Two Senses of Sense-Making and the AI Case

Abstract (forthcoming talk)

Froese (2026) poses what he calls the “AI dilemma”: either frontier LLMs are capable of sense-making despite lacking biological embodiment, or the linguistic competence they exhibit does not require sense-making. Both disjuncts are his, and three responses are available. He takes it as forcing the first, on the basis of a “technologically-mediated embodiment” whose enabling properties he concedes are “in need of further theoretical analysis.” Thompson (2026) denies it force by denying its premise: there is no competence, and the fluency is “in the eye of the beholder.” I give the third. The dilemma is ill-posed, because one word in it is doing two jobs.

Froese is not confused, and what moves him should be stated before it is set aside. If systems that produce what looks like understanding are denied sense-making outright, then within the framework of Frank, Gleiser and Thompson (2024) lived experience is at risk of demotion — scaffolding that cognition needed on the way up and can now dispense with. That is a cost, not a mistake, and I discharge it at the end rather than pay it.

Neither response reaches the equivocation, and the stall is not local to this exchange. Independent literatures have been circling the same object and stopping in the same place: that what LLMs lack is basic rather than high (García-Valdecasas 2026; Barandiaran and Almendros 2025; Seth 2025); that no unified position occupies the system (Shanahan et al. 2023); that the meaning is at the human end (Bender and Koller 2020); that the products circulate without the capacity that made them (Harnad 2025; Farrell et al. 2025); and that communication was never transfer (Reddy 1979; Sperber and Wilson 1986). They agree on no thesis and mostly do not cite one another. Each names an absence, and none says what is absent.

The convergence is structural. These accounts describe cognition at two levels and not at the level between them — functional criteria such as autonomy and precariousness, and phenomenological description in terms of concern and significance — with nothing saying what an organism builds through encounter. A missing level does not leave a silence. It leaves a word doing too much work: “sense-making” has no referent of its own, so it drifts to both its neighbours.

Naming that level means looking at what separates basic minds from humans. Take a cat that has hunted. Mice look and move a certain way, call for a certain readiness, and carry a certain charge; across a history of hunts these settle into one configuration, activated as one event. Call that integrated preverbal residue a Pattern-Constellation — \(\{MOUSE\}\) for this cat, \(\{X\}\) in general — laid down through serial, concern-laden encounter from a persistent locus. Humans have \(\{X\}\) too, and language besides. Take the structure of the use of a word as a single speaker has acquired it, tied to encounters or not: the totality of that structure for “mouse” in that speaker I write \(\langle mouse \rangle\), and in general \(\langle x \rangle\), a Linguistic Pattern-Constellation. Both are individual, because both sediment one history — of encounter in the first case, of use in the second.

The two levels correspond to two different operations on sense. Biological sense-making (Bio-SM) is the bacterium moving up the gradient: nothing is manufactured, and the activity is its own point. Across the enactivist literature “sense-making” is close to a synonym for doing what is needed to survive and reproduce.

Making sense is the other operation. Manufacturing sense is an activity that generates a product, and the product — the sense — is a conclusion reached against everything one already holds. This is Cognitive making-sense (Cog-MS), and it is bound to language: its product is what language exists to transmit.

Both operations answer to something already laid down, and neither could answer to it without a who. In Bio-SM what is laid down is the history that makes some encounters matter and others not; nothing is held, and nothing is concluded. In Cog-MS what is laid down is a body of holdings, and the conclusion is measured against it. Holdings are owned. The standpoint is therefore not a condition added to sense-production. It is inside it.

What a cat does when it appears to infer is activate the constellations it has, so it manufactures nothing and stands, on this axis, with the bacterium. Basic minds do not manufacture sense. Nor does biological sense-making supply the standpoint that manufacturing requires: it presupposes a locus separated from the world, and contributes the concern that makes encounters selective.

A standpoint is therefore built, and built slowly — through encounter, not through language. Language is not sense either. Sense lives in the associations of Linguistic Pattern-Constellations, which are unique to each individual. Text, language’s medium, runs the other way: it makes the product portable, not by compressing it but by standing for it. In a conversation nothing called sense travels from speaker to hearer; there are only tokens that activate sense at the hearer’s end. What circulates in this paragraph is nobody’s \(\langle sense \rangle\) but a token standing for it, which in a competent reader activates his own. That communication is two acts rather than a transfer is not my claim but a century old, and mostly buried today. What the units add is why the second act can fail outright: reconstruction needs an accumulated \(\langle x \rangle\) to activate and a standpoint to conclude against, and a system may have neither.

An LLM has no \(\langle x \rangle\) of its own and no \(\{X\}\); it renders from the aggregate of everyone’s \(\langle x \rangle\), in which incompatible structures sit side by side with nothing to reconcile them.

So a conversation with an LLM is not a conversation with a human. Sense is reconstructed from tokens at the human end only, because there is no standpoint at the other. A prompt selects a region of the model rather than supplying a standpoint, and that selection is what produces the illusion of one. The illusion does not survive counting. The same model holds thousands of such regions at once, in thousands of simultaneous conversations, with nothing anywhere at which two of them would have to be reconciled. An LLM cannot have a standpoint because it has all of them at the same time.

None of this convicts the tradition that made sense-making central. In us Bio-SM and Cog-MS never come apart, and a dependence that never fails looks like an identity; that is what it was recorded as. As in REC, there is no Cog-MS in basic minds. Froese’s cost is discharged on the same ground: lived experience belongs to the producing side, so denying the models the biological side demotes nothing.

The result is that an LLM has neither Bio-SM nor Cog-MS. It is a third thing: the collective space of Linguistic Pattern-Constellations, which anyone who brings a standpoint can interrogate and no one inhabits.

One consequence reaches past the machines. Life–mind continuity holds that perspective comes free from accumulated adaptive regulation (Thompson 2007); the vector-grounding argument holds that grounding comes free from operations on already-grounded inputs (Coelho Mollo and Millière 2026). Both treat a property of a developmental history as closing under what is built on it. The sense-making equivocation is that error one level down, which makes the enactive debate and the AI debate one mistake at two altitudes.

Supply the level and the convergences above stop being separate findings and become consequences of one structure. A cat has biological sense-making and manufactures no sense, and nothing runs the other way. What LLMs lack is not at the top but at the bottom: they fail where the bacterium succeeds.