Neither Player nor Tool: Large Language Models as Quasi-Participatory Linguistic Devices in Asymmetric Human–AI Language-Games

Large language models (LLMs) have made hallucination and alignment into two focal problems, usually treated as engineering defects to be fixed by more data, better architectures, and stronger alignment procedures. This paper offers a Wittgensteinian redescription on which both are better understood as structural: as two faces of a single condition in which language, once detached from the corrective practices that would otherwise constrain it, continues to run on its own momentum. Its central move, however, is not to apply Wittgenstein to LLMs but to ask what kind of thing an LLM is. Against the dichotomy of player versus tool, the paper argues that an LLM is neither a full participant in a language-game—it possesses no form of life, bears no consequences, and answers to no community that could correct it—nor a mere instrument like a database, since its outputs occupy the second-person response-position and are taken up as answers. It is, rather, a quasi-participatory linguistic device. On this basis, the paper reconstrues both problems. Model hallucination stands in a structural analogy—not a conceptual identity—to the grammatical illusion Wittgenstein diagnosed; the shared feature is not the word “illusion”/“hallucination” but the absence of a corrective practice. And the rule-following paradox, together with family resemblance, shows that final, complete alignment to a determinate human value-function is not merely difficult but conceptually confused. What a non-subjective device in the response-position changes is the human side of the game. The paper names this asymmetric structure co-play, and argues that alignment is therefore better understood not as fitting a model to a value-function but as governing the assertability of its outputs within an asymmetric human–AI language-game: a task realized in the design of interfaces, roles, evidence-marking, and abstention.