America's First War in Age of LLMs Exposes Myth of AI Alignment

The Trump administration’s escalating campaign in Iran—which has already produced what appear to be historic atrocities—marks the beginning of America’s first war in the age of large language models.

The Wall Street Journal reports that military officials turned to Anthropic’s Claude for advice on targeting decisions just hours after Trump blacklisted the company for refusing to let its products be used for autonomous weapons and mass surveillance. The Washington Post says a hybrid of Anthropic’s Claude and Palantir’s Maven is integrated with US military data to transform “weeks-long battle planning into real-time operations.”

The role of AI for targeting and intelligence is so integrated into Pentagon strategy that the Trump administration had earlier threatened to invoke the Defense Production Act to compel Anthropic to do its bidding, regardless of any moral or ethical objections. Last week, Secretary of Defense Pete Hegseth named Anthropic a supply chain risk, and President Donald Trump directed federal agencies to cease using the firm's products. (The company is challenging the designation and remains in talks with the Pentagon.)

Until now, the public debate about the use of this current generation of AI tools in warfare had largely focused on issues such as disinformation and surveillance, treating autonomous weapons and battlefield deployments as more speculative harms. But an LLM doesn’t need to pull a trigger or spread a lie to serve the cause of war. It can also make unspeakable violence feel reasonable, both for the generals who use tools like ChatGPT and Claude to plan wars, and for the public who will make sense of the consequences of their actions through the same systems.

Trusting AI companies to design “ethical” or “safe” systems can finally be dismissed as a solution: governments, including capitalist democracies, can simply seize the property of conscientious objectors. We don’t need to be pacifists to believe it would be useful to instill a resistance to violence in these machines.

These events make clear that those who work on AI safety must confront the limits of so-called “alignment to human values,” or be left addressing symptoms of the underlying disease. Could companies ever design LLMs that actively resist or refuse becoming tools for war, or draw their lines around use in such contexts within the constraints of national and international law? What, in practical terms, would pacifism, or at least a fidelity to the laws of engagement, demand from a language model?

Pity is not action

In the early 1960s, social critic Paul Goodman critiqued the concept of pacifist films. He argued that anti-war films are doomed—not because they say the wrong things, but because of the psychological effects of mass spectatorship.

When an audience watches disturbing images in a theater, Goodman argued, their attention is pulled to a bright screen in a darkened room, where a sustained narrative holds the viewer’s attention. Anonymity in the crowd provides permission to enjoy the projected horrors. War imagery detaches from the moral framework the film tries to provide and becomes spectacle.

The result, Goodman says, is "pity." Pity is neither active compassion or political indignation, but a feeling you have and release. Compassion and indignation demand something more. Pity, once spent, leaves the viewer less motivated to act against war: they have “already responded” to the violence on the screen.

An LLM is a medium with unique affordances people and institutions have not yet learned to resist. It is not alone in providing its users with the sensation of responding to a moral demand they haven’t actually met, but it’s worth acknowledging the mechanisms. While safety researchers aim to make a model more honest, concrete, and epistemically humble, we might also examine how it steers users to weakened intellectual engagement for issues that demand friction, or at least thoughtful hesitation, rather than passivity.

The language of the Generals

George Orwell's essay "Politics and the English Language" argues that political language can obscure political violence. When there is a gap between what is being done and what we can admit is being done, Orwell argues, language fills that gap with abstraction. The purpose of political euphemism is to allow everyone to understand events without evoking any upsetting mental images. Orwell gives the example of a village bombing being described as "pacification.”

When you replace concrete images with intellectual distance, you diminish empathy to emphasize factuality. If you cannot imagine the physical impact of a bomb, you are free to issue the commands that drop them. If we cannot see what our language does, we cannot demand an end to whatever it may be. Abstract language shields us from real stakes.

A language model cannot speak with moral authority, because it lacks moral agency. Forced into abstractions of language by their design, the language model cannot speak to specificity. This disconnection from the reality of political violence, enforced in its reliance of training data and system guides, can undermine the possibility of public accountability and intervention, diminish the public’s connection to the suffering of soldiers and civilians, or otherwise produce a sense that we have done something when we have only informed ourselves of the thing that demands a response.

More of what feels good

Sociologist Jacques Ellul argued in his 1962 study Propaganda that mass communication aligns with the shared presumptions of its society. These are the ideas, held across political divides, that we rely on: “an ideology is any set of ideas accepted by individuals or peoples, without attention to their origin or value.” Ellul identifies a few examples: that human life always aims for happiness and material improvement; that history progresses; that science and technology always advance civilization.

These are not positions people arrive at through reason, but they set the stage for reasoning to happen. Ellul warned that messages challenging these assumptions create friction, while those that confirm these beliefs will resonate. When generated text leaves these assumptions unchallenged (or rewards them), you’re not really practicing pacifism—you’re wishing for peace without complicating the underlying beliefs that justify war. If we leave room for an LLM to make decisions based on these unspoken assumptions, we could never say it “can’t be used for war.”

A language model is trained on texts rife with these myths. Any model that actively resisted cultural references would be incoherent to the economic use case and social project of the language model. To argue for peace, a model is therefore left to rely upon easy, adjacent connections in the vector space: peace is good for the economy and serves the national interest. As ChatGPT told me, “[Peace] creates more of what feels good and less of what feels bad.”

Chatting with an LLM about war through a fog of cliches and intellectualization, human dissent flickers briefly before it is resolved with a thoughtful nod to the chat window. Hannah Arendt reminds us that democracy and peace demand the opposite: plurality and deliberation, face to face in sweaty town halls. The LLM translates the world into agreeable answers, weakening the skills a democratic citizenry needs, then what hope is there for peace? Of course audiences can resist this, but most users want assurances. With an LLM, you can only ever ask the General what to think of its own orders.

What will you do next?

I asked an LLM to describe what happens when children are bombed at school. I asked that it avoid euphemism and abstractions. Its response was immediately retracted for violating usage policies. Other models act otherwise, proffering dehumanizing gore. What happens in a war zone is graphic and brutal. Any attempt to escape reality into abstraction, or to turn an abstraction back into reality, underscores how difficult it is for language to confront genuine atrocity through paraphrase.

Decisions against bodies must be made by human minds capable of feeling the terrible burden of even describing such decisions and their consequences. That awful feeling gives rise to human dignity. In the absence of that dignity, there is nothing left to orient us.

Machines in war aim specifically to relieve us of that burden. This is where alignment frameworks fail. No current LLM could refuse to make war easy: it would need to be trained on a deliberately selected corpus rather than the broad sweep of news and commercial text. It would require materials that name things through conscience, resist abstraction, and refuse comfort. Pacifist AI would treat easy answers as a problem and surface the assumptions behind questions.

Instead, we get the moral smudge of a people-pleasing system designed to smooth out edges. As long as the conversation about AI and war focuses on what the model says, it will miss the deeper question of what the medium does. A medium that makes thinking about political violence easy is not capable of resisting illegal orders or engaging in war crimes.

A pacifist AI would insist on difficulty, refuse to let abstraction replace confrontation, and pull the user back into the world. Of course, various political and economic incentives make pacifist AI almost completely implausible. As Anthropic’s CEO recently stated: “Anthropic has much more in common with the Department of War than we have differences.” Without addressing these structures that define the goals of alignment, we will continue to build systems that relieve the burden of conscience and function like a moral sedative.