A military decision-support tool, powered by artificial intelligence, built largely by Palantir and used by America's armed forces, including CENTCOM and NATO. Maven takes information from open sources, such as social-media feeds, and classified sources, such as satellites, fusing them together to generate targets, recommend optimal weapons for each strike and assess damage afterwards. It also serves as a "digital twin" of the real world, allowing commanders to simulate how a particular decision might play out. Arnel David, the NATO officer in charge of the programme, says the aim is to turn military command into a "machine-aided, predictive science".
Joe O'Callaghan, a retired colonel who led the development of Maven at the US Army's XVIII Airborne Corps, said one classified study showed how Palantir's system allowed military staff to plan an operation on the scale of the Iraq war with one-tenth of the manpower. A former NATO general involved with Maven said what previously took dozens of people tens of hours "could be boiled down to two minutes." A European general described what he witnessed as "alchemy": "We are moving from ten targets a day to 300. The aspiration is 3,000 a day."
Maven has been employed in aid of Ukraine from 2022. Anthropic's Claude model is used to some extent inside Maven but not for geospatial tasks like identifying objects.
In both the American and Israeli armed forces, humans approve each target, outside extreme circumstances such as air-defence systems engaging incoming projectiles. But the increasing scale and tempo of strikes has created incentives to give computers greater latitude in firing on the targets they have generated.
Maven was created by Drew Cukor, a US Marine, as a project to find objects in drone footage. Early algorithms tested in Somalia in 2017 mislabelled clouds as flying school buses; in Afghanistan the next year they identified trees as people. Google walked out in 2018 after employee protests; Palantir later became the central contractor. Maven was used during the 2019 operation to kill Abu Bakr al-Baghdadi in Syria and the 2020 drone strike on Qassem Suleimani. Models with 70% success rates in Afghanistan dropped to 30% in the Philippines, where targets walked against jungle rather than dust. Training a single algorithm typically requires 10,000 accurately labelled images. Maven's targeting in Ukraine reportedly cost America $1m a month in cloud bills. Long after deployment in Ukraine, Maven still produced about ten incorrect detections per square kilometre assessed. Two new weapons built around Maven algorithms are "Goalkeeper", a loitering munition, and "Whiplash", an explosives-laden jet ski whose early versions were smuggled into Ukraine by the CIA. Katrina Manson's 2026 book Project Maven documents the programme.
A closed mouth gathers no foot.