Congressional Hearing Testimony: AI Targeting Systems Tied to Gaza, Iran Civilian Deaths Despite Accuracy Claims

At a hearing before the Tom Lantos Human Rights Commission, lawmakers heard testimony that the AI systems now guiding U.S. and Israeli military targeting have not delivered on their central promise: fewer civilian deaths.

Rep. Jim McGovern (D-Mass.) opened by restating the government's pitch, that AI could sharpen assessments of military operations and cut civilian casualties. The witnesses who followed disagreed, pointing to Gaza, Iran, Ukraine, and Yemen as evidence the claim does not hold.

The Pentagon has already put the technology to work. Using Palantir's Maven Smart System to identify targets and select weapons, the U.S. military struck 13,000 targets in the first 38 days of a recent conflict, according to testimony cited at the hearing.

Steven Feldstein of the Carnegie Endowment for International Peace challenged the industry's headline accuracy figures directly. Systems marketed at 90 percent accuracy, he said, may in practice perform closer to 50 percent. He added that even genuine gains in precision are offset by scale: when a system generates vastly more targets, the total number of civilians harmed can rise even as per-strike accuracy improves.

The Israeli military's Lavender system, used in Gaza, was cited as the clearest example. The IDF claims a 90 percent accuracy rate for Lavender, which has been used to flag 37,000 Palestinian men as targets, according to testimony and prior Guardian reporting referenced at the hearing. An intelligence source told the Guardian that staff were pressured to keep generating targets, with the threshold for what counted as a "Hamas operative" shifting constantly, and that the IDF approved strikes on low-level targets even when they were expected to kill 15 to 20 civilians as collateral damage.

Amanda Klasing of Amnesty International USA told the commission that accuracy claims are beside the point when proportionality fails, as when a 2,000-pound bomb is dropped on a residential building. Feldstein made a related point about command decisions: if a military sets a threshold accepting five civilian deaths per targeted individual, that ratio is what the system is instructed to execute. The machine follows rules humans set.

Deborah Brown of Human Rights Watch was blunt about the record so far.

"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 also raised questions about whether AI targeting played a role in the February 28 strike on a girls' elementary school in Minab, Iran, which killed more than 170 people. Palantir's Shyam Sankar, asked about the company's alleged role, called the deaths a tragedy but argued that technology has historically reduced collateral damage in warfare.

Meanwhile the institutional infrastructure meant to check civilian harm is shrinking. The Pentagon's Civilian Protection Center of Excellence, created in 2023 to train combatant commands on harm-reduction methods, is set to fall from 40 staff in January 2025 to nine by year's end.

Brown told the commission the Pentagon has not produced evidence supporting its own claims. If AI is actually limiting civilian harm and improving precision, she said, the government should show the data.