Does AI language generation undermine human judgment and responsibility?

In "Owning Our Words: Sounding the Depths of Language," L. M. Sacasas (The Convivial Society, 2026-02-07) argues that the rise of LLM "language machines" forces a reckoning not just with what the technology can do but with what language itself is for — specifically, with three interrelated human capacities he says are now at stake: "the importance of human judgment, responsibility, and language." Sacasas frames these as "a useful set of lenses through which to consider the impact of artificial intelligence as it takes the form of a language machine to the degree that it undermines our capacity to judge well, encourages the evasion of responsibility, and outsources the vital labor of articulation." His starting premise, following Iris Murdoch, is that "words are the most subtle symbols which we possess and our human fabric depends on them," so handing over their production is not a neutral convenience.

Sacasas's reasoning works through borrowed authorities rather than original argument. He takes from Marilyn Chandler McIntyre the idea that language must be tended like soil, through sharpened reading, precise speech, and "poesis" (being "makers and doers of the word"); from Tolkien's refusal of the question "what makes you tick" ("I don't tick. I am not a machine") the point that metaphors "mediate our self-understanding"; and most heavily from Wendell Berry the claim that "the disintegration of communities and the disintegration of persons" tracks "the disintegration of language" over roughly a century and a half. Berry's dissected example — Nuclear Regulatory Commission officials who become "virtually languageless" when trying to describe meltdown risk to the public, their professional "objectivity" eliminating themselves as moral agents so that "public responsibility becomes public relations" — supplies Sacasas's clearest mechanism: specialized or evasive language doesn't just fail to communicate, it lets the speaker off the hook for judgment, which Sacasas links to Arendt's distinction between stupidity and an unwillingness to think.

This is a philosophy-of-language argument about AI, not an empirical one, and it sits closest to notes built from similarly non-empirical sources. It shares ground with Does language create subjects or express them? in treating language as constitutive of the speaker's self-understanding and moral standing rather than as a neutral tool for transmitting pre-formed thought — Sacasas's Tolkien anecdote and Murdoch's metaphor claim make the same instrumentality-denial from a cultural-criticism angle rather than a linguistic-theory one. It also extends Does polished AI output trick audiences into trusting it?: Berry's NRC officials show the pre-AI version of the same substitution, where correct-sounding, specialized diction stands in for the judgment and responsibility it should carry, which is the human precedent Sacasas worries AI-generated language will scale. And it presupposes the distinction drawn in Are language models and human speakers doing the same thing?: Sacasas's worry about "outsourcing" articulation only has force if producing language is itself an act that implicates the producer's judgment and responsibility, which is precisely what he and Berry say LLM output does not carry.

The excerpt is argumentative cultural criticism — a chain of borrowed quotations (Murdoch, McIntyre, Tolkien, Berry, Arendt) assembled into a thesis, with no new evidence about how LLMs actually affect judgment, responsibility, or articulation in practice. Sacasas does not show that users of language machines in fact judge less well, evade responsibility more, or lose facility with articulation; he asserts the risk and illustrates it with a pre-AI case (the NRC transcript) rather than an AI one. The implication he draws — that the two legitimate responses to language machines are critiquing their risks and relearning what is worth affirming about human language — is a call to a practice (reading, precise speech, "poesis") rather than a finding, and should be read as a normative program, not a measured effect.