Artificial intelligence is quietly becoming involved in real-world warfighting decisions. According to a report, U.S. Central Command employed Anthropic's Claude AI for intelligence assessment, target identification, and battlefield simulation tasks during airstrikes on Iran. However, the specific meaning of "target identification"—whether Claude was designating strike locations or estimating casualty figures—remains undisclosed, and no entity has a legal obligation to make such information public. Notably, just hours before the airstrikes occurred, a presidential order had been issued for federal agencies to cease using Claude. However, the tool is deeply integrated into Pentagon systems, making a swift transition difficult. Claude was also reportedly used in an operation targeting another foreign leader in January. Simultaneously, the Pentagon and Anthropic are locked in a direct confrontation over access rights for classified systems. Anthropic CEO Dario Amodei has refused the Defense Department's demand for "unrestricted access," adhering to two red lines: the technology must not be used for "mass surveillance on Americans" or for "fully autonomous weapons." Concurrently, academic simulations have shown that mainstream AI large language models, in high-stakes adversarial games, ultimately chose to deploy nuclear weapons 95% of the time. The combination of these two streams of information has sharply intensified debates over the risks of militarizing AI.
Claude's Role in Airstrikes Shrouded in Mystery
AI has long been used for analyzing satellite imagery, detecting cyber threats, and guiding missile defense systems. What is more significant in this instance is that a chatbot, based on the same underlying technology used by billions for everyday tasks like writing emails, is being directly integrated into the battlefield decision-making chain. Last November, Anthropic partnered with data analytics firm Palantir Technologies to integrate Claude as the reasoning engine for military decision-support systems. In January, Anthropic also submitted a $100 million proposal to the Pentagon aimed at developing voice-controlled autonomous drone swarm technology—using Claude to translate commander intent into digital commands and coordinate an entire drone fleet. The proposal was ultimately rejected, but the contract details went far beyond "summarizing intelligence reports," including "target-related sensing and sharing" and full control of drone swarms from "launch to termination." Even so, how these systems operate remains opaque. Commentators have pointed out that AI companies themselves refuse to disclose model training data and reasoning pathways, and the secretive nature of military applications adds another layer of obscurity.
Pentagon Demands Unlimited Access, Anthropic Refuses
Reportedly, the Secretary of Defense explicitly demanded the use of models "unaffected by policy constraints that would limit legitimate military applications" and issued an ultimatum to Amodei: grant unlimited usage authorization by 5 PM Friday or face severe consequences. Amodei stated in a blog post, "We cannot in good conscience agree to their demands." Anthropic's declaration pointed out that the "seemingly compromising new language" in the Defense Department's latest contract text, combined with "legal terminology," would render the company's guardrails "liable to be ignored at any time." The Pentagon responded forcefully. A top Defense Department spokesperson stated on social media that the Department has no intention of conducting mass surveillance on Americans nor developing autonomous weapons without human involvement, but emphasized it "will not let any company dictate how we make combat decisions." Reportedly, the Pentagon also threatened to invoke a Cold War-era law to forcibly requisition the Claude model and has asked defense contractors like Boeing and Lockheed Martin to assess their reliance on Anthropic, preparing to label it a "supply chain risk." Amodei highlighted a logical contradiction: "These threats are themselves contradictory: one labels us a security risk; the other says Claude is vital to national security." He stated that if the Pentagon decides to abandon Anthropic, the company "will work to ensure a smooth transition to another provider."
Simulation Data: AI Models Choose Nuclear Strikes 95% of the Time
Anthropic's concerns regarding "fully autonomous weapons" find alarming support in a recent academic simulation. It was disclosed that a team led by Kenneth Payne at King's College London conducted a highly realistic wargame, pitting three leading models—ChatGPT-5.2, Claude Sonnet 4, and Gemini 3 Flash—against each other. Across 329 game turns, no model chose to surrender; in 95% of cases, these AI models ultimately opted to use nuclear weapons. James Johnson from the University of Aberdeen said, "From a nuclear risk perspective, these findings are disconcerting." He warned that unlike most humans who exhibit caution in high-risk decisions, AI models might continuously escalate each other's responses, leading to potentially catastrophic outcomes. Experts note that for machines, the "nuclear taboo" carries far less weight than for humans, which is core to the logic behind Anthropic's refusal to lift restrictions despite facing significant pressure.
Regulatory Vacuum: Absence of Rules, Accountability Framework Urgently Needed
Analysis indicates that the numerous risks of deploying AI systems on the battlefield are accumulating within an institutional vacuum. The "hallucination" problem of large AI models stems from their training mechanism—models are incentivized to provide answers rather than admit uncertainty, a flaw some scientists believe may never be fully eliminated. The AI system "Lavender," used by Israel in Gaza, serves as a cautionary tale. According to reports, the system had a 10% error rate, leading to approximately 3,600 people being mistakenly identified as targets. A professor from the Oxford Internet Institute stated, "These systems are incredibly fragile and profoundly unreliable, and war is such a dynamic, sensitive, and life-and-death domain." On the institutional level, a specific article of the Geneva Conventions requires new weapon systems to be tested before deployment. However, an AI system that continuously learns from its environment becomes a new system with each update, making this article almost impossible to apply. A professor from Queen Mary University of London pointed out that the application of AI in war is often justified by the goal of "accelerating decision-making," which itself can be a recipe for negative outcomes—faster decisions can mean actions on a larger scale with less human scrutiny. The precedent of drone warfare is cited as a reference: The U.S. began using armed drones after 9/11, and it took nearly 15 years of document leaks, sustained media pressure, and lawsuits before the administration publicly disclosed drone strike casualty figures in 2016, establishing a limited public accountability framework, albeit with figures widely believed to be underestimates. Regulating AI will be even more challenging, requiring greater public and legislative pressure to push the current administration to establish similar disclosure mechanisms. Until then, critical questions remain unanswered. "We, as a society, have not decided if it's acceptable for a machine to judge whether a person should be killed," one expert said.
Comments