DOD's AI strategy leans on high-quality data

Positioning high-quality data as the foundation of the Department of Defense’s work to scale artificial intelligence technologies across its mission will require significant partnerships with the private sector, the head of the Pentagon's Chief Digital and Artificial Intelligence Office — or CDAO — said on Tuesday. 

During a keynote address at CDAO’s defense data and AI symposium, Chief Digital and Artificial Intelligence Officer Craig Martell said his office is trying to strike a delicate balance between providing tools for “the fight tonight” and long-term efforts to adopt interoperable technologies and more holistic automated solutions across DOD. 

In terms of developing AI solutions, Martell said DOD “should be paying industry to build models because they're always going to be on the cutting edge.” But he added that the department is pursuing an “AI scaffolding” approach, where it creates “everything on the left and the right of the model.”

While Martell said DOD could simply “contract out a stovepipe solution” that does not use high-quality data to address some of its immediate needs, he said that “data quality is the most important aspect” of the department’s long-term adoption of AI technologies — particularly when it comes to streamlining solutions across the various combatant commands. 

DOD’s data, analytics and AI adoption strategy — developed by the CDAO and publicly released last November — placed quality of data as the base of the department’s overarching “AI hierarchy of needs.”

DOD is still working to fully implement its combined joint all-domain command and control — or CJADC2 — initiative, which is focused on promoting interoperability between disparate military domains. Martell said proper utilization of high-quality data — which he defined as data that are “discoverable, accessible and available” — will help realize this effort.

“Imagine a world where those combatant commanders aren't getting that information via PowerPoints, aren't getting that information via email, or the turnaround time for situational awareness for what's going on shrinks from a day or two to 10 minutes,” Martell said, adding that making that quality data available will enable combatant commanders to, for example, “build an app on top of it very quickly.” 

To make this future a possibility, however, Martell said the Pentagon will need to rely even more on the private sector to drive change across the department — particularly when it comes to ensuring that quality data is being properly used as part of DOD’s adoption of data-mesh capabilities, intended to make information shareable across domains.

Martell said his interpretation of data mesh is that “the data stays where the data is,” adding that “the best place for the data is with the owners of the data.”

“This is where it's going to be really valuable for industry to help us out,” Martell added, noting that CDAO previously issued a request for information — or RFI — last year that asked for industry input on data mesh capabilities to underpin its CJADC2 effort. He said more RFIs are on the way in 2024, including one related to model monitoring. 

“We want to build this AI scaffolding in such a way that vendors can show up to play understanding the rules, and model builders in the department can come to us and say, ‘I need to build the following kind of model,’” Martell said. “And we have the playbook, which includes a list of vendors who know how to follow the rules or are abiding by the right rules. And we’re working very hard on this.”

Martell said CDAO is also working this year to, in part, publish its application programming interfaces — or APIs — for government data and for its federated data catalog to developers, as well as create opportunities for developers to sit side-by-side with DOD users to get direct feedback on their products.