Is Language the Horizon of Thought, or Its Scaffold? — Reference, Use, Reflection, and Normativity

philosophy of language · reference · use · normativity · inner speech · linguistic relativity · Exploration v0.2 · 2026-09-24

1. What did the linguistic turn actually turn?

The analytic “linguistic turn” need not be read as the ontological claim that the world is made by language. Its thinner core is methodological: instead of trying to grasp thought or concepts directly, philosophy can approach them through explicit and analysable linguistic structures.

In Frege, Russell, early Wittgenstein, and Carnap, the dominant movement was from ordinary surface form toward logical articulation.

With later Wittgenstein, Ryle, and Austin, attention shifted toward how expressions are actually used in practice.

This exploration adds a third question: might language be not merely an external expression of thought, but a device by which a finite agent converts its own cognitive states into reusable representations and feeds them back into further reasoning?

2. Ideal language, ordinary language, and normative reconstruction

Ideal-language philosophy seeks to map ordinary language into a more logically transparent system. Ordinary-language philosophy, by contrast, treats actual use as philosophically informative.

Sellars and Brandom fit neither category neatly. They begin from actual discursive practice but reconstruct it in terms of what licenses assertions, what commitments they incur, what follows from what, and how correction is possible.

The Sellars–Brandom line is therefore best treated as a third position: an abstraction of reasons, inference, and corrigibility from within actual practice, rather than mere description of ordinary speech.

3. Referring to an individual and applying a concept are different operations

A simple form such as F(a) already separates two functions. a belongs to singular reference and re-identification of an object; F belongs to predication and common classification across different objects.

The central question for singular reference is whether appearances across time or perspective concern the same individual.

Predicates such as “red,” “cat,” or “good” instead place distinct objects under one classificatory treatment.

Natural language often compresses individual identity, sameness of kind, common property, and mere similarity under words such as “same.” Philosophical analysis should keep them apart.

4. Can reference survive incomplete understanding?

Frege's Morning Star / Evening Star case shows that expressions with the same referent can differ in cognitive significance.

Externalist approaches after Kripke and Putnam emphasize that incomplete or partly false descriptions need not destroy reference to the same object or natural kind.

One can model the ancient case as two object files, each tracking Venus under different appearances, later discovered to concern one object. Representation can be partial while reference remains stable.

5. Object-side refinement and use-side refinement

For a term A, at least two kinds of inquiry should be distinguished.

These projects are independent. A use profile can be studied without assuming that a matching real property exists.

Quinean assent/dissent, Davidsonian radical interpretation, and Horwich-style use theories approach the use side in different ways, while Kripke–Putnam externalism leaves room for a referent whose nature outruns the descriptions shared by a linguistic community.

6. Actual use, idealized use, and truth

Use itself has layers. Actual speakers are affected by factual error, inconsistency, inferential mistakes, memory limits, and social bias.

U_actual concerns actual usage; U_idealized concerns usage under some information correction, reflection, or coherence improvement; P* concerns what actually holds on the world side. Carnapian explication, Dummettian competence and manifestation, and contemporary conceptual engineering all move beyond raw usage in different ways. But idealized human use need not be identical to truth.

7. How many kinds of linguistic “ought” are there?

Linguistic normativity is not one thing. At the weakest level there is descriptive regularity: members of G usually use A in a certain way. No ought follows yet.

There can then be an instrumental reason tied to communication.

If one wants to be understood, one has an instrumental reason to use expressions in ways the audience can recognize. Stronger theories such as Dummett's or Brandom's place a correct/incorrect distinction inside meaning itself.

Even so, semantic correctness does not automatically generate robust normative reason.

8. Sellars and Brandom — where does the space of reasons come from?

Sellars distinguishes language-entry transitions from perception, intralinguistic inference, and language-exit transitions into action. Concept use belongs not merely to causal response but to a practice of giving and asking for reasons. Brandom reconstructs this through commitment, entitlement, incompatibility, and deontic scorekeeping.

The relevant normativity is not an order to adopt an ideal language. It is the constitutive normativity internal to discursive practice: making one assertion can undertake commitments, expose one to requests for reasons, and constrain what else one may coherently claim.

This project can accept that internal structure while refusing to infer world-tracking success or R_n from it.

9. But are the premises even shared?

Once linguistic norms are located in a community, a problem appears immediately: do speakers really share the same inferential rules or premises? If content is determined by total inferential role, then when Inference_A ≠ Inference_B, are A and B disagreeing about one concept or merely using different meanings?

Brandom resists an I–we picture in which the community acts as a single super-subject. He instead emphasizes I–thou relations in which speakers keep score on one another's commitments and entitlements. Total premise-sharing is therefore unnecessary.

Yet the problem remains: how much overlap is required for genuine disagreement? And even if a community-internal standard of correct use is stable, the whole community may still be wrong about the world.

10. Is private language really impossible?

It is too strong to read Wittgenstein's private-language argument as showing that a language used by only one agent is impossible. The target is not merely a personal code or inner speech. It is a system whose application criterion depends on intrinsically private experience such that seeming correct cannot be distinguished from being correct.

A one-agent symbol system can still contain error if it can compare present with past use, generate predictions, track downstream success, retain memory, and correct itself. Recent LLM systems provide an artificial case in which intermediate representations can be used for later computation without being identical to a user-facing answer. This helps distinguish non-public working language from an in-principle uncorrectable private language.

11. What remains under maximal skepticism?

We can stress-test linguistic function under skepticism extending all the way toward solipsism.

If all three are suspended, ordinary accounts of public communication become uncertain. Yet if the present cognition is distinguishing A from not-A, manipulating same/different, or reusing a prior symbolic occurrence in a later cognitive step, then some difference, retention, and relation is constitutively involved in the present inference.

