Before the Pentagon can hold artificial intelligence accountable, it will have to find it. Auditing the Department of Defense is a daunting task to begin with, and trying to figure out within that budget the full scale of existing spending on AI likely requires the audit as a prerequisite. At an Oct. 3 roundtable with Pentagon Comptroller David Norquist, defense officials discussed the massive scope of even finding all the AI programs in order to form a budget.

Asked how much the Pentagon was spending on AI, Deputy Defense Secretary Patrick Shanahan immediately moved to expand the category.

“You have to start with the question of how much are we spending on the cloud, right?” said Shanahan. “And then there’s the whole notion of how much are we spending on artificial intelligence.”

Cloud computing, which is to say the process of storing information used in one locating in servers in a different location and then accessing that information or software remotely sometimes, is a common ingredient of AI. Something like a customer service chatbot can work really well when run through the cloud, as machine learning algorithms train on the asked questions and iterate better answers, but the cloud itself is neither necessary nor inevitable to make functioning AI possible.

“The biggest effort underway right now that we have is to really make sure we can account for money that's being spent on the cloud,” continued Shanahan. “I can't give you a finite estimate on artificial intelligence. It's such a broad definition. The degree to which you and I think about, you know, real artificial intelligence, I don't have an estimate for that.”

Shanahan mentioned the difference in how AI is defined in apps, where it might just describe a chatbot with machine learning, and the more elusive “real artificial intelligence,” a term used by the Pentagon since at least 1983 to describe performing tasks with little human intervention, even autonomously.

Between the extremes is where most AI efforts fall, as some form of learning software that takes raw inputs from sensors and modifies it in some way to make the tasks of the human processing this information easier. Except for a few obvious parts of the budget, like the dedication Joint Artificial Intelligence Center, extracting the AI from the audit will mean getting into the weeds of every software acquisition and the ways in which that software modifies itself based on changing information.

However the audit turns out, one thing is certain: Whatever the Pentagon’s spending on AI, it’s much more than JAIC.