Language, Cognition & World Models: Why the Linguistic Layer Is Irreducible
The Inquiry: The AI world model discourse contains a paradox: LeCun and colleagues argue that text is a "crutch" and that animals build world models without language, while simultaneously describing world model components (planners, configurators, goal-setters) that require linguistically constituted structures to function in organizational domains. Is language merely a modality for world models (replaceable by better perception), or does it occupy an irreducible architectural position — one where certain world model ingredients can only be constituted through linguistic acts?
Falsifiable formulation: If the three world model ingredients claimed to be linguistically constituted (relational structure among governance roles, normative constraints, and declared organizational purpose) can be fully instantiated in organizational domains without any linguistic mechanism — without declarations, contracts, policies, commitments, or other speech acts — then language is not architecturally irreducible and the claim here does not hold.
The field's binary dissolves — but not into mush.
The AI world model discourse frames a false binary: language models vs. world models. LeCun argues world models should be learned from sensory experience, not text. The language model community argues that scaling text prediction produces emergent world understanding. Both positions contain insight and error.
The synthesis: language's role in world models is not uniform but ingredient-specific. For physical-world prediction (object permanence, spatial reasoning, dynamic simulation), language is unnecessary — animals demonstrate this, and LeCun is correct that text is not the optimal training signal. For organizational governance (authority delegation, constraint constitution, purpose declaration), language is irreducible — governance structures are linguistically constituted and cannot exist without speech acts.
The three-category mapping (F7) — irreducible, beneficial, unnecessary — provides the formal resolution. It does not disagree with JEPA for physical domains. It extends the world model concept to domains where institutional reality is linguistically constituted. The resolution is architectural, not polemical.
The Austin → Searle → Winograd & Flores lineage is the direct intellectual ancestor.
The speech act tradition provides the philosophical foundation. Austin (1962) established that some language constitutes rather than describes. Searle (1969, 1995, 2010) formalized institutional facts through "X counts as Y in C" and status function declarations. Winograd & Flores (1986) applied this directly to organizational computing through conversations for action. The lineage is direct: organizational governance is constituted through speech acts, and governance infrastructure must operate at the linguistic level where constitution occurs.
The mutable-institutional-reality argument resolves the evolutionary objection.
Animals have world models without language — correct for fixed governance (pack hierarchy, colony roles). Animals cannot change governance through declaration — this is exclusively human and requires language. Organizational governance is mutable — created, modified, and replaced through speech acts. Therefore language is irreducible for organizational world models. The argument is not that animals lack world models (they don't) but that mutable institutional reality requires linguistic constitution (it does).
LLMs are cognitive infrastructure, not cognitive replacement.
The extended mind thesis (Clark & Chalmers), distributed cognition (Hutchins), and the history of cognitive technology (Ong) establish that cognitive processes extend beyond the brain through tools. Writing extended memory. Printing extended distribution. Computing extended calculation. LLMs extend linguistic processing. They are the natural cognitive extension for the linguistic layer of world model operation — not replacing human governance cognition but amplifying it beyond individual bandwidth.
The AI world model discourse frames a false binary: language models versus world models. LeCun argues that animals build world models without language, therefore text is unnecessary. This research resolves the binary through a domain-sensitive analysis grounded in speech act theory. Austin, Searle, and Winograd & Flores establish that certain organizational structures — authority delegations, normative constraints, declared purpose — are linguistically constituted: they are brought into existence through speech acts, not described by them. A three-category mapping emerges: language is irreducible for three world model ingredients in organizational domains, beneficial for four, and unnecessary for three. The mutable-institutional-reality argument permanently resolves the evolutionary priority objection: animals have fixed governance (genetically channeled); only humans constitute and change governance through declaration. LLMs are cognitive infrastructure extending linguistic processing in the lineage of writing, printing, and computing.
Institutional facts — money, property, marriage, corporations, legal authority — exist only because of collective acceptance of systems of constitutive rules. — paraphrasing Searle (1995), The Construction of Social Reality