FT: 20 Soldiers Now Do the Work of 2,000 as AI Targeting Outpaces Human Checks

The US military can now generate targets fast enough to require an approval every 3.6 seconds, and the people meant to be authenticating those decisions do not have the infrastructure to keep up. That is the underlying picture in Financial Times reporting on how AI-assisted targeting has scaled inside US Central Command.

The scaling number comes from Scarlet Dragon, a Pentagon exercise series running since 2020. Inside it, 20 soldiers from the 18th Airborne Corps, using Palantir's Maven Smart System, handled a targeting workload once managed by 2,000 personnel during the 2003 invasion of Iraq. CENTCOM's stated objective is now 1,000 tactical decisions per hour, or one every 3.6 seconds, a benchmark CENTCOM head Brad Cooper has said the command reached.

Operation Epic Fury, the Iran campaign that began earlier in 2026, was the first at-scale run. US forces struck more than 3,000 targets in Iran with the Maven system helping pick them, and roughly 4,000 within the first four days. On the first day of strikes, a Tomahawk cruise missile hit the Shajareh Tayyebeh elementary school in Minab, in southern Iran; the death toll was later given at more than 168, over 100 of them children under twelve.

The named voices in the reporting separate speed from accuracy. Retired Air Force Lt. Gen. Jack Shanahan, the Pentagon's first AI chief, frames the underlying gap: "The Department of Defense was built as a hardware company in the industrial age." Emelia Probasco, senior fellow at Georgetown's Center for Security and Emerging Technology, is blunter on the review side: "That infrastructure is underinvested in now."

At the Tom Lantos Human Rights Commission hearing on September 16, reported by C4ISRNet, Carnegie Endowment senior fellow Steve Feldstein put the point plainly: "A bigger question is to what extent is the Pentagon collecting data and verifying the target accuracy of all these 1000s of strikes?" The Pentagon's civilian-casualty reporting rules do not currently require identifying whether AI contributed to a specific incident, which means the answer to Feldstein's question may not be reconstructable after the fact. This is the fourth military-AI alert we have tracked this week.