Critical Analysis of Language Models and Institutional Legitimacy

Language models are often evaluated through familiar technical criteria: accuracy, hallucination rates, robustness, safety, and bias. These measures are necessary, but they leave a broader institutional question unresolved. What happens when generated language enters settings in which wording itself carries authority?

In administrative, legal, financial, academic, and governmental environments, legitimacy is partly produced through recognizable forms. Reports, summaries, classifications, recommendations, notices, and procedural statements are not neutral containers. They signal competence, hierarchy, obligation, and permissible action. When language models begin producing those forms, they participate in the conditions through which authority is recognized.

Legitimacy through form

Institutional legitimacy rarely depends on content alone. It also depends on formatting, register, sequence, modality, attribution, and procedural consistency. A statement that appears in the correct institutional form can acquire force before its assumptions are examined. This is especially important in AI-mediated environments, where users may encounter outputs that resemble expert judgment without seeing the chain of decisions that produced them.

The effect is not that language models independently become legitimate authorities. Rather, institutions can lend legitimacy to machine-generated outputs by embedding them in procedures that already carry authority. Once this occurs, the distinction between generation and decision becomes politically significant. A model may only recommend, but an institution may treat that recommendation as the default basis for action.

The problem of displaced responsibility

AI-generated discourse can also complicate attribution. A consequential statement may be shaped by training data, model architecture, system instructions, organizational policy, user prompts, post-processing, and human approval. The final text may nevertheless appear as a singular output. This compression of authorship can make responsibility harder to trace precisely when institutional consequences become more significant.

That does not eliminate responsibility. It redistributes it. The analytical task is therefore to identify where responsibility remains recoverable and where formal design obscures it. Agentless constructions, abstract nouns, procedural language, and automated classifications are especially relevant because they can separate action from actor while preserving the force of the result.

From model critique to institutional analysis

A critical analysis of language models should not stop at the model boundary. The most consequential effects often emerge from the interaction between generated language and institutional procedure. The same output can be trivial in one context and authoritative in another. What changes is the structure in which the language is received, validated, and executed.

For this reason, AI governance requires a theory of discourse as well as a theory of computation. The central question is not whether machines possess legitimacy, but how institutions may confer legitimacy on machine-mediated forms and then allow those forms to shape decisions, visibility, and accountability.