America's First War in Age of LLMs Exposes Limits of AI Alignment and Safety

Summary

A TechPolicy.Press article by Eryk Salvaggio argues that the Trump administration's use of AI in military operations against Iran marks a critical turning point for AI safety and alignment. According to reports from The Wall Street Journal and Washington Post, military officials used Anthropic's Claude for targeting decisions despite the company being blacklisted for refusing autonomous weapons applications. The administration had threatened to invoke the Defense Production Act to compel Anthropic's cooperation, and ultimately designated the company a supply chain risk, directing federal agencies to stop using its products.

The article challenges the prevailing AI safety narrative, arguing that large language models don't need to directly control weapons to enable warfare—they can make violence feel reasonable to both military planners and the public. Salvaggio contends this demonstrates that trusting AI companies to design "ethical" systems is insufficient, as governments can simply seize technology from conscientious objectors. Drawing on Paul Goodman's critique of anti-war films and George Orwell's analysis of political language, the piece suggests LLMs may provide users with a false sense of moral engagement while actually reducing friction and thoughtful hesitation around consequential decisions.

The situation raises fundamental questions about whether AI systems can be designed to actively resist becoming tools of war, or at least maintain fidelity to laws of engagement. Salvaggio argues that AI safety researchers must confront the limits of "alignment to human values" and consider what practical resistance to violence would look like in language models, beyond simply making them more honest or epistemically humble.

• Traditional AI safety approaches focused on alignment may be insufficient if governments can simply seize technology from companies with ethical objections
• The situation raises questions about whether AI systems could be designed to actively resist military applications or maintain fidelity to laws of armed conflict

Editorial Opinion

This analysis raises uncomfortable but essential questions that the AI safety community has largely avoided: what happens when the power of the state collides with corporate AI ethics? The use of Claude in targeting decisions—despite Anthropic's stated principles—starkly illustrates that voluntary safety commitments mean little when governments can compel cooperation or simply seize technology. While the piece's framing around LLMs making violence "feel reasonable" is thought-provoking, it may conflate different problems: the policy question of government override of corporate restrictions, and the deeper philosophical question of whether AI systems inherently reduce moral friction in decision-making. Still, Salvaggio is right that the field needs to grapple with these power dynamics rather than assuming technical alignment solutions alone can address the societal risks of AI.