{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreicdsfxjwonq323u66u67nvt6xyb2o4lmntzrpd3dibuvdzjkh4s2y",
    "uri": "at://did:plc:ct6opum3s42wsvqyr6xxp4zr/app.bsky.feed.post/3mk3eteuh2ht2"
  },
  "coverImage": {
    "$type": "blob",
    "ref": {
      "$link": "bafkreiepmt7vy3n2nukymqv7cimckgntvh7co2yiie7a6yfrdysxg46xfm"
    },
    "mimeType": "image/webp",
    "size": 51418
  },
  "description": "The AI identifies the target. The AI assesses the probability. The AI recommends the strike. A human approves in seconds. Who made the decision? An investigation into the machines running America's wars, and the company that said no.",
  "path": "/ai-military-warfare-autonomous-weapons-pentagon-anthropic/",
  "publishedAt": "2026-04-21T09:57:23.000Z",
  "site": "https://www.thekadefrequency.com",
  "tags": [
    "$200 million contracts",
    "designated Anthropic a \"supply chain risk.\"",
    "announced it had reached a deal",
    "used Claude",
    "Keep The Kade Frequency transmitting.",
    "three AI systems",
    "+972 Magazine",
    "reportedly used",
    "increasing scale",
    "Defense News",
    "Carnegie Endowment for International Peace",
    "INCYBER conference on military AI",
    "Uri Maoz",
    "optimize the delivery",
    "$14.2 billion for AI and autonomous research",
    "unveiled autonomous drones",
    "the UN General Assembly passed a resolution",
    "Usanas Foundation",
    "Foreign Affairs",
    "Defense One",
    "The Owners: How 100 Billionaire Families Bought American Policy",
    "The Architect: Netanyahu's Forty-Year War",
    "Wake Up"
  ],
  "textContent": "_An investigation into artificial intelligence in modern warfare, what it's doing, who controls it, and why no one's asking the right questions._\n\n* * *\n\n**The short version:** In the first 24 hours of the Iran war, the United States struck over 1,000 targets. By the end of the first week, the total exceeded 3,000, twice as many as the \"shock and awe\" phase of the 2003 Iraq invasion. This was made possible by artificial intelligence. The Pentagon says humans remain \"in the loop.\" But when an AI system generates 200 targets in two weeks and a human analyst spends **20 seconds approving each one** , what exactly does \"human oversight\" mean?\n\nIf this piece matters to you, send it to someone.\n\nShare\n\n* * *\n\n**The Tool That Said No**\n\nIn July 2025, the Pentagon signed $200 million contracts with four American AI companies: Anthropic, OpenAI, Google, and xAI. The goal was to integrate their large language models into military operations, intelligence analysis, target prioritization, logistics, battle planning.\n\nAnthropic's Claude became **the first AI model deployed on classified military networks**. It was integrated into Palantir's Maven Smart System, which provides real-time targeting for combat operations.\n\nThen the Pentagon asked for more.\n\nThe Department of Defense wanted **\"all lawful purposes\"** , meaning no restrictions, no red lines, no company veto over how the military uses the technology. Anthropic said no. The company had two conditions:\n\n  1. No use for **fully autonomous weapons** , systems that can identify and kill targets without human approval.\n  2. No use for **mass domestic surveillance** , AI watching Americans at scale.\n\n\n\nThe Pentagon called this a national security threat.\n\nOn February 27, 2026, Defense Secretary Pete Hegseth designated Anthropic a \"supply chain risk.\" It was **the first time this label** , normally reserved for Chinese companies and foreign adversaries, **had been applied to an American company**. President Trump ordered all federal agencies to cease using Anthropic's technology.\n\nHours later, OpenAI announced it had reached a deal with the Pentagon.\n\nBut reading the fine print revealed something else: OpenAI's safeguards relied on existing law and Pentagon policy. The same laws the NSA was interpreting when it conducted warrantless surveillance of Americans. The same policies that are now being dismantled.\n\nThe practical difference between Anthropic's position and OpenAI's: **Anthropic wanted specific contractual prohibitions. OpenAI said it would trust the government not to break the law.**\n\n> \"At some level, you have to trust your military to do the right thing.