Pentagon AI targeting claims clash with Gaza casualty data

Claims vs. reality in AI-assisted warfare

The U.S. military's adoption of artificial intelligence for targeting has been accompanied by consistent promises: AI will reduce civilian casualties through improved precision. But testimony before the Tom Lantos Human Rights Commission this week challenged that narrative with evidence from active conflict zones.

Palantir CEO Alex Karp argued in 2024 that technology could "drastically" reduce collateral damage and civilian deaths. The Pentagon has echoed similar claims as it deploys AI systems at scale — striking 13,000 targets in the first 38 days of one recent conflict using Palantir's Maven Smart System to identify targets and select weapons.

Yet experts told lawmakers the data doesn't support these assurances.

The accuracy paradox

Steven Feldstein, a senior fellow at the Carnegie Endowment for International Peace, disputed industry claims of 90% targeting accuracy, suggesting real-world performance may be closer to 50%. More fundamentally, he explained that even improved accuracy can increase total civilian harm when the volume of targets expands dramatically.

"Because the number of targets is so much vaster, the ultimate number of civilians harmed is likely far higher as well," Feldstein testified.

Deborah Brown of Human Rights Watch pointed to conflicts in Gaza, Iran, Yemen, and Ukraine as evidence that AI deployment hasn't limited civilian casualties. The Israel Defense Forces' Lavender system, which claims 90% accuracy, has been used to identify 37,000 Palestinian men as targets. Intelligence sources told the Guardian that pressure to generate more targets led to constantly shifting definitions of who qualified as a combatant.

Rules of engagement matter more than algorithms

Amanda Klasing from Amnesty International USA noted that accuracy becomes irrelevant when proportionality standards permit dropping 2,000-pound bombs on residential buildings. AI systems execute instructions based on human-determined rules of engagement — and those rules may not prioritize civilian protection.

"If a military decides that it's acceptable to have five civilians die for every suspect who is targeted, you're telling the machine what's acceptable," Feldstein said.

Reporting indicates the IDF approved strikes on low-level targets even when models predicted 15 to 20 civilian deaths as collateral damage. Questions have also emerged about whether AI played a role in a February 28 attack on a girls' school in Minab, Iran, that killed over 170 people.

Why it matters

As the Pentagon scales AI deployment in combat operations, the gap between vendor promises and battlefield outcomes raises fundamental questions about accountability and oversight. The military's Civilian Protection Center of Excellence is simultaneously cutting staff from 40 to just nine employees — reducing training capacity even as AI systems expand. Without transparent evidence that these technologies actually reduce harm, their rapid adoption represents an uncontrolled experiment with civilian lives at stake.

Transparency deficit

Brown highlighted a core problem: the Pentagon hasn't shared evidence supporting claims that AI mitigates civilian harm. "If this is the case that it's actually limiting civilian harm and being more precise and effective, then let's see the evidence," she testified.

The details were first reported by Responsible Statecraft, which covered the congressional hearing on AI use in military operations.