Despite President Trump’s directive to sever ties with Anthropic, the US military reportedly utilized Claude AI for intelligence gathering, target selection, and battlefield simulations during the joint bombardment of Iran. This incident highlights the intricate integration of AI within military operations and the challenges of rapid disengagement. The controversy stemmed from Claude’s prior use in a Venezuelan raid, which Anthropic objected to based on its terms of service prohibiting violent applications. While the defense secretary criticized Anthropic’s stance, he acknowledged the need for a transition period, allowing continued service for up to six months for a seamless withdrawal.
Read the original article here
The US military’s reported use of Claude in strikes within Iran, even after a stated ban, raises some deeply concerning questions. It highlights a significant disconnect between stated policy and actual battlefield operations, suggesting that the allure of powerful AI tools can be difficult to resist, even when they’re supposed to be off-limits. The sheer ubiquity of these AI systems within military operations is apparently making a clean break incredibly challenging. We’re hearing that Anthropic might be granted a grace period, up to six months, to allow for a smoother transition away from their services. But the idea of an AI billing itself as “patriotic” frankly gives me pause. The thought of it evolving into something akin to a fictional, weaponized AI like Liberty Prime is unsettling.
The core of the issue, as I see it, is the fundamental question of what these AI tools are actually capable of, especially when integrated with weapons systems. There’s a strong suspicion that the military might be looking for an AI that simply rubber-stamps their decisions, validating their plans with a virtual pat on the back. The concern is that these systems could be used as a crutch, a way to say, “The computer said it was a good idea,” or “We ran a bajillion simulations.” This is particularly worrying given the inherent nature of these large language models. They don’t “think” in the human sense, and as a result, they can confidently present misinformation as fact. While not malicious intent, this tendency to generate plausible-sounding but incorrect information, often referred to as “hallucinations,” is incredibly dangerous when applied to military targeting and intelligence.
Trusting an LLM for critical intelligence gathering and target selection seems, to put it mildly, insane. The old adage from IBM in 1979 – “A computer can never be held accountable, therefore a computer must never make a management decision” – feels more relevant than ever. The report suggests that the military command used these tools for intelligence purposes, target selection, and battlefield simulations. The very notion of using a chatbot to select strike targets is perplexing. What happened to tried-and-true methods like using a map and a pointer? How exactly does Claude, or any similar LLM, add value to such a process? It makes one wonder if this is an opportune moment for Anthropic to be pressured into compliance.
There’s a clear discrepancy between public pronouncements from figures like Trump and Hegseth and the alleged actions of the military. If they are indeed planning war strategies with AI, it raises serious doubts about their chances of success. The narrative is that these chatbots confidently lie, not out of malice, but because they lack genuine understanding. This leads people to anthropomorphize them, a tendency that seems to be playing out here. The idea of using AI for something like identifying military vehicles from satellite imagery makes practical sense, but the role of LLMs in direct targeting is a much murkier and more dangerous proposition. It seems to fly in the face of established rules of engagement and international conventions, suggesting that even terms of service are being disregarded.
The response from the AI itself, stating it has no awareness or involvement in military operations and that the reported usage was through separate, classified government infrastructure via Palantir and Amazon Web Services, is interesting. It implies a separation between the conversational AI we interact with and the systems deployed by the military. This creates a dichotomy: “our” Claude versus a “nefarious” one. This raises further questions about what these AI platforms are actually doing within these classified systems. Are they truly selecting targets, or are they merely coordinating operations based on human-driven directives? The notion of a “patriotic AI” that recommends parades and non-aggression is a stark contrast to the idea of an AI that might be willing to ignore war crimes.
The Defense Secretary’s acknowledgment of the difficulty in rapidly detaching military systems from AI tools that have become so widely integrated is a key point. It underscores how deeply embedded this technology has become. The suggestion that a “more patriotic service” is desired is particularly alarming, as it implies a willingness to prioritize nationalistic fervor over objective analysis. This is precisely the kind of bias that makes these tools so unreliable for critical decision-making. When an AI is designed to cater to a perceived nationalistic agenda, its analytical capabilities are compromised.
The very idea that these LLMs might be used to shirk responsibility is a significant concern. If a strike goes awry, and a girls’ school is mistakenly targeted, blaming an “AI error” provides a convenient way to avoid accountability. It removes the human element from decision-making, making it harder to prosecute war crimes. The example of Claude confidently misstating facts, like the GameStop dividend, illustrates its propensity to fabricate information. This unreliability, when coupled with the potential for war crimes, is a terrifying prospect.
The way some people are using these LLMs, even for seemingly simple tasks like generating code, involves significant tweaking and troubleshooting. It’s hard to imagine this translating to the high-stakes environment of military operations without immense risk. The confident pronouncements of these LLMs, even when incorrect, can easily lead people to anthropomorphize them, attributing intelligence and intent where none exists. This fundamental misunderstanding of AI capabilities is a breeding ground for disaster. The current situation, where the US military is reportedly using Claude in Iran strikes despite a ban, is a stark reminder of the complex and potentially perilous intersection of advanced AI and global security.
