Language Is Not Neutral Infrastructure: Why AI Power Begins in Form

We still tend to speak about artificial intelligence as if its political effects begin after the model has produced an answer. That is too late. Power is already present in the form of the answer itself: in what is named, what is omitted, who appears as an agent, what is rendered inevitable, and which alternatives are made to sound unreasonable before a reader has consciously evaluated them.

This is one of the central problems I have been pursuing across my recent work on language, authority, and artificial intelligence. The issue is not only whether a model is biased, factually wrong, or politically aligned. Those questions matter, but they do not reach the deeper level at which discourse begins to organize reality.

A system can produce a sentence that is technically balanced and still distribute agency unevenly. It can describe suffering while erasing perpetrators. It can present a decision as procedural necessity rather than as the outcome of human choices. It can turn power into grammar.

The disappearance of the speaker

In older models of authority, the speaker was usually visible. Governments ordered. Judges ruled. Editors selected. Administrators denied. The structure of responsibility was imperfect, but the institutional voice could often be located.

AI-mediated systems complicate this architecture because decisions increasingly arrive through impersonal formulations: content was flagged, risk was detected, a request could not be processed, a recommendation was generated, eligibility criteria were not met.

None of these sentences is empty. Each performs an action. Yet each can obscure the chain of agency behind that action. The interface becomes the visible surface while policy, model design, training data, institutional priorities, thresholds, and human decisions recede into the background. Authority remains operative even as authorship becomes diffuse.

Why grammar belongs inside the study of power

Political analysis often treats language as a vehicle that carries a decision already made elsewhere. I think that assumption is increasingly inadequate. In automated environments, linguistic form is part of the mechanism itself.

Passive constructions can suppress agency. Nominalizations can transform contested actions into abstract processes. Risk labels can convert uncertainty into administrative fact. Deontic language can narrow the range of legitimate responses without ever issuing a direct command.

These are not merely stylistic choices. When repeated across platforms, institutions, and automated systems, they become structural. A sentence can organize visibility. A category can determine access. A default can become policy in practice. A formulation that appears descriptive can quietly become executable.

From bias to executable legitimacy

This is why I believe the next stage of AI criticism must move beyond the language of bias alone. Bias asks whether a system treats groups differently. A theory of executable legitimacy asks a different question: how does a form acquire enough authority to shape conduct, distribute responsibility, and constrain alternatives?

The answer may lie not in any single decision-maker, but in recurrent formal patterns that institutions learn to trust and reproduce. The consequence is methodological as well as political. If power travels through form, then form can be audited.

We can measure agent deletion, responsibility loss, nominalization, procedural abstraction, deontic density, risk classification, and the linguistic conversion of judgment into necessity. These phenomena are observable. They can be compared across systems and domains. They can become evidence rather than impression.

The real problem is not that machines speak

Machines do not need to possess political authority in any philosophical sense. Institutions only need to let machine-generated forms participate in the production of decisions. Once that happens, grammar enters the architecture of governance.

The important question is no longer whether an AI system has power by itself, but whether institutions allow its outputs to become conditions of action. That distinction matters. It relocates the problem from speculation about machine intention to the observable mechanics of institutional language.

And it suggests a harder conclusion: the future politics of artificial intelligence may be decided not only by what systems know, but by the formal structures through which their statements become authoritative.