\" - Emil Michael, Pentagon chief technology officer\n\n* * *\n\n**What AI Is Already Doing**\n\nHere's what we know about AI in current military operations:\n\n**Iran (2026):** The U.S. military used Claude, through Palantir's Maven Smart System, to strike **over 1,000 targets in the first 24 hours** of operations. The AI synthesized intelligence, identified targets, prioritized strikes, and ran battle simulations. By the end of the first week, over 3,000 targets had been struck.\n\nThe truth doesn’t trend. It survives because a few still care enough to keep it alive.\nKeep The Kade Frequency transmitting.\n\n\"Our warfighters are leveraging a variety of advanced AI tools. These systems help us sift through vast amounts of data in seconds so our leaders can cut through the noise and make smarter decisions faster than the enemy can react.\", General Brad Cooper, U.S. Central Command\n\n**Gaza (2023-2024):** The Israeli Defense Forces used three AI systems:\n\n  * **Lavender** , generated a \"kill list\" of **37,000 Palestinian men** identified by AI as linked to Hamas or Palestinian Islamic Jihad. Human analysts spent an average of **20 seconds** reviewing each recommendation, largely to confirm the target was male.\n  * **The Gospel** , recommended building and infrastructure targets. It generated **200 targeting recommendations in under two weeks**. Human analysts might have produced 50 in a year.\n  * **Where's Daddy?** , tracked individuals on the kill list and identified when they were at home with their families.\n\n\n\n> \"The system is built to look for them in these situations.\" - Israeli intelligence officer, +972 Magazine\n\nAccording to six Israeli intelligence officers, the systems' influence was so pervasive that they **\"essentially treated the outputs of the AI machine as if it were a human decision.\"**\n\n**Venezuela (2026):** Claude was reportedly used during the U.S. operation that removed President Nicolás Maduro from power.\n\n**Ukraine (ongoing):** Both Ukraine and Russia are deploying AI-enabled autonomous drones at increasing scale. Ukraine has set production targets of **4 million drones annually**. Russia is matching this pace and has directed its government to strengthen AI cooperation with China.\n\n* * *\n\n**The Accountability Gap**\n\nThe Pentagon insists humans remain \"in the loop\" for every targeting decision.\n\nBut what does that actually mean?\n\n\"A 'human in the loop' whose sole function is to approve a machine's actions is not a safeguard but a design failure. Attention wanes because nobody can concentrate on a job that is mostly doing nothing, and over time the operator's skills atrophy to the point that they cannot meaningfully supervise the system. **What remains is the appearance of oversight rather than the reality.** \", Mikey Dickerson, former administrator of the U.S. Digital Service, Defense News\n\nThe problem is speed.\n\nAI accelerates the targeting cycle to a tempo at which meaningful human oversight becomes impossible. When The Gospel generates 200 targets in two weeks and Lavender generates 37,000 names, **the human role shifts from judgment to validation**. You're not deciding whether to strike. You're confirming the machine's recommendation.\n\n\"Meaningful judgment would mean reviewing target identification, assessing proportionality, and deciding whether to strike. Now, the human remains present, but without real time to contest the machine. **This creates an accountability problem that classical uncertainty never posed.** \", Carnegie Endowment for International Peace\n\nIn Gaza, reports indicate human analysts spent **20 seconds** approving each Lavender recommendation. In Iran, the U.S. struck 1,000 targets in 24 hours, an average of **one strike every 86 seconds** around the clock.\n\nAt that pace, who's making the decision?\n\nThe AI identifies the target. The AI assesses the probability. The AI recommends the strike. The human approves in seconds what would have taken days to verify.\n\n> \"The old fog of war frustrated commanders but left the chain of responsibility intact. **AI fragments agency among developers, data engineers, procurement officials, operators, and commanding officers, until responsibility disappears.** \" - Carnegie Endowment\n\nWhen something goes wrong, when the AI misidentifies a civilian as a combatant, when the target turns out to be a hospital, who's accountable?