New Navy-Marine AI and data strategy coming this fall, seeking rapid adaptation for wartime

WASHINGTON — The US Navy is drowning in data, but it can’t turn that intelligence into new tactics or tech updates fast enough for a 21st century war, warned the new Chief Data and AI Officer for the Department of the Navy Stuart Wagner. But, he said, a forthcoming strategy for “data and AI weaponization” aims to accelerate all that.

How? In part, he said, by automating security classification procedures, and in some cases, bypassing the laborious Authority To Operate (ATO) process to get the right data to the right algorithms ASAP.

“We need to spin this really fast,” Wagner told the Hampton Roads chapter of AFCEA Tuesday. “Our strategy principally … is focused on expediting and making this easier.”

“A final draft of our data and AI weaponization strategy … is going to be signed out in the next month or so,” said Wagner, who took his post just this spring. It will lay out six broad goals, each with several subheadings: improving data management and infrastructure; rapidly turning AI pilots into operational systems; improving the DoN workforce’s AI and data skills; and increasing collaboration with outside partners like industry, academia, and foreign allies. But the common theme is that everything the Navy and Marine Corps aim to achieve with AI and data analytics is “informed by the speed at which we can learn and adapt.”

It’s “a digital OODA loop,” he said: an information-age version of the classic “Observe-Orient-Decide-Act” cycle originally developed by Col. John Boyd to analyze jet dogfights over Korea but later applied to all levels of conflict, from the tactical to the strategic.

And in 2025, that cycle moves faster than ever. “In Ukraine, we’re seeing this with the measure/countermeasure cycle, which is sub-24 hours,” Wagner said. That means one side will deploy some novel technology or tactic — perhaps a new way to deploy a drone, perhaps a technique to get radio communications through hostile jamming — only to have the adversary analyze the innovation, come up with a countermeasure, and deploy it in less than 24 hours, usually in the form of an overnight software update.

RELATED: Digitize or die: Ukraine’s war is a wake-up call for 20th century militaries

Rapidly analyzing data, using it to develop new tactics, techniques, and software updates, and applying those improvements to the current force is crucial to deter or, in the worst case, defeat a potential Chinese invasion of Taiwan in 2027, Wagner argued. (That’s the date by which Xi Jinping has famously ordered the People’s Liberation Army to be ready to subdue the island republic.)

“’27 is not far off. We’re not going to rebuild the fleet in a year,” Wagner said. “I suspect we’ll be able to move the needle more to prepare by refining our existing capabilities. … If that’s the case, then we should be investing into that and getting data off the ships and off our operational systems.”

The problem, he said, is that the Department of Defense is not currently set up to do that. While the military has sensors on manned aircraft, drones, surface ships, and other forward “platforms” that record tremendous amounts of data, much of it stays on those platforms and is never looked by either a human or an AI. Forward-deployed warships have limited communications capabilities to relay information to headquarters, Wagner said, often to the point that it’s faster to physically fly a hard drive back to the US than to transmit the data. Combat aircraft have limited storage, so in many cases the data collected on one flight is just recorded over on the next one, without ever being downloaded anywhere.

“We can’t learn and adapt to the adversary if their data is located in drives that we’re recording over,” Wagner said. “We were treating it like trash.” By contrast, his aim for the forthcoming strategy is to get data moving rapidly from the forces that collect it to high-powered AIs that can analyze it, helping human tacticians and software developers come up with new solutions quickly and then send them back out to the frontline.

That requires not only improving connectivity — something the Navy’s Project Overmatch and the all-service CJADC2 effort have worked hard on, Wagner says — but also automating key bottlenecks where information has to pass through slow-moving humans.

“Harnessing automatically collected data and AI to expedite learning and adaptation is table stakes,” he said. “It’s not going to work in real time if we have humans that are reading security classification guides and other policy documents that tell you the rules of sharing.”

Nor is it possible to simply feed all the official classification policies into a Large Language Model, then have the LLM either power a chatbot to advise the human classification officers or just make the decisions for them, Wagner warned, because there are too many different regulations, which can be too unclear, and often they conflict or contradict one another — all of which leads humans to reflexively mark everything as classified information just to be on the safe side. Instead, Wagner argues, the DoD needs to use older, more deterministic forms of artificial intelligence like “knowledge graphs” that follow strict, pre-programmed rules and therefore give the same answer to the same question every time. That, however, requires developing security classification policies that are clear and consistent in the first place.

In his prior job with the US Air Force, “I started a program called ‘Battering Ram,’ which is focused on increasingly automating security classification determinations,” Wagner said. “What we needed to do was basically develop non-contradictory policy which is clear to a human or machine, and then apply it directly on the data. … I know INDOPACOM and other combat commands are now looking really closely at this and starting to do pilots” of similar automated-classification approaches.

While automated security classification algorithms can speed up the flow of data, Wagner went on, the DoD also needs to work on the other end of the pipeline, where the data feeds AI algorithms and other analytic tools. Today, such software has to go through a cybersecurity vetting process and get an ATO before being deployed on DoD networks — a process that can take months. That’s not fast enough to get the latest and greatest software tools from the commercial world in the hands of military analysts trying to rapidly develop new software and tactics, Wagner argued: “If you’re waiting on an ATO to get started on learning and adapting, you’re not going to hit the sub-24 hour measure/counter-measure cycle.”

At the same time, it’s not safe to simply download unvetted software, let alone to military systems. So Wagner proposes a “sandbox” system, where the new tools are deployed only onto specialized, isolated networks, which should prevent any malware or bugs from affecting the rest of DoD or from transmitting classified data to outside parties. “I’m optimistic about the investments that we’re making into sandboxing right now,” Wagner said.