Military vets patent hallucination-resistant, explainable AI technology

A technology startup founded by a group of retired U.S. military service members has received a patent for new tech that aims to bring explainability and hallucination resistance to artificial intelligence capabilities.

In receiving the patent, which was issued Tuesday by the U.S. Patent and Trademark Office, the leaders behind Data² hope to help open the black box of AI to inject more trust in the adoption and use of the technology, particularly in high-stakes mission sets like defense, intelligence and national security.

“These are organizations that, at the end of the day, have a low tolerance for failure, especially in the defense and in the intel space. Wrong answers have real consequences,” CEO Jon Brewton told DefenseScoop. “But what this means is they can finally build and deploy AI that they can trust and use to make mission-critical decisions. And I think that’s a real differentiator, because we’re not talking about just generic chatbot features. What we’re really talking about is building systems that are getting more reliable and have a systematic process for creating trustworthy outputs.”

Brewton, a former airman, founded Data² with Chris Rohrbach, a retired Navy SEAL commander, and Eric Costantini, a Marine, as well as Jeff Dalgliesh, with whom he worked in oil and gas, another industry that was key to the company’s inception. Additionally, the organization’s advisory team features members like former Federal CIO Suzette Kent and Nancy Morgan, a former chief data officer of the U.S. intelligence community. Early on in 2023, when the group was launching what would become Data², they took the idea through the Catalyst Accelerator program with the Air Force Research Lab and Space Force, and introduced a proof of concept associated with the intel community.

That subject matter expertise has been key to the company’s journey, particularly in understanding the gaps it is trying to fill for the services Brewton and his co-founders once served in.

“We rely on our subject matter expertise and our backgrounds to point us and narrow our focus around what we really pay attention to,” Brewton said, adding that they are targeting “high reliability industry applications” in sectors like defense, intelligence and finance.

He continued: “The future we’re trying to build is one where a military commander on the field can trust the AI-generated battlefield assessment, so cybersecurity analysts can rely on AI to respond to threats in near real time. We’re really trying to make AI trustworthy for governments, so that they can use it in those mission-critical areas, and so that people and machines can really start to make better decisions together… [and] so that they can start generating value from the data that they already have in this technology and capability.”

Brewton explained that the patented capability was built to be technology agnostic and can work with any AI model or ecosystem, because it uses “knowledge graph” tech on top of an organization’s existing data, “grounding the AI in a fact base” so that “every AI-generated answer is anchored in verifiable, traceable source data with citations down to the data record level.”

Despite major progress, today’s foundational AI models still have major limitations that cause hallucinations and are often unreliable, which is a concern for defense officials.

For that reason, Brewton said, Data² took a data-focused approach to “solve for the lack of sort of trust, transparency and explainability.”

“Ultimately, what we found out very, very early on in the process, which informed the patent that we developed, is that better data architecture, better data structure, is really the key to unlock how you can grow in explainability, traceability and transparency — not better models,” he said.

Currently, the business approach of Data² is to partner with larger technology vendors, like Amazon, Microsoft, Dell and Nvidia, with large footprints through which it can offer its services in the broadest way possible with some “brand equity” and association with “some of the industry’s most trusted technology partners, top-tier technology providers to the U.S. government at scale.”

“It’s not limited to a data architecture, it’s not limited to a data type, it’s not limited to a large language model, it’s not limited to an environment,” Brewton said. “We’ve tested fully cloud-hosted environments, all the way down to a completely air-gapped deployed edge kit.”

While the plan moving forward is to continue to lean into that partner ecosystem to make headway with defense and intelligence activities, Brewton said, acquiring the patent “really substantiates what we believed all along was a really transformational approach to how to scale AI, especially in these mission-critical spaces.”

“Our patent really represents sort of a first step towards making AI usable by governments and other highly regulated organizations,” he added.