Policy

Project Maven AI linked to surge in civilian casualties in Somalia

Airwars analysis finds civilian deaths tripled after Pentagon deployed computer vision targeting in 2017, with no evidence AI reduced harm.

Omega Editorial · November 18, 2026 · 3 min read

Key takeaways

  • Civilian deaths in Somalia more than tripled after Project Maven AI deployment in December 2017, rising from 25 to 79 minimum killed over comparable two-year periods.
  • U.S. strikes increased from 64 to 153 after Maven introduction, with the proportion causing civilian harm rising from one-in-seven to one-in-five.
  • Maven's computer vision algorithms showed detection accuracy below 30% in early deployments, sometimes dropping under 10% in new terrain.
  • Somalia was chosen as a Maven testing ground because military planners viewed it as lower-risk for "messing up" compared to Iraq, Syria, or Afghanistan.
  • The Pentagon has made Palantir's Maven Smart System a permanent program across U.S. military services despite no evidence it reduced civilian casualties.

The Pentagon's first large-scale deployment of artificial intelligence for military targeting coincided with a dramatic increase in civilian casualties, according to new analysis that challenges claims AI makes warfare more precise.

When Project Maven—the Defense Department's algorithmic warfare program—arrived at a U.S. airbase in Somalia in December 2017, it marked a watershed moment for AI in combat. The system used computer vision algorithms trained on thousands of hours of drone footage to detect potential targets in real-time video feeds. Within two years, both the pace of strikes and civilian harm had surged.

Airwars, a civilian casualty monitoring organization, documented 64 U.S. strikes in Somalia during the two years before Maven's introduction. In the two years after, that number jumped to 153. Civilian deaths rose even more sharply: from a minimum of 25 killed in 2016-2017 to at least 79 in the following two years—a more than threefold increase.

The proportion of strikes causing civilian casualties also climbed. Before Maven, roughly one in seven alleged U.S. strikes resulted in reported civilian deaths or injuries. After deployment, that ratio rose to more than one in five.

How Maven transformed targeting

Project Maven represented two fundamental shifts in military operations. First, it created a unified platform—initially built by Google, then Palantir after employee protests—to fuse data from drones, satellites, radar and other sources into a single decision-support system. Second, it deployed machine learning algorithms to automatically detect objects and potential targets in imagery.

The system's accuracy was notably poor in early deployments. According to reporting by Katrina Manson in her book "Project Maven," detection accuracy sometimes fell below 30 percent, and at times dropped under 10 percent in new terrain or challenging conditions. Somalia was deliberately chosen as a testing ground because, as one integration lead told Manson, it was "more of a free-for-all" where "the operational risks of messing up weren't as high."

The goal wasn't initially precision—it was volume. Maven aimed to find more targets and accelerate the targeting cycle, objectives it appears to have met.

Policy changes compounded AI impact

Other factors contributed to rising civilian harm alongside Maven's introduction. President Trump designated part of Somalia an "area of active hostilities" in 2017, loosening strike approval requirements. Command structures were reorganized to enable faster strike authorization, and private contractors became more extensively involved.

Yet the spike in strikes aligns precisely with Maven's stated purpose: identifying more targets faster. And the data provides no support for industry claims that AI reduces civilian casualties or improves targeting accuracy.

In one documented case from August 2018, less than a year after Maven's Somalia deployment, U.S. Africa Command struck a vehicle, killing four people it labeled "terrorists." Amnesty International later determined three civilians had died. The military acknowledged it didn't know the identities of those killed.

Why it matters

This analysis offers rare empirical evidence about AI's real-world impact on civilian harm, countering marketing narratives from defense contractors. As the U.S. military makes Palantir's Maven Smart System a permanent program across all services, and as autonomous weapons negotiations continue at the United Nations this week, the Somalia data suggests AI-accelerated targeting may be increasing—not decreasing—risks to civilians. With the current administration already launching nearly four times as many strikes in Somalia as its predecessor, understanding AI's role in civilian casualties has become urgent.

The findings were first reported by Airwars researcher Namir Shabibi, drawing on the organization's strike documentation and details from Manson's "Project Maven" book.

  • #project maven
  • #military ai
  • #civilian casualties
  • #autonomous weapons
  • #palantir
  • #computer vision

This is an original analysis by the Omega editorial team. Source reporting: AI Watch.

Want systems like this working for your business?

Book a Call

More in Policy