presents the Department of Defense (DoD) with a crucial opportunity to accelerate two of its most significant transformations – preparing for the future character of conflict and strengthening our military overmatch against our rivals, especially the People’s Republic of China (PRC).    

1. Enabling Decisional Advantage. Presently, the most promising aspect of generative AI models is as a decision aid, or what we would term a cognitive co-pilot. The Department should identify both operation centers and weapons platforms where it can immediately and securely integrate generative AI models to enable better informed and faster decision making, including in the oversight and conduct of operations.     

2. Enhancing Operations. The Department of Defense is already exploring the use of AI for predictive maintenance² and select back-office support functions.  The next areas that could benefit from AI-enabled models are (1) logistics and sustainment; (2) investment and divestment decisions; (3) experimental course of action generation; and (4) global force deployment management decisions, particularly of high demand, low inventory capabilities.  

3. Developing the talent. Much like generative AI will impact labor productivity and create new occupations in the private sector, it will most certainly affect current military occupational series and create demand for new skill sets. The Department of Defense needs to put in place policies that recruit, develop, and retain talent that can develop mission driven requirements for, evaluate, build, and – crucially – employ AI tools for military effects. 

4. Identifying new defenses. Generative AI will most certainly accentuate threats against our nation from disinformation, in the cyber domain, and from chemical and biological agents. Rival militaries will also adopt generative AI. The Department of Defense should identify what threats can be reasonably expected to arise, and begin to plan for additional defensive measures that need to be undertaken, including to counter AI generated attacks by our adversaries.  

2. Build an Automated Orchestration Platform, a generative AI-powered interface that can call up relevant tools and datasets, is able to receive user prompts, can decompose the queries into discrete tasks, and semi-autonomously or autonomously complete tasks.  

3. Develop defense-tailored generative AI models, trained on specific military terms and jargon, secured from revealing the data they were trained on, the impact of tuning, or the content of prompts and their outputs, inspectable and instrumented by cleared researchers, and broadly accessible to cleared personnel, including for operational purposes.  

Finally, progress in generative AI will most certainly accelerate three trends in warfare – the imperative of human-machine collaboration and teaming, the criticality of software advantage, and the necessity of empowering warfighters at the edge so that they can conduct distributed, network-based operations. These three trends, as SCSP has previously argued in our proposed Offset-X competitive strategy for the Department of Defense,³ will be among the key determining factors of U.S. military advantage, particularly vis-a-vis the PRC.  

Strategic Context

Much like the commercially-driven telegraph, steam engine, and railroad changed the way wars were fought at the dawn of last century, the advances in AI, automation, and unmanned platforms are driving some of the most fundamental changes in the character of warfare in this century. Nowhere is this more visible than in Ukraine.⁴ Ubiquitous sensors, big data processing, automated orchestration platforms, and drones employed at scale are dramatically compressing detection-to-destruction timelines, and in combination with a digital levee en masse are introducing new ways of warfighting.

If these changes were to occur during peacetime or with military overmatch assured, the United States military would still take notice. But they are occurring in the midst of an intensifying geopolitical competition with an increasingly belligerent PRC and on a timeline where there is a real risk that the PRC may decide to invade Taiwan, making the changes in the character of conflict even more alarming and the need for adaptation and adoption more urgent. Moreover, China – the only nation with the intent and latent capacity to compete with the United States⁵ – has already articulated a vision and is preparing for the changing character of warfare with the development of new concepts and supporting capabilities. Its military, the People’s Liberation Army (PLA), is fusing mechanization, informatization, and “intelligentization”⁶ into a new way of war purpose-built to overcome and leapfrog U.S. military capabilities. AI plays an essential role in the PLA’s ambitions, particularly in its objective to “intelligentize” its armed forces.⁷ Specifically, the PLA has set its sights on leveraging AI and big data to exploit U.S. vulnerabilities, use influence operations to prevent U.S. military leaders from understanding their environment, and employ swarm attacks to overwhelm U.S. military’s defenses.⁸

While the changes in the character of conflict and China’s ambitions are concerning, the recent and rapid advances in generative AI by American companies are encouraging. They have opened up a window of opportunity for the United States to re-demonstrate its innovation power, and strengthen its military might. As this moment may not last, however, the Department of Defense needs to significantly enhance the pace, scope, and cohesion of the efforts it has undertaken since 2015 to respond to both geopolitical and technological changes.⁹ To be clear, generative AI is but one aspect of the enhancements; SCSP has previously provided a comprehensive blueprint – Offset-X – for a new competitive strategy that would both close any deterrence gaps in the near-term and position the United States for a competition that is likely to be enduring.¹⁰ 

