DoD Has Embraced AI. Now What?

The U.S. Defense Department is starting to get its reps in with AI.

Thus far, the military’s embrace of this formidable new technology has been, for all its complexity and challenges, both measured and maturing. Safety has been a top focus, as Deputy Secretary Hicks underscored when releasing the strategy: “Safety is critical because unsafe systems are ineffective systems.”

The full promise of AI to empower organizations with greater efficiency, effectiveness, understanding – and enable faster decisions relative to our adversaries – will impact every process, from back office functions, to warfighting across all domains. We won’t get it right at first. This will be an iterative process from which we’ll have to learn as we go. So, in the spirit of relentless improvement, what are some of the foundational questions we should be thinking about as it applies to warfighting and the application of AI/ML?

Winning with AI requires deploying it at the edge because the battlefield will be non-contiguous and disconnected. Nearly all large language models (LLMs), open source and proprietary, are housed in the cloud, but expecting that operators will be able to keep their access in a denied or degraded network environment is not realistic against a peer adversary. We must therefore anticipate and prepare for the loss of reach-back capability to the cloud – and still have the warfighter be able to benefit from rapid insights and leverage AI as close to the fight as possible.

First responders have a saying about defibrillators and other lifesaving equipment: “If you don’t have it on you, you don’t have it at all.” This same logic holds for AI in combat. If it isn’t already on the warfighter – that is, capable of being deployed in a DDIL environment – then the operator is fighting without it. So how do we address this challenge?

One pathway is developing Small Language Models or SLMs. While trained on smaller datasets than LLMs, they offer efficient and effective performance for straightforward tasks and can be quicker to deploy and require less computational resources. Critically, they can operate independent of a network. Less expensive and tailored to discrete functions, they can operate on mobile devices at the edge without any interconnectivity. They require less storage space.

SLMs also require less power, a vital attribute as forthcoming power demands for AI continue to skyrocket. Consider: a generative AI platform like ChatGPT uses 10 times the energy for a search compared to a simple Google search. Looking ahead, Wells Fargo has calculated that by 2026, growth in AI power demand will rise by more than 8,000% from their 2024 projections.

For a military as reliant as we are on energy requirements, this makes SLMs a necessity. Just as we cannot assume network access in a conflict, neither should we assume a reliable or unlimited energy supply.

As former Secretary of Defense James Mattis was fond of saying, “Operations only move at the speed of trust.” 

Whether large or small, LLMs’ biggest barrier to operator trust has been inaccurate results provided by a generative AI tool. What’s the use of a model that can work in a DDIL environment if the operator cannot trust its outcomes? While high trust drives faster decisions, low trust put us back into more arcane processes that eat up precious time.

Models are only as good as the data fed to them, yet with retrieval augmented generation (RAG),  industry is beginning to witness AI software that mitigates hallucinations, capturing 99% or more of RAG errors produced before they reach the operator in some cases.

Higher trust drives operational speed at the tactical edge, and that’s crucial. But we are thinking about executing these capabilities at scale – in combat scenarios, in cases where these models will be used for targeting, albeit with a human in the loop. AI is not only at the edge; the models must be explainable to the commander, on whose shoulder’s responsibility will ultimately rest. Driving down hallucinations will be as essential at the edge as it is at headquarters.

It is no great insight to say the defense acquisitions process – writ large, but especially for AI and large language models – lags the private sector. Traditional procurement cycles take years from concept to deployment; private sector innovation timelines are measured in weeks.

But DoD knows this and has made progress as for example with the Army’s pursuit of flexible and modular contracting through the Software Acquisition Pathway announced earlier this year. This is a good step and reflects a broader awareness that there is a distinct difference between hardware and software acquisitions. It is also in line with recent arguments that have been advanced to treat AI models as a commodity, to “to think about models not as exquisite systems but as consumables.” There is merit here, and such market-placed solutions should be part of the defense enterprise’s collective maturation around AI. All agree the defense acquisitions process must be streamlined. I am optimistic we will get there. The global threat environment demands it.

We also need to be able to assess our return on investment. Defined metrics for evaluating the enhancements in productivity or efficiency as a result of AI investments is essential for DoD, Congress, and finally, the American people.

These questions have focused on near-term priorities: keeping access in degraded environments, trusting the tech while in those environments, and how to get that tech more quickly into the hands of warfighters at the edge.

This is crucial, but we must also understand that acquiring AI is not the one-time purchase of a “thing,” but rather an iterative process that will require continuous investment to keep pace as the technology evolves. This means sustained, committed investment into the capabilities themselves, the energy they require, and the on-going research and education of both military and civilian personnel who will be tasked with deploying them.

The U.S. Defense Department is starting to get the reps in with AI. We are building the muscle memory. It is incumbent upon us now to think through the next step of responsible deployment.

LTG Bob Ashley (USA, ret.)is former Director of Defense Intelligence Agency and an adviser to Primer AI.