This does not place “language” or a symbolic sequence inside the Momentary Epistemic Minimum itself. The minimum remains the category-neutral present occurrence or constraint. Once comparison, reuse, or sequence is being described, we are already talking about relational conditions that make inference possible. Explaining those relations as linguistic or quasi-linguistic symbolic processing belongs to a thicker functional and physical model.

What survives maximal skepticism is therefore not a public language community or any particular linguistic format. If linguistic articulation appears to be carrying the present inference, its thinnest epistemic implication is that some relation enabling comparison is present; language is a higher-level candidate scaffold that realizes, stabilizes, and reuses such relations.

Even if interpersonal communication came first in language evolution, internalized language may later scaffold subsequent inference. That does not make language itself part of the epistemic minimum.

12. Reflection as cooperation among temporally separated cognitive states

A finite agent cannot transmit the full high-dimensional state of one moment directly to a later state. Let a cognitive state at one time be X_t1. Perception, affect, association, implicit relations, and competing hypotheses may have to be serialized into inner speech, writing, notation, diagrams, or concepts.

Here m ≠ X_t1. Linguistic articulation is often lossy serialization, not exact copying.

Yet lossiness is not merely a defect. Compression requires deciding what matters, what structure should be preserved, and what can be treated as one unit.

Turning a vague sense that “something is off” into “this is a problem about predicate boundaries rather than object identity” creates a new representation that itself becomes input to further reasoning. Functionally, this is state_t1 → representation → state_t2: cooperation among temporally separated cognitive states.

Within the ordinary world-model, treating both states as belonging to one diachronic self is an exceptionally strong explanation. But that subject identity is a thicker attribution than the information transfer itself. In the vocabulary of Transitions in Inheritance Systems and Variable Individuality, the thinner requirement is a directed successor relation from the earlier state to the later state; “the same self” can be modeled at a higher level as a particularly dense instance of that relation.

This temporal cooperation is not identical to cooperation with a simultaneously existing other agent. An earlier state can constrain a later one through memory and records, while the later state normally cannot causally answer the earlier one. Temporal division of cognitive labor therefore has a one-way asymmetry absent from ordinary reciprocal interaction.

13. Linguistic compression can also discard what matters

The same mechanism cuts in the opposite direction.

Compressing an experience into an existing label such as “anger,” “rationality,” “good,” or “consciousness” can erase subtle differences, context, unconceptualized features, and simultaneously live hypotheses. If that compressed output then guides later self-interpretation, a feedback loop appears.

Language is therefore not merely a medium that stores cognition. It is also a filter that helps generate the next cognitive state.

14. Is language thought itself, or a scaffold for thought?

No strong identity claim is needed. On a Language of Thought view, a non-natural-language representational system may precede verbal articulation. Natural language could then partly externalize or re-encode prior cognition.

The important weaker claim is that even if natural language is not all of thought, intermediate linguistic representations can be re-entered into cognition and thereby support serial, recursive reasoning that would otherwise be difficult. LLM chain-of-thought and scratchpad techniques are not identical to human inner speech, but they are useful cases showing that intermediate symbolic representations can play computational roles independently of interpersonal communication.

15. Weak linguistic relativity — what does a language compress?

If a language L maps a high-dimensional cognitive state X into C_L(X), weak linguistic relativity can be reconstructed as a difference between compression schemes.

This does not require strong Sapir–Whorf determinism. At least three levels can be separated.

• Encoding relativity: languages differ in what they habitually make explicit or omit during linguistic encoding.
• Recursive cognitive relativity: encoded representations are fed back into reflection and reasoning, so encoding differences can re-enter cognition.
• Habitual representational relativity: repeated loops may reshape habits of attention, categorization, or memory.

Slobin's “thinking for speaking” is closest to the first level. The present model extends the question to inner speech, where X_t → C_L(X_t) → X_t+1 can recur without any outward utterance.

16. Inventing a concept can mean inventing a new compressor

Introducing a new concept A can alter more than vocabulary. It can modify a cognitive compression scheme.

A structure that previously required several propositions may become manipulable as one reusable unit in working memory. Concept invention can therefore extend cognition through compression + abstraction + reusability.

But bad concepts can compress heterogeneous phenomena into a false unity. A useful concept may be understood not merely as a short label, but as a representation that achieves high compression while preserving structure relevant to inquiry.

17. Implications for “good” and value inquiry

The same distinctions apply directly to evaluative language. One inquiry concerns when a person or community actually uses “good.”

Another concerns how that use would change under some correction for factual error, inconsistency, or cognitive limitation.

V* is the separate question of what objective value structure, if any, actually exists. Even if strong moral error theory is true and V*=∅, research into U_G(good) remains. Conversely, even if realism is true, present human “good” may track only part of the true structure, conflate distinct structures, or discard important dimensions that current vocabulary cannot yet express.

Value inquiry therefore cannot be reduced to choosing among answers inside an already fixed conceptual vocabulary.

If the true value structure may lie outside not only our current theory set but our current representational space—V* ∉ H_t—then inventing new evaluative concepts, axes, and representational formats becomes part of inquiry itself.

Provisional conclusion — language is a scaffold, not necessarily a horizon

Language is not the world itself. Linguistic articulability does not entail linguistic constitution, and the fact that humans now perform sophisticated reasoning through language does not imply that every possible rational agent must share our format.

For finite agents, however, language can scaffold object re-identification, conceptual classification, preservation of inferential relations, reflection on one's own use, information transfer among temporally separated cognitive states, integration with external memory, and revision of the conceptual scheme itself.