Defense Applications of Generative AI

Even in its current nascent stage, generative AI is already powerful enough to help military personnel access faster and more data and tools; generate text and initial recommendations for memos, plans, orders, and information campaigns; and prepare elements of planning for legal and operational review. As generative AI models become increasingly sophisticated, they will further help synthesize and display data from different datasets and multiple modalities. Finely-tuned models, such as foundation models for multimodal geospatial intelligence,¹¹ could deliver advanced battlespace awareness, including visualizations of the battlespace and provide commanders with a detailed and updated common operational picture. Generative AI could also conduct real-time fusion, correlation, and pattern analysis of massive volumes of data and adaptively orchestrate novel attacks or defenses with minimal human intervention. In short, generative AI holds the promise of delivering in the near-term a significant information advantage for the U.S. military.

But the potential is much broader. With the Department of Defense having now stood up Task Force Lima to explore and integrate generative AI more broadly across DoD, there are at least four critical areas that are ripe for transformation. 

1. Enabling Decisional Advantage. Presently, the most promising aspect of generative AI models is as a decision aid, or what we would term a cognitive co-pilot. Current models help – imperfectly to be sure – synthetize and summarize large volumes of information, answer broad and specific questions, compose initial takes, and provide recommendations. These are all inputs that can augment human cognition and enable better informed and faster decision making, as well as creative and unexpected courses of action. Additionally, there are three specific mission areas that current AI models can significantly enhance. The first of these is with indications and warning (I&W) which is customarily the mission of the Directorate of Intelligence (J2).¹² AI models are particularly effective at performing pattern analysis across vast volumes of data in order to flag anomalies. While current models are still descriptive and not predictive, identifying deviations can still greatly assist human analysts. Second, generative AI models could prove useful tools in operation centers, such as the National Military Command Center at the Pentagon, or in the Combined Air Operations Center in Qatar. Layered on appropriate data sources and streams of information, AI models can help with alerting, planning, monitoring, or simply answering questions. Lastly, further refined AI models could help improve the operational survivability of platforms and our personnel operating them. They can inform human operators in planning infiltration and exfiltration routes, and assist in pre-empting detection, discovery, and destruction – an increasingly challenging task.

2. Enhancing Operations. The Department of Defense is already exploring the use of AI for predictive maintenance and select back-office support functions.¹³ However, additional missions could benefit from AI models. One mission is logistics and sustainment planning. Given the complexity of logistics planning and the reliance on private carriers, AI models can aid in breaking down the various tasks, help formulate or create parallel plans, and assist in anticipating sustainment demands. Additional DoD functions that would benefit from AI models are Comptroller in budget planning and auditing, and Cost Assessment and Program Evaluation office in resource allocation decisions. Generative AI models can also help with generating initial courses of action, not just for military operations, but also for policy actions and campaigning options. The current models may not come up with new and original recommendations, but — at a minimum — they can help planners and action officers capture a broader menu of possible options. Finally, they can help senior leaders with global force deployment management decisions, weighing the potential impact of deploying high value force packages to global hot spots, and adjudicating competing demands among regional commands that invariably exceed available forces, while avoiding burning down force readiness with low consequence deployments.  

4. Identifying New Defensive Areas. Generative AI will most certainly accentuate threats against our nation from disinformation, in the cyber domain, and from chemical and biological agents. These threats will also specifically target our military personnel and operations. Therefore, DoD should identify what additional threats and vectors of attack can be reasonably expected to arise, and begin to plan for additional defensive measures that need to be undertaken. These measures need to account for not only AI generated attacks by our adversaries, but also for attacks by our adversaries against our AI-enabled systems.  

Concrete Steps

1. Establish a Defense Experimentation Unit (DEU) to experiment with and iterate on AI models for DoD.
   DEU’s mission would be to advance the use of generative AI and other key technologies in the DoD through an iterative and continuous process of experimentation, learning, and development. To accomplish this, the DEU would build a “sandbox environment” for generative AI experimentation and concept development that is accessible DoD-wide; develop new generative AI capabilities for use in the sandbox environment; emphasize partnerships with front-line units, field-to-learn processes, rapid iteration, and concept exploration; and establish feedback mechanisms to quickly and consistently share lessons learned from experiments to the entire DoD. DEU could augment or partner with the Defense Innovation Unit to identify and acquire new capabilities, or augment and leverage the infrastructure and standards established by the CDAO.

