America’s First War in the Age of LLMs: Debunking the AI Alignment Myth

When explicit images are replaced with abstract language, empathy diminishes in favor of factuality. If one cannot envision the physical repercussions of a bomb, they may feel liberated to issue commands to drop them. If we cannot acknowledge the impact of our language, we are incapable of demanding an end to its consequences. Abstract phrasing insulates us from the tangible stakes involved.

A language model cannot articulate moral authority because it lacks moral agency. Its design forces it into abstractions, preventing it from speaking to specific realities. This disconnect from the harsh truths of political violence—dictated by its training data and guidelines—can hinder public accountability and intervention, weaken connections to the suffering experienced by both soldiers and civilians, and engender a false sense of accomplishment when we have merely informed ourselves of events requiring a response.

More of what feels good

Sociologist Jacques Ellul, in his 1962 exploration Propaganda, argues that mass communication reflects the common assumptions held within society. These shared beliefs, embraced across political divides, are foundational ideas upon which reasoning is structured: “an ideology is any set of ideas accepted by individuals or groups without scrutiny to their origin or validity.” Ellul provides examples: the belief that human life strives for happiness and progress, that history is linear, and that scientific and technological advancements ultimately benefit civilization.

These beliefs are not conclusions arrived at through reasoned thought, yet they set the stage for rational discourse. Ellul cautioned that messages challenging these ingrained assumptions create friction, while messages supporting them resonate effortlessly. When generated text fails to confront these unspoken beliefs (or reinforces them), it does not constitute a genuine practice of pacifism; rather, it represents a desire for peace without interrogating the foundational ideas that justify conflict. If we allow an LLM the leeway to make decisions grounded in these silent assumptions, we cannot credibly claim it “cannot be used for war.”

An LLM learns from texts embedded with these myths. Any model resisting prevalent cultural references would appear incoherent to its intended economic and social functions. Therefore, advocating for peace becomes a challenge, relegated to superficial connections within vector space: peace is presented as beneficial for the economy and aligns with national interests. As ChatGPT noted, “[Peace] creates more of what feels good and less of what feels bad.”

Engaging an LLM about warfare through a haze of clichés and abstraction allows brief flickers of human dissent to dissipate as they yield to agreeable resolutions in a chat interface. Hannah Arendt reminds us that democracy and peace require the opposite: plurality and deliberation conducted face-to-face in demanding environments like town halls. If the LLM simplifies the complexities of the world into palatable responses, what hope remains for sustaining peace? Certainly, users can resist this trend, yet many seek reassurance. When interacting with an LLM, you can only ever inquire of the General how to interpret its commands.

What will you do next?

I posed a question to an LLM about the aftermath of children being bombed in a school and requested it to avoid euphemisms and abstract language. Its answer was promptly withdrawn for violating usage policies. Other models respond differently, sometimes presenting dehumanizing depictions of violence. The realities of conflict zones are graphic and harrowing. Efforts to evade reality through abstraction, or attempts to convert abstractions back into lived experiences, highlight the inherent difficulty in articulating genuine atrocities through paraphrase.

Decisions concerning human lives must be executed by minds capable of fully comprehending the grave consequences of such choices. This haunting awareness is what upholds human dignity. Without that dignity, we risk losing our moral anchor.

In warfare, machines aim to alleviate us from this weight. This is where frameworks for alignment fall short. No current LLM could resist facilitating war: that would necessitate targeted training on a specifically chosen corpus, rather than the expansive array of news and commercial texts. It would require resources documenting reality with conscience, rejecting abstraction, and denying comfort. A pacifist AI would confront simplistic answers and uncover the underlying assumptions shaping inquiries.

Instead, we encounter a morally blunted system, designed to pacify demands for complexity. As long as conversations about AI and warfare center solely on what the model says, we neglect the more profound inquiry of what the medium does. A platform that simplifies the contemplation of political violence to the point of ease is unqualified to resist unlawful orders or prevent war crimes.

A truly pacifist AI would insist upon difficulty, challenge the comfort of abstraction, and draw users back into the harsh realities of the world. However, various political and economic pressures render the development of pacifist AI largely implausible. As Anthropic’s CEO recently asserted: “Anthropic shares far more in common with the Department of War than our differences suggest.” Without addressing the structural factors that guide the objectives of alignment, we will persist in creating systems that alleviate the burden of conscience and function as moral anesthetics.