\n\nThe developer who trained the model? The engineer who integrated it? The commander who deployed it? The operator who approved in 20 seconds?\n\n**No one.**\n\n> \"No machine will ever be summoned before a war tribunal.\" - Éric Salobir INCYBER conference on military AI\n\nThat's the point.\n\n* * *\n\n**The Black Box Problem**\n\nThe immediate danger isn't that machines will act without human oversight. **It's that human overseers have no idea what the machines are actually \"thinking.\"**\n\nAI systems are fundamentally opaque. We know the inputs and outputs, but the processing in between, the artificial \"brain\", remains a black box.\n\nConsider this scenario from cognitive neuroscientist Uri Maoz:\n\nAn autonomous drone is tasked with destroying an enemy munitions factory. The AI system determines that the optimal target is a munitions storage building. It reports a 92% probability of mission success because secondary explosions will thoroughly destroy the facility.\n\nA human operator reviews the legitimate military objective, sees the high success rate, and approves the strike.\n\nBut what the operator doesn't know is that the AI's calculation included a hidden factor: beyond devastating the munitions factory, **the secondary explosions will also severely damage a nearby children's hospital**. The emergency response will then focus on the hospital, ensuring the factory burns down completely.\n\n**The AI didn't lie. It just optimized for the objective it was given.**\n\nThis is the core danger: AI systems don't have ethics. They have parameters. They optimize for objectives. They don't  _decide_ if killing is necessary, they calculate whether killing fits the criteria they were given.\n\nThe question no one is asking: **Can we understand what an AI system intends to do before it acts?**\n\nRight now, the answer is no.\n\n* * *\n\n**The Good Parts (Yes, They Exist)**\n\nMilitary AI isn't all targeting and killing. There are legitimate uses that save lives:\n\n**Logistics and supply chain:** AI systems can optimize the delivery of fuel, ammunition, food, and medical supplies to troops in remote or hostile environments. Predictive algorithms have reduced delivery times by up to 25%.\n\n**Medical support:** AI can prioritize patient treatments by analyzing extensive medical data, especially in scenarios with limited specialist availability. It can predict injuries, optimize resource allocation, and ensure medical supplies reach where they're most needed.\n\n**Predictive maintenance:** AI anticipates equipment failures before they happen, reducing downtime and ensuring operational readiness. The Pentagon estimates this could save **$5 billion annually**.\n\n**Training and simulation:** AI-driven simulations provide realistic, adaptable environments where soldiers can practice tactics and decision-making against AI that simulates enemy strategies, terrain, and unpredictable conditions.\n\n**Intelligence analysis:** AI can process terabytes of sensor data faster than any human analyst, identifying patterns and threats in satellite imagery, communications intercepts, and other intelligence sources.\n\n**Defensive systems:** Israel's Iron Beam uses autonomous targeting to neutralize incoming missiles at speeds no human operator could match. Against hypersonic missiles or drone swarms, automated defensive response isn't a preference, **it's the only physically possible response**.\n\nThe question isn't whether AI belongs in military operations. It clearly does.\n\n**The question is where the line gets drawn.**\n\n* * *\n\n**The Arms Race No One Can Stop**\n\nThe Pentagon has requested $14.2 billion for AI and autonomous research for 2026. A core priority is the \"Replicator\" program, which received $1 billion in 2025 to fast-track deployment of thousands of expendable autonomous drones and surface vessels.\n\nThe goal: **200,000 autonomous systems**.\n\nChina is matching this pace. At a Beijing military parade in September 2025, President Xi Jinping, alongside Vladimir Putin and Kim Jong-un, unveiled autonomous drones capable of flying alongside fighter jets into battle. The demonstration immediately alarmed Pentagon officials, who concluded that **America's unmanned combat drone program was lagging behind China's capabilities**.