This isn't sci-fi, AI is already killing people

Experts told a Tom Lantos Human Rights Commission hearing on Wednesday that the Pentagon's military use of AI is not reducing civilian harm, contradicting claims by Palantir CEO Alex Karp and other proponents. Carnegie Endowment senior fellow Steven Feldstein said the real targeting accuracy of AI systems may be closer to 50% rather than the 90% often cited, and Human Rights Watch deputy director Deborah Brown said evidence from Gaza, Iran, and elsewhere does not show AI limiting harm given high civilian death tolls. The Pentagon struck 13,000 targets in the first 38 days of a war with Iran using Palantir's Maven Smart System, and lawmakers have questioned whether AI played a role in the February 28 attack on a girls' elementary school in Minab, Iran, that killed over 170 people.

Palantir CEO Alex Karp proclaimed https://breakingdefense.com/2024/05/mark-milley-civilian-death-ai-technology-alex-karp-palantir/ on a panel in 2024 that “technology should be able to reduce drastically the amount of collateral damage, civilian deaths.” Former Google CEO Eric Schmidt and former Deputy Director of the CIA David Cohen agreed. Proponents of artificial intelligence use in the military have long made some version of this argument — that the military’s adoption of AI will lead to a reduction in civilian harm. Rep. Jim McGovern D-Mass. summarized the pitch during a congressional hearing about AI use in the military on Wednesday. “Our government has argued that AI could improve assessments of the likely effects of military operations and reduce the risk of civilian casualties,” McGovern said. The question is no longer theoretical, as the U.S. military is already deploying AI. The Pentagon struck https://www.armscontrol.org/act/2026-05/news/ai-plays-major-role-war-iran 13,000 targets in just the first 38 days of war, aided by Palantir’s Maven Smart System to identify targets and choose weapons used for attacks. But experts at the hearing, hosted by the Tom Lantos Human Rights Commission, said the argument that AI reduces civilian harm simply isn’t supported by the facts on the ground, pointing to Iran, Ukraine, Yemen, and Gaza as evidence. “I disagree with that,” said Steven Feldstein, a senior fellow in the Democracy, Conflict, and Governance Program at the Carnegie Endowment for International Peace. While some tout targeting accuracy within AI systems being up to 90%, Feldstein says the real number may be closer to 50%. Deborah Brown, deputy director of technology, rights & investigations at Human Rights Watch wasn’t convinced by the argument either. “Whether it's Israel in Gaza, whether it's Iran, the evidence doesn't suggest that this is limiting the harm because of the high level of civilians dead.” Lawmakers have raised questions over whether AI played a role in the February 28 attack https://www.armscontrol.org/act/2026-05/news/ai-plays-major-role-war-iran on a girls’ elementary school in Minab, Iran, which killed over 170 people. Palantir’s Shyam Sankar made the compassionate case for the military using Maven when asked https://www.foxnews.com/media/palantir-executive-says-ai-enabling-rapid-battlefield-planning-high-speed-us-strike-operations about his company’s alleged role in the attack. "It's a tragedy for those girls for sure,” Sankar said. “The thing I would say is that in the history of warfare, technology has always reduced collateral damage.” But Feldstein explained that “even if there is more accuracy, because the number of targets is so much vaster, the ultimate number of civilians harmed is likely far higher as well.” For instance, the Israel Defense Forces claims to have a 90% accuracy https://www.theguardian.com/world/2024/apr/03/israel-gaza-ai-database-hamas-airstrikes rate with Lavender, its AI-targeting system used in Gaza. But Amanda Klasing, national director of government relations and advocacy at Amnesty International USA, pointed out that accuracy doesn’t matter when proportionality is “blown out of the water,” such as when a 2,000-pound bomb is dropped on a residential building. AI systems are also carrying out instructions based on rules of engagement determined by humans, which could be based on instructions that do not prioritize civilian harm reduction. 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,” said Feldstein. Lavender has been used to identify https://www.theguardian.com/world/2024/apr/03/israel-gaza-ai-database-hamas-airstrikes 37,000 Palestinian men as targets. But an intelligence source told https://www.theguardian.com/world/2024/apr/03/israel-gaza-ai-database-hamas-airstrikes the Guardian that they were constantly being pressured to bring more targets. “The numbers changed all the time, because it depends on where you set the bar of what a Hamas operative is,” they said. The IDF also approved https://www.theguardian.com/world/2024/apr/03/israel-gaza-ai-database-hamas-airstrikes these attacks on low-level targets, even if it meant killing 15 to 20 civilians as collateral damage. Even as proponents promise that AI will create a more humane approach to warfare, training for civilian harm reduction has been hollowed out. The Pentagon’s Civilian Protection Center of Excellence, established in 2023 to train combatant commands on civilian harm reduction methods, is reducing https://theintercept.com/2026/07/20/hegseth-civilian-harm-deaths-war/ its staff from 40 in January 2025 to just nine by the end of this year. Witnesses pointed out that part of the issue is that the Pentagon has not shared evidence that AI systems mitigate 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,” said Brown.