The Corrective Layer: Why LLM Use Presupposes the Six Commitments
The Corrective Layer: Why LLM Use Presupposes the Six Commitments
Theoretical foundations: Grant C. Sterling (Eastern Illinois University). Analysis and synthesis: Dave Kelly. Prose rendering: Claude (Anthropic). 2026.
The Problem Stated
A large language model produces fluent, structured, internally consistent output. It does so whether the output is correct or not. The dangerous failure is not the visible error — the garbled sentence, the broken format — but the coherent falsehood: output that reads as a genuine application of a governing framework while being a pattern-matched imitation of one. The two are superficially identical. No inspection of the output’s fluency, structure, or vocabulary distinguishes them, because the imitation is built from the same fluency, structure, and vocabulary as the genuine article.
It follows that anyone who uses an LLM for serious work faces a question that is not technical but philosophical: what makes correction possible? Something outside the model must check its output against the standard the output claims to meet. Call this the corrective layer. The question this essay addresses is what the corrective layer must be — and the answer is that its operations presuppose six classical philosophical commitments, and that their modern replacements not only fail to ground it but describe the machine that makes it necessary.
What Correction Requires
Consider what the human checker actually does when he catches a coherent falsehood in LLM output. He compares the output not to other output, and not to what sounds right, but to the framework as it actually is — to the source text, the defined procedure, the fact of the matter. That is a correspondence operation: it treats the output as a set of propositions that represent something beyond themselves, accurately or inaccurately (the Correspondence Theory of Truth). He recognizes, sometimes immediately, that a claim is wrong — a recognition that arrives as a direct grasp, prior to and often generative of the argument he then constructs to document it (Ethical Intuitionism, in its epistemic register: some truths are apprehended, not derived). He traces the error to its warrant, asking not whether the claim fits the surrounding claims but what it rests on, following the dependency chain down to what needs no further support (Foundationalism). He then refuses assent to the erroneous output — and this refusal must originate in him. If his verdict were merely the determined product of his own training, his own distribution of exposures and reinforcements, it would be one pattern overruling another with no standing to call itself correction (Libertarian Free Will). For that refusal to be his act and not an event that happens through him, the checker must be an agent distinct from the mechanisms he audits — a rational faculty that has the model’s output as its object rather than being one more output stream alongside it (Substance Dualism). And where the work carries normative weight — where the question is not only whether the output is accurate but whether producing or publishing it is right — the standard invoked must be an objective one, or “the output is wrong to release” reduces to “I would prefer otherwise” (Moral Realism).
Six operations; six commitments. Remove any one and the corresponding operation loses its object. This is not a claim that the checker must consciously affirm the commitments as doctrine. It is a claim that his practice presupposes them: every act of genuine correction is an exercise of the capabilities they name, and a person who denied them consistently would have no coherent account of what he was doing when he corrected anything.
The Replacements Describe the Machine
Now set the modern replacements beside the LLM, and something arresting appears. Physicalism holds that there is no rational faculty distinct from mechanism: the LLM satisfies this — it is the weights. Determinism holds that output is fixed by prior state plus input: the LLM satisfies this exactly. Coherentism holds that justification is fit among beliefs with no foundational layer: the LLM’s entire epistemic operation is fit with its training distribution — coherence with no floor. Deflationary and consensus theories of truth hold that “true” marks assertibility rather than correspondence: the model asserts what was assertible in its corpus. Constructivism about value holds that norms are built from preference and agreement: the model’s normative outputs are distilled human preference data. The large language model is not merely compatible with the replacement commitments. It is their first complete instantiation — the replacement picture of a mind, built and running.
This settles the question of whether the replacements could serve as the corrective layer. To audit a coherence engine with a coherentist standard is to run the same operation twice. Such a check can catch inconsistency — output that fails to fit — but the failure that matters is the coherent falsehood, which fits by construction. Detecting it requires stepping outside the web of mutually supporting text to something the text is about, and that step is precisely what the replacements deny exists. A second model checking the first does not escape this; it multiplies the distribution without transcending it. Under the replacements, there is no corrective layer available to anyone — not because humans lack the capabilities, but because the replacements assert that no one has them. The human reviewer, on that picture, is only another pattern-completion process with a different training history, and his “correction” is one distribution’s output displacing another’s, with no fact of the matter between them. The word correction survives; the concept does not.
The Order of the Argument
One guard must be stated. The argument is not that the commitments should be held because holding them makes LLM use go well. That inference is pragmatist — true-because-useful — and it concedes the correspondence theory in the act of defending it. The order runs the other way. The commitments are held because they are true. What the LLM era has done is expose their load-bearing role in a domain where it can no longer be ignored. For centuries a person could hold the replacements in the seminar room and live by the commitments at his desk, borrowing the capabilities his stated philosophy denied, and the loan went unnoticed. The LLM calls the loan. Here, for the first time, is a producer of language that actually is what the replacements say a mind is — and the immediate, universal, practical response is to insist that a human being check its work. Every organization that mandates human review of model output is affirming, in policy, what the replacement commitments deny in theory: that there exists a kind of agent who can do what the machine cannot — compare representation to reality, grasp truth directly, trace warrant to foundation, originate assent, stand outside the mechanism, and answer to an objective standard.
Conclusion
The practical asymmetry, restricted to LLM use, is therefore total. The user who holds the six commitments has a coherent account of his own role: he is the corrective layer, and his review is an exercise of real capabilities against a real standard. The user who holds the replacements can still perform review, but his own philosophy has evacuated it — under his commitments, the review is the same kind of process as the thing reviewed. He governs the model, if at all, on borrowed metaphysics. The six commitments are not one framework among others for working with these systems. They are the conditions under which working with them — rather than merely alongside them — is possible.
Theoretical foundations: Grant C. Sterling (Eastern Illinois University). Analysis and synthesis: Dave Kelly. Prose rendering: Claude (Anthropic). 2026.
0 Comments:
Post a Comment
<< Home