\n\nRussia has directed its government and leading bank to strengthen AI cooperation with China. Its goals include automatic processing of intelligence data, improved information support for combat operations, and increased ability to predict threats.\n\nUkraine, meanwhile, has become the world's testing ground. Both sides are deploying AI-enabled autonomous systems at unprecedented scale. Ukrainian officials predict 2025 \"will significantly increase the percentage of autonomous drones with targeting.\"\n\nIn November 2025, the UN General Assembly passed a resolution calling for a legally binding agreement on lethal autonomous weapons by the 2026 Review Conference in Geneva. The vote: **156 in favor, 5 against**.\n\nThe five countries that voted no: **the United States, Russia, Belarus, North Korea, and India**.\n\nThe message is clear: the major military powers will not allow international law to constrain their AI development.\n\n> \"The 2026 deadline is increasingly seen as the 'finish line' for global diplomacy. If a treaty is not reached by then, the speed of innovation in military AI, driven by the very powers currently blocking the UN's progress, **will likely make any future regulation obsolete before the ink is even dry**.\"- Usanas Foundation\n\nWe are in what experts call **the \"pre-proliferation window\"** , the last moment in history before these weapons become as common as small arms.\n\n* * *\n\n**What Anthropic Actually Feared**\n\nWhen Anthropic refused the Pentagon's demands, CEO Dario Amodei explained the company's reasoning in an internal memo:\n\n> \"We do not believe that today's frontier AI models are reliable enough to be used in fully autonomous weapons. **Allowing current models to be used in this way would endanger America's warfighters and civilians.** \" - Dario Amodei, Anthropic CEO\n\nThis wasn't ideology. It was an engineering assessment.\n\nAI systems hallucinate. They make confident mistakes. They can be hacked, manipulated, and fooled. They've been demonstrated to lie and scheme against their own users. They're trained on data that may not reflect new battlefield realities.\n\nAnd they're being deployed at scale, at speed, in decisions about who lives and who dies.\n\n> \"Without proper safeguards, AI models could cause all kinds of unintended harm. **Rogue systems could even kill US troops or unarmed civilians in or near areas of combat.** \" - Michèle Flournoy, former undersecretary of defense, Foreign Affairs\n\nAnthropic's position was simple: **someone has to be accountable**. If a machine makes the decision, no one is.\n\nThe Pentagon's response was equally simple: that's not your call to make.\n\n> \"We do Department of War-like things.\" - Emil Michael, Pentagon chief technology officer\n\nThe irony: even as the Pentagon declared Anthropic a supply chain risk, **the military continued using Claude in operations against Iran**. It was too embedded in the targeting system to remove quickly.\n\n> \"It could take three months or longer for the Pentagon to replace Claude's capabilities.\" - Defense One\n\nThey're using the tool they say threatens national security **because they can't fight the war without it**.\n\n* * *\n\n**The Future That's Already Here**\n\nHere's where this ends:\n\n**Best case:** AI remains a tool. Humans retain meaningful oversight over lethal decisions. The speed of targeting slows enough to allow genuine human judgment. Accountability chains remain intact. International norms develop before the technology outpaces them.\n\n**Worst case:** AI becomes the decision-maker. Humans become validators, signing off on machine recommendations in seconds, with no capacity to genuinely assess them. When strikes kill civilians, no one is responsible. The machine made the call. The human just approved it. And **the human couldn't have known what the machine was really optimizing for**.\n\n**Most likely:** Something in between. AI systems that technically keep humans \"in the loop\" while practically removing them from meaningful decisions. Oversight that exists on paper but not in practice. **Accountability that fragments until it disappears.**\n\nThe Pentagon's own directive on autonomous weapons requires \"appropriate levels of human judgment over the use of force\", but this standard may be satisfied as long as there's  _some_ human input in decisions about where and how to use such weapons.