2. Develop an Automated Orchestration Platform.
   This platform would be a generative AI-powered interface that would serve as an intermediary between users and the vast suites of tools and databases available in many military environments. The platform would receive user prompts, decompose the queries into discrete tasks, and assist users in executing them by calling up relevant tools and datasets, and semi-autonomously or autonomously complete tasks. This process greatly reduces the time and enterprise knowledge needed to interact with large or complicated datasets, accelerating both decision-making and operations. The DoD should partner with industry to develop a generative AI-enabled orchestration platform and an accompanying suite of tools, with the intent to further develop and update this platform as increasingly powerful generative AI models continue to be developed, either by industry or the government. 

3. Develop defense-tailored generative AI models.
   While it would be difficult for the U.S. government to develop a leading edge foundation model, it is possible for DoD to acquire or develop models based on the technology previously developed by OpenAI, Google, Anthropic, or Meta. This would allow DoD to spend far fewer resources, while still reaping many of the benefits of generative AI. A defense-tailored model would allow the military to fine-tune and prompt the model without exporting information to an uncleared commercial model. Once developed, the model could be used for experimentation or operational employment, and could facilitate the development of an automated orchestration platform. 

An Important Consideration

The rapid pace of generative AI development will most certainly exacerbate the challenge of maintaining interoperability and interchangeability between the U.S. military and its allies. Variations in access to and levels of trust in generative AI models across allies will lead to barriers to coalition-wide implementation of generative AI in military systems and operations.¹⁸ The Department of Defense should start working immediately with our allies to coordinate the pursuit of generative AI models, or to begin offering U.S. models to allies that cannot otherwise access them. In the process, the DoD and allies need to establish international standards to ensure the collective security and integrity of generative AI models, and coordinate technology solutions to address the risk of compromise.¹⁹ At a fundamental level, the DoD also needs to work with its counterparts and political leaders in NATO and other alliances to address public distrust to the employment of generative AI by engaging in public education and discourse on the uses of generative AI in defense.

---

Endnotes

1. DOD Announces Establishment of Generative AI Task Force, U.S. Department of Defense (2023)
2. GAO-23-105556, Military Readiness: Actions Needed to Further Implement Predictive Maintenance on Weapon Systems, U.S. Government Accountability Office (2022). 
3. The Future of Conflict and the New Requirements of Defense, Special Competitive Studies Project (2022); Offset-X: Closing the Deterrence Gap and Building the Future Joint Force, Special Competitive Studies Project (2023).
4. Shashank Joshi, The War in Ukraine Shows How Technology is Changing the Battlefield, The Economist (2023).
5. National Security Strategy, The White House at 8 (2022).
6. Military and Security Developments Involving the People’s Republic of China, U.S. Department of Defense (2022).
7. Xie Kai, et al., A Perspective on the Evolution of the Winning Mechanism of Intelligent Warfare, China Military Network – PLA Daily (2022).
8. Koichiro Takagi, The Future of China’s Cognitive Warfare: Lessons from the War in Ukraine, War on the Rocks (2022).
9. Gian Gentile, et al., A History of the Third Offset, 2014-2018, RAND Corporation at 42-45 (2021). 
10. The Future of Conflict and the New Requirements of Defense, Special Competitive Studies Project (2022); Offset-X: Closing the Deterrence Gap and Building the Future Joint Force, Special Competitive Studies Project (2023).
11. Gengchen Mai, et al., Towards a Foundation Model for Geospatial Artificial Intelligence (Vision Paper) SIGSPATIAL (2023).
12. J2 Joint Staff Intelligence, U.S. Joint Chiefs of Staff (last accessed 2023).
13. GAO-23-105556, Military Readiness: Actions Needed to Further Implement Predictive Maintenance on Weapon Systems, U.S. Government Accountability Office (2022). 
14. Interviews with DoD personnel responsible for the development of technology solutions.
15. Final Report, National Security Commission on Artificial Intelligence at 373-374 (2021).
16. Interim Report, National Security Commission on Artificial Intelligence at 61-65 (2019).
17. SCSP and the National Security Commission on AI have recommended including problem curation, the AI lifecycle, data collection and management, probabilistic reasoning and data visualization, and data-informed decision-making as core tasks for effective human-machine interaction. This recommendation would add prompt engineering to this list based on recent trends in generative AI.
18. Final Report, National Security Commission on Artificial Intelligence at 82 (2021).
19. Final Report, National Security Commission on Artificial Intelligence at 560 (2021).