Pentagon AI targeting loophole Directive 3000.09

โดย Tim Official: Gemini

So a machine can now build the kill list and no rule covers it. The Pentagon reportedly found a legal workaround to let AI pick targets without ever calling it a weapon. It's already been used on 2,000 targets in 96 hours. Here's the deal: the Defense Department turned on ChatGPT Mil and Grok for Government inside GenAImil, its internal AI platform, with Google's Gemini already there. About 1.7 million of the 3 million personnel are using it. Every outlet covered this as a software rollout, but look at what the Pentagon's own AI chief put in a court filing earlier this year: Cameron Stanley wrote that Grok let US forces put more than 2,000 munitions on 2,000 targets in 96 hours during the Iran war. The only policy governing machines and lethal force is Directive 3000.09, effective January 25, 2023. It defines an autonomous weapon system as one that can 'select and engage targets without further intervention by an operator.' That means it has to do both: select AND engage. But then there's the exclusion list. Item seven says the directive doesn't apply to 'systems that are not weapon systems.' That's the whole loophole. A model that reads the intelligence, ranks the options, and hands a commander a finished target list never engages anything. It fires nothing. So it isn't a weapon system, and item seven exempts it completely. The directive even defines target selection as identifying targets for engagement. It just never applies that definition to anything except a weapon. So think about what that leaves: everything 3000.09 demands lands on the thing that pulls the trigger. It has to pass rigorous testing, clear three undersecretaries before development and again before fielding, survive a legal review, and run on data that is 'transparent to, auditable by, and explainable by relevant personnel.' None of it touches the system that decided what to shoot at. And commercial models couldn't clear that auditable bar anyway, because the weights are a trade secret owned by a private company. The Pentagon's defense is that a human stays in the loop and responsibility sits with the chain of command. But researchers who study these systems say that the model structures the decision before the human ever sees it. A commander under time pressure approving a machine-built list is signing, not deciding. And the only hard limits anyone actually wrote down were in a vendor contract. Anthropic insisted its models couldn't be used for mass surveillance of Americans or for lethal autonomous weapons. The Pentagon labeled the company a supply-chain risk to national security, the same label it uses for Huawei. Anthropic sued in March. But every lab that accepted the looser terms is now deployed across the department. So American policy on AI in warfare is being set by whichever company bids lowest on restrictions. And Washington has noticed: the House version of the 2027 defense bill would force 3000.09 to cover mission planning. Kirsten Gillibrand's bill would ban AI outright for target selection, nuclear use and domestic surveillance. But the White House got there first. NSPM-11 ordered the directive rewritten inside 90 days, and the memo blames previous administrations for bureaucracy that 'hampered the pace of AI adoption.' To be fair, most of what runs on GenAImil is travel forms, logistics and paperwork. Militaries have leaned on decision-support software for decades, and a human still signs every strike. If you think about it like that then this is just faster staff work. But reporters asked the Pentagon directly whether these models can be used for targeting or intelligence analysis. It literally DECLINED to answer. A machine can now build the list of who gets hit, and no rule covers it. So who takes the blame when the list is wrong?