Anthropic Defies Pentagon Demand to Loosen Military AI Restrictions, Faces Legal Threat

The Defense Production Act Threat

The Defense Production Act, signed into law in 1950 during the Korean War, gives the president broad authority to direct private industry to prioritize contracts and orders deemed necessary for national defense. It has been invoked periodically for various purposes — most recently during the COVID-19 pandemic to compel production of medical supplies — but using it to force an AI company to provide its technology for military applications would represent an unprecedented application of the law.

Legal experts are divided on whether such an invocation would survive judicial scrutiny. The DPA was designed for physical goods manufacturing — steel, ammunition, medical equipment — not for compelling a software company to alter its terms of service. The question of whether AI model access constitutes a "product" that can be commandeered under the act raises novel legal questions that courts have not yet addressed.

• Pentagon demands Anthropic remove restrictions on AI use in autonomous weapons and surveillance
• Anthropic has refused, citing its foundational commitment to AI safety principles
• Defense Department threatens to invoke the Defense Production Act by Friday
• Legal experts question whether the DPA can compel a software company to change its policies
• The standoff could set precedent for government authority over AI companies

Industry Implications

The confrontation between Anthropic and the Pentagon sends shockwaves through an AI industry that has been navigating an increasingly complex relationship with national security agencies. Google, Microsoft, Amazon, and OpenAI all have significant defense contracts, and each has faced internal and external pressure over military applications of their technology. Google famously withdrew from Project Maven, a Pentagon AI program, after employee protests in 2018, though the company has since expanded its defense work.

If the Defense Production Act is successfully used against Anthropic, it would establish a precedent that any AI company operating in the United States could be compelled to provide its technology for military purposes regardless of its own ethical guidelines. That prospect could chill AI safety research, push safety-focused companies to relocate outside US jurisdiction, or create a bifurcated industry where companies must choose between government contracts and safety commitments.

Conversely, if Anthropic successfully resists the order — whether through legal challenge or political negotiation — it could strengthen the principle that AI companies have the right to set ethical boundaries on how their technology is used, even when the customer is the US government.

The Broader Tension

The standoff reflects a fundamental tension that has been building since large language models and other advanced AI systems began demonstrating capabilities with clear military applications. The US government views AI dominance as essential to national security, particularly in competition with China, which is pouring resources into military AI applications with fewer ethical constraints. From the Pentagon's perspective, allowing a leading AI company to opt out of defense applications is a luxury the nation cannot afford.

From Anthropic's perspective, the restrictions exist precisely because the company believes unconstrained military application of powerful AI systems poses catastrophic risks — risks that are not eliminated simply because the user wears an American uniform. The company's position is that some applications of AI are too dangerous to enable, regardless of who is asking.

How this standoff resolves will likely shape the relationship between the AI industry and the US government for years to come. It is a test case for whether AI safety commitments can withstand the gravitational pull of national security imperatives — and whether the government will use its most powerful legal tools to ensure they cannot.

This article is based on reporting by The Decoder. Read the original article.