- US intelligence agencies and the military are developing AI programs.
- But they need specialized, secure systems to rein in sometimes-chaotic AI models.
- Microsoft has created an AI system cut off from the internet for US intelligence.
The US military and intelligence services are eager to harness the potential of AI, and firms are developing new technology to enable them.
While many industries can experiment with AI freely and use public tools, the high stakes and sensitivity of intelligence work and warfare represent a big barrier.
For companies that can keep the data safe enough and defend against the well-documented mistakes and hallucinations of AI models, a significant new market awaits. Its tasks are as varied as sorting through reams of National Security Agency intercepts for terror threats to guiding battlefield decisions in real time.
Firms like Microsoft have built walled-off AI products for the intelligence community, and Palantir has also staked out its ambitions. Similar efforts years ago created an uproar inside Google.
An emerging business
This month, a senior Pentagon official focused on AI, Radha Plumb, pointed to the small amount of classified computing power as a hurdle as the Pentagon prepares to carry out new tests, Defense One reported; Plumb has since stepped down.
As demand from defense and intel agencies grows, so should the business opportunity.
Officials hope that AI can supercharge tasks from analyzing swaths of secret data to battlefield targeting, an approach Israel’s Defense Force used in its withering aerial war on Hamas-led Gaza.
“The US is planning to integrate AI into a wide range of national security-related tasks,” said Ian Reynolds, a postdoctoral fellow for the Futures Lab at the Center for Strategic and International Studies.
He said the Pentagon had around 800 AI-related projects in the works, and was rolling out uses of the technology identified in a 2023 testing program called Project Lima.
“There are some indications that the technology is operational in some circumstances even today,” said Reynolds.
Defense One reported that the US military was trying to figure out how AI could help its leaders make decisions faster in a potential conflict with China with tests in the Pacific region.
“The idea is to quicken the decision-making process and achieve what the DoD is calling ‘decision advantage’, or the capacity to make faster, better decisions,” Reynolds said.
Among the Pentagon’s chief aims is to improve the flow of information within different parts of the military.
Not just the US, but nations including China and Gulf states are racing to dominate the new technology and experimenting with how it can be used by spies and the military.
Reynolds said that one of the core functions would be to analyze troves of classified data.
“I think the goal here is to get at the most critical data, information, or broader patterns across data, at a quicker rate than an analyst,” he said.
Power and danger
The dangers, though, are many and severe — classified data could accidentally drift into non-classified uses for an AI. It could leak or be stolen.
AI models could also display bias in ways that are difficult for humans to pick up on or could misunderstand nuances in communications reports, distorting the decision-making process.
“We are not fully sure of the degree to which human decision-makers may be nudged toward certain decision pathways by AI-enabled decision support systems,” Reynolds said.
And the secrecy of the programs being rolled out is another concern for critics.
Amos Toh, a senior counsel in the Liberty and National Security Program at the Brennan Center for Justice, told Business Insider that “the little we know about military uses of commercial AI indicates a real risk of exposing classified information to adversaries.”
“Using AI in intelligence analysis may also sweep up vast amounts of personal and sensitive data while amplifying discriminatory predictions about who poses a national security threat,” he added.
Microsoft in December said it had created a solution: a walled-off AI that could handle classified data safely.
It said it was the world’s first time a major AI model had operated wholly severed from the internet — signaling the start of a new kind of spy-friendly AI.
