The Pentagon and Anthropic continue to feud over the implementation of guardrails for artificial intelligence (AI). The Department of Defense insists that the AI company drop its restrictions of oversight and grant military officials “all lawful use” (whatever that means) of its models. If Anthropic refuses to comply with these demands, then Secretary of Defense Pete Hegseth will allegedly either classify Anthropic as a “supply chain risk” or use the Defense Production Act to force Anthropic to give into the Pentagon’s demands.
The Pentagon’s move here is reckless. It relies on the assumption that AI can be trusted to target, survey, and kill people. Although no one can ever control the unfolding of a conflict, with war being akin to a “game of cards,” the little control that remains needs as much human agency as possible.
Recent war games used major AI models, such as Claude and ChatGPT, to understand how AI might react during a geopolitical crisis between two nuclear powers. It resulted in the “near-universal” use of nuclear weapons. This is beyond disturbing and makes the prospects of using AI in warfare even more worrisome. The courage and caution of policymakers and defense officials to avoid launching these weapons prevented nuclear Armageddon in past crises. To take this human agency away would be to increase the chances of miscalculation and war.
Using AI to make judgments about war and peace, the most dangerous and pressing issue of humankind, detaches responsibility from military officials, the Pentagon, and ultimately the president himself. This makes the Pentagon’s motive to remove Anthropic’s safeguards unacceptable in a nation where its most powerful leaders must be held responsible for their actions. And this accountability is needed more than ever in the realm of war and peace.