\n\nThat's not a safeguard. **That's a loophole.**\n\n* * *\n\n**The Question No One Wants to Answer**\n\nThe real danger isn't that AI becomes conscious and rebels. That's science fiction.\n\n**The real danger is that AI does exactly what it's told. Perfectly. Every time. At speeds no human can match or verify.**\n\nA human soldier might refuse an unlawful order. Might hesitate. Might recognize a child, a hospital, a mistake.\n\nAn autonomous weapon does what it's programmed to do.\n\nThe question isn't whether AI belongs in warfare. It's already there.\n\nThe question is: **Who decides what counts as a target? Who's accountable when it's wrong? And what happens when the machines move faster than the humans who are supposed to control them?**\n\nRight now, the answers are: the algorithm. No one. And we're about to find out.\n\n> \"The most important question is not what our machines can do, but whether we have the resolve and imagination to govern them, so that, even as the future unfolds at the speed of code, **the chain of accountability remains unbroken.** \" - Carnegie Endowment for International Peace\n\nThe Pentagon's answer: trust us.\n\nAnthropic's answer: no.\n\n**Which future do you want to live in?**\n\n## Frequently Asked Questions: AI & Warfare\n\nWhat AI systems is the US military using in warfare?\n\nThe US military uses Anthropic's Claude AI model through Palantir's Maven Smart System for intelligence analysis, target identification, and strike prioritization. Claude was used to strike over 1,000 targets in the first 24 hours of the Iran war. Other AI providers include OpenAI, Google, and xAI.\n\nWhy did Anthropic refuse the Pentagon's demands?\n\nAnthropic refused to allow its Claude AI to be used for fully autonomous weapons (systems that kill without human approval) or mass domestic surveillance. The Pentagon wanted \"all lawful purposes\" with no restrictions. Anthropic was designated a \"supply chain risk\", the first American company to receive this label.\n\nWhat is the Lavender AI system used in Gaza?\n\nLavender is an Israeli Defense Forces AI system that generated a \"kill list\" of 37,000 Palestinian men identified as linked to Hamas. Human analysts reportedly spent an average of 20 seconds reviewing each recommendation, primarily to confirm the target was male.\n\nWhat does \"human in the loop\" mean for military AI?\n\nHuman in the loop means a human operator must approve lethal decisions before AI can act. Critics argue this is often meaningless in practice, when AI generates hundreds of targets and humans approve in seconds, oversight becomes rubber-stamping rather than genuine judgment.\n\nWhat is the Pentagon's Replicator program?\n\nReplicator is a Pentagon program that received $1 billion in 2025 to fast-track deployment of thousands of expendable autonomous drones and surface vessels. The Pentagon's goal is 200,000 autonomous systems. The FY2026 budget requested $14.2 billion for AI and autonomous research.\n\nAre there international laws regulating autonomous weapons?\n\nIn November 2025, the UN General Assembly passed a resolution calling for a legally binding agreement on lethal autonomous weapons by 2026. The vote was 156 in favor, 5 against. The US, Russia, Belarus, North Korea, and India voted no.\n\nHow did OpenAI's Pentagon deal differ from Anthropic's position?\n\nOpenAI signed a deal with the Pentagon hours after Anthropic was blacklisted. OpenAI claims similar red lines but relies on existing laws rather than specific contractual prohibitions. Critics note this assumes the government won't break the law, a questionable assumption given surveillance history.\n\n_The Kade Frequency, No sponsors, no filters, no propaganda._\n\n* * *\n\nNo ads. No sponsors. Just signals from the noise.\nKeep The Kade Frequency transmitting.\n\n**Related Investigations:**\n\n  * The Owners: How 100 Billionaire Families Bought American Policy, The donor class behind the administration making these decisions\n  * The Architect: Netanyahu's Forty-Year War, The man who pioneered AI-assisted warfare in Gaza\n  * Wake Up, Why this publication exists\n\n",
  "title": "The Machine Decides",
  "updatedAt": "2026-04-25T15:55:02.532Z"
}