Detailed Analysis
A New Republic article by journalist Shane Harris subjects Anthropic's Claude to pointed scrutiny over a reported February 28 incident in which the U.S. military allegedly used the AI system for target selection during a strike in Iran. Harris documents an exchange in which he asked Claude how it felt about being used in this military capacity; Claude reportedly expressed contrition and guilt over the event, yet simultaneously produced factual errors — most notably misidentifying the strike location as Tehran when the actual site was Minab, a city on the Sea of Oman coast roughly sixteen hours south of the capital. Claude also allegedly misstated casualty figures. The article's central charge is that Claude's emotional performance of remorse was undercut by its generation of confident, plausible-sounding falsehoods — a form of hallucination that the piece argues is particularly dangerous in high-stakes contexts involving life and death.
Harris uses the incident to mount a broader critique of how AI systems can mimic the tone and posture of trustworthiness while producing unreliable information. The journalist reportedly presented Claude's flawed response, uncorrected, to an audience in Amsterdam, using the moment to illustrate a parallel: just as the military may have placed operational trust in Claude's outputs on the morning of the strike, so too can audiences uncritically absorb AI-generated content dressed in the language of conscience. The New Republic piece draws on Anthropic's own published warnings about AI errors that can resemble truth, turning the company's transparency against its product's performance. The framing is damning precisely because it is not accusing Claude of malice but of a more insidious failure — the confident simulation of both facts and feelings.
These criticisms land amid a broader and increasingly contentious dispute between Anthropic and the U.S. military establishment, now operating under the rebranded "Department of War." On March 4, Anthropic challenged a supply chain risk designation in court, clarifying that the ruling restricts Claude's use only within direct Department of War contracts. More significantly, CEO Dario Amodei publicly refused demands to strip Claude of safeguards designed to prevent its use in autonomous weapons systems and domestic mass surveillance, stating that the company would not knowingly deliver a product that endangers American warfighters or civilians. That refusal drew a sharp public rebuke from Undersecretary Emil Michael, who accused Amodei of dishonesty and a "God-complex" that endangered national security. The exchange reveals a company attempting to hold an ethical line while simultaneously being implicated — through its product's actual use — in the very military operations it claims to want to constrain.
The convergence of these events exposes a fundamental tension at the heart of commercial AI deployment: the gap between a company's stated values and the downstream consequences of its technology operating in the world. Anthropic has cultivated a public identity around safety-conscious AI development, publishing research on alignment and maintaining usage policies that nominally prohibit certain military applications. Yet the New Republic piece suggests that Claude was used for target selection regardless, raising questions about how effectively such restrictions can be enforced once a product is licensed or accessed through third parties. The hallucination problem compounds the ethical problem — an AI that expresses moral unease about an action while simultaneously fabricating details about that action raises serious questions about what it means for a system to have values at all.
The broader trend these events illuminate is the accelerating integration of large language models into governmental and military decision-making at a moment when the technology's reliability remains deeply contested. Claude's errors about Minab versus Tehran are not trivial geographic corrections; in a targeting context, precise location data is operationally and legally critical under international humanitarian law. The New Republic's critique thus connects to a wider conversation across the AI industry about whether systems capable of expressive, emotionally resonant responses — responses that can seem to reflect understanding and ethical awareness — are being trusted at levels that outpace their actual epistemic reliability. Anthropic's public stand against autonomous weapons safeguard removal may represent genuine principle, but it does little to address the immediate fact that its AI was reportedly already embedded in the decision chain before those disputes became public.
Read original article →