DOD to convene conference on generative AI amid concerns about ‘hallucinations’
The Department of Defense will host a conference in June to look at ways that the U.S. military can leverage generative artificial intelligence for “decision support and superiority.” But the Pentagon is well aware of the technology’s current shortcomings, according to Deputy CTO for Critical Technologies Maynard Holliday.
Generative AI has made headlines recently with the emergence of ChatGPT and other capabilities that can generate new content, such as text or images, based on their training, which can include the use of large language models.
The upcoming DOD meeting, which is being spearheaded by Undersecretary of Defense for Research and Engineering Heidi Shyu, has been scheduled as the Pentagon has both hopes and concerns about the technology.
“On the trusted AI and autonomy front, we are working with companies that are developing these large language models. And we recognize that, as all you should recognize that these things don’t have any common sense and they are word completion programs that use statistics to complete thoughts, you know, based on prompts,” Holliday said at the Potomac Officers Club’s annual R&D Summit on Thursday.
The Pentagon is concerned about the “hallucinations” associated with this type of technology, or the generation of misinformation.
“It still hallucinates greatly and it does not have a corpus of defense-specific information that would give us decision advantage yet. And so, you know, we recognize that we’re going to have to develop that corpus of data so that these large language models can be useful — and then to minimize, you know, the hallucinatory side effects of what these logic language models do,” Holliday said.
Leaders of OpenAI, which created ChatGPT, and others in industry have acknowledged the hallucination issues even though they see great promise in the systems’ capabilities.
Holliday declined to provide more specific information about who will be attending the upcoming DOD conference and what the expected outcomes are, noting that it’s Shyu’s initiative.
But more broadly, “DARPA has an advisory group, of which I talked to some of the members last week when I was in Northern California, to kind of demystify and debunk … some of the things about these large language models that I talked to you about earlier. And so, again, we want to get at the crux of what is necessary to develop these to leverage for defense,” Holliday told DefenseScoop.
Updated on March 28, 2023, at 4:15 PM: On March 28, a DOD spokesperson told DefenseScoop that “we are still in the early stages of planning” but the AI conference is scheduled for June 20-22 in McLean, Virginia. “Attendee list, session chairs, agenda, etc. are all being worked,” the spokesperson said.
