Recent conflict in Gaza demonstrates that artificial-intelligence driven targeting is no longer a theoretical risk. After the October 2023 Hamas assault, the Israeli military deployed algorithms that selected individuals for lethal action. These systems, operating without fully autonomous weapons, nevertheless determined who would be killed, effectively placing algorithmic judgment at the core of the kill chain.
International negotiations in Geneva have sought to curb such developments. In early September, representatives of 128 states convened to endorse a draft that could lay the groundwork for a binding treaty on autonomous weapons. However, legal teams linked to former leaders of the United States and Russia succeeded in diluting a clause that would obligate human review of AI-generated target lists before any strike.
The diplomatic text concentrates on fully autonomous “killer robots” while leaving the use of AI to choose targets for conventional munitions largely untouched. This omission threatens the long-standing principle of maintaining a “human in the loop,” meaning that meaningful human control over weaponised algorithms could be reduced to a mere formality.
Israel’s operational system, dubbed “Lavender,” scanned phone metadata after the October attacks to flag patterns associated with Hamas operatives. A second tool, informally called “Where’s Daddy?”, extended the search to residential locations, often resulting in the death of family members who were not combatants. Although the process routed AI-selected names through intelligence officers for final approval, the review was described as perfunctory, reducing human oversight to a rubber-stamp function within the targeting cycle.
The Israeli defence establishment later acknowledged that roughly one-tenth of the individuals marked for elimination were not members of the Hamas military wing. Moreover, commanders accepted an anticipated twenty civilian casualties to eliminate a single low-level suspect, a ratio that contravenes the proportionality requirement of international humanitarian law, which bars attacks expected to cause excessive civilian loss relative to the concrete military advantage sought.
To prevent algorithmic complacency, the International Committee of the Red Cross highlights the danger of “automation bias,” where operators over-rely on machine recommendations. Genuine human control therefore demands rigorous interrogation of AI outputs and clear accountability for senior officers when lower-level staff merely endorse algorithmic decisions. Only such scrutiny can ensure that the final act of killing remains a deliberate human choice rather than a mechanical by-product.