The US military is accelerating its adoption of generative AI to support intelligence operations, with recent deployments highlighting both the potential and the risks of this shift. In a recent report by MIT Technology Review, it was revealed that Marines from the 15th Marine Expeditionary Unit used cutting-edge generative AI tools during a months-long deployment across the Pacific. These tools, developed by defense-tech firm Vannevar Labs, were employed to scan and interpret massive volumes of open-source intelligence—including foreign news articles, social media content, and sensor data.
According to the report, officers used large language models to rapidly translate and summarize information, saving considerable time compared to traditional manual methods. Captain Kristin Enzenauer, for example, used the tools to track sentiment in foreign news coverage of US military exercises, while Captain Will Lowdon leveraged AI to draft daily intelligence reports. The system, while not without hiccups—such as unreliable connectivity at sea—was hailed by officers as significantly more efficient.
Vannevar Labs, which recently secured a $99 million production contract from the Pentagon’s Defense Innovation Unit, uses a mix of proprietary and commercial language models, including some from OpenAI and Microsoft. The company claims to ingest terabytes of data across 180 countries daily, using that information to power AI systems that can translate, detect threats, and analyze political sentiment.
However, experts quoted in MIT Technology Review also voiced concern. Heidy Khlaaf, chief AI scientist at the AI Now Institute, warned that reliance on generative AI for critical decisions could be dangerous, particularly given the known limitations of large language models. She highlighted sentiment analysis as a particularly fraught application, arguing that such judgments are inherently subjective and prone to error—even for human analysts.
Chris Mouton, a senior engineer at RAND, echoed these concerns. In his evaluations of models like GPT-4, he found that while AI could help with many intelligence tasks, it struggled with subtle forms of propaganda and nuanced interpretation. Mouton stressed that the central issue is not whether AI can assist, but whether its outputs should be trusted in decisions involving real-world consequences.
With the Pentagon planning to spend $100 million over the next two years on generative AI pilots, the technology is poised to become a fixture in military operations. Yet as the MIT Technology Review article suggests, the challenge now is deciding where to draw the line between helpful automation and dangerous overreliance.
