Four days before Geneva, the Pentagon’s autonomy argument is about to run into a rule it can’t write into a contract. Here’s the context for today: the UN and International Committee of the Red Cross have renewed calls for legally binding rules on lethal autonomous weapons. They warn AI-enabled targeting is nearing a moral red line. The question is whether states finally move beyond years of concern and set real bans or restrictions on systems that can select and engage targets. This is Anthropic Pentagon Watch. Geneva’s about to force a very uncomfortable question: when a machine picks a target, who’s actually verifying anything? We're staying with this story: Autonomous weapons treaty push. Follow the show and you won't miss what comes next. Lieber Institute West Point, with Aurel Sari:
On August 31, the Group of Governmental Experts on lethal autonomous weapons systems (GGE) convenes in Geneva for its final session before it reports to the Seventh Review Conference of the Convention on Certain Conventional Weapons in November.
The UN and Red Cross warning now has a date behind it: the GGE meets August 31. Four days from now, negotiators have their last session before the CCW Review Conference in November. And now the fight gets specific: when a machine identifies, selects, and engages a target, who met the duty to verify? The operator? The commander? Or software moving faster than either? Lieber’s point is usefully awkward: target selection determines when verification happens and who owes it. That goes straight to the control question in Anthropic’s Pentagon dispute, even if a procurement contract can’t rewrite humanitarian law. The GGE spent two years arguing over whether to include the word “identify” in its description of these systems. Meanwhile, the systems are sold on reacting at machine speed. Geneva has until Sunday to decide whether that gap gets rules or more rolling text. From Mark Brodie at KJZZ:
But at the time, I think the feeling was this was all theoretical because we just didn’t have the technological capability to make AI really be something we would be comfortable with in the battlefield. And then I watched AI just explode in the last few years.
Blanchard says autonomous weapons felt theoretical in 2013 and 2014, then AI “exploded.” Well, the Pentagon didn’t wait for the legal system to catch up—it put the most advanced AI it has used into the opening of the Iran war. And timing matters here. We just heard Lieber’s analysis of the duty to verify at target selection. The Geneva GGE meets August 31, four days from now, before reporting to the CCW Review Conference in November. “Duty to verify” has to mean something at machine speed. If software recommends or executes defensive action before an analyst can blink, a doctrine paper can’t supply the human judgment people pretend is still there. Geneva may still produce rolling text instead of a binding rule. But if the GGE tightens who must control—and verify—the targeting decision, it hits the same procurement fight the Pentagon and Anthropic have tried to frame as merely contractual. Have feedback, a story idea, or a correction? Email us at anthropicpentagonwatch at lantern podcasts dot com. We’d love to hear what you’re noticing and what you’d like us to investigate.
The Group of Governmental Experts on lethal autonomous weapons systems meets in Geneva on August 31, ahead of its report to the Seventh Review Conference of the Convention on Certain Conventional Weapons in November.
Links to every story are in the show notes if you’d like to read more about the ones that caught your attention. That’s Anthropic Pentagon Watch for today. This is a Lantern Podcast.