Uncovering our disagreement about representational explanation

After almost two years, and important step of our adversarial collaboration was to identify remaining disagreements. We developed a procedure to do so which we describe here. We started with the question:

How is representational content explanatory, and based on what notion of explanatory relevance?

We learned a lot from this discussion and were really surprised about where the actual disagreement lies.

Here is our AI-structured dialog of our 1hour discussion, in which--to not name names :)--PROPERTY discusses with DESCRIPTION.

1. The Challenge

PROPERTY: Let us begin with the problem that brought us here. Cognitive science routinely explains systems representationally. A perceptual state represents an object as red; a neural state carries information about orientation; an internal variable represents a spatial location. Critics then ask whether representational content is really doing explanatory work.

For a reductionist, there is a natural answer. Suppose representational property (R) is identical to some natural property (N)—perhaps an informational relation, a functional relation, a structural relation, or some combination of these. If we can show that (N) plays the relevant role in producing, constituting, or otherwise explaining the phenomenon, and if (R=N), then we have shown that the representational property plays that role. What more could reasonably be demanded?

DESCRIPTION: I agree with much of that, but not with your last sentence.

PROPERTY: Then where exactly do you disagree? You accept the reduction?

DESCRIPTION: Potentially, yes. Let us suppose for present purposes that the ontological reduction succeeds. Representations are not an additional ingredient in the furniture of the world. Whatever representational states there are, they are entirely realized by the physical, functional, informational, or structural organization of the system.

My question is different. I deny that an ontological reduction by itself establishes the explanatory significance of representational description.

PROPERTY: But if the representational property simply is the natural property, how could the first be explanatory while the second is not—or vice versa?

DESCRIPTION: Because explanation may be sensitive not only to what a description refers to but also to how the explanans is described.

That is the possibility we need to investigate.

2. Two Questions That Look Like One

PROPERTY: I think we should distinguish some things immediately. There is a perfectly intelligible question: Is property (P) explanatorily relevant to phenomenon (E)? In causal cases, for example, we can ask whether (P) is among the causally relevant properties of an event rather than merely accompanying it accidentally. What I reject is the thought that, having identified the relevant property, we must ask a further metaphysical question: Is (P) explanatorily relevant under description F but not under description G? If (F) and (G) genuinely pick out the same property, I do not see how the property could acquire or lose its relevance merely because we redescribe it.

DESCRIPTION: That formulation already makes your position clearer. You are primarily giving an account of the explanatory relevance of properties. I am asking about explanations. Those need not be the same question.

PROPERTY: But surely explanations are constrained by explanatory relevance. An explanation cannot correctly appeal to a property that has nothing to do with the explanandum.

DESCRIPTION: Agreed. The ontic relation is indispensable. What I deny is that it is sufficient. The relevant causal, constitutive, counterfactual, or mechanistic relations may be truth-makers for explanatory claims without being, by themselves, explanation-makers. An explanation is presented under some description, in response to some explanatory question, and within some epistemic context. Different descriptions of one and the same worldly item may therefore differ in explanatory value.

PROPERTY: Then we may have located the disagreement. I think the condition that a theory of representation needs to satisfy is fundamentally a condition on the property. You think it is at least partly a condition on the description.

DESCRIPTION: Yes. And that difference changes what counts as solving the problem of representational explanation.

3. The Reductionist's Test Case

PROPERTY: Consider a simple analogy. Suppose we ask whether temperature has causal or explanatory relevance. A successful reduction identifies temperature with mean kinetic energy. We then show that variations in mean kinetic energy make a difference to some outcome. Surely we do not need an additional demonstration that the very same property produces its effect under the description “temperature.” Once we establish: temperature is mean kinetic energy; and mean kinetic energy has the relevant causal role, we have vindicated temperature's causal role. To require a third step— show that mean kinetic energy acts qua temperature— looks like an artificial obstacle created by language.

DESCRIPTION: For causal relevance, perhaps. But you have moved quickly from causation to explanation.

PROPERTY: Deliberately. Causal explanation is the clearest case.

DESCRIPTION: And precisely for that reason it may hide the issue.

Let us grant that the same physical property is causally operative irrespective of what we call it. That does not yet show that every true description of that property yields the same explanation. Suppose somebody asks why a group of people took off their jackets. Saying that the temperature was high appears explanatory in a fairly direct way. A description couched in terms of molecular mean kinetic energy might be true and might identify the same underlying physical state. Yet there remains a question whether it explains the behaviour in the same way.

PROPERTY: Surely that is just a matter of appropriateness. One answer is easier to understand in the context. The other is more technical. But both can be perfectly true explanations.

DESCRIPTION: “Appropriateness” is precisely where the dispute starts. You classify the difference as pragmatic after you have already decided what the substantive explanatory relation is. I am not yet convinced that we are entitled to do that.

4. Superman Enters the Laboratory

DESCRIPTION: Consider a familiar type of case. Someone asks: Why did the victim survive the attack? You answer: Because Clark Kent was there. Suppose the questioner does not know that Clark Kent is Superman. The answer may fail to explain to that person why the victim survived. Answering—Because Superman was there—may succeed immediately. The referent is the same. The epistemic contribution is not.

PROPERTY: That shows that explanations as communicative acts can succeed or fail depending on what an audience knows. I have no wish to deny that. But it does not establish that there are two different causal properties. Superman's intervention does not become causally efficacious because someone describes the intervener as “Superman” rather than “Clark Kent.”

DESCRIPTION: Again, I agree concerning causation. That is why this debate should not be confused with a crude view according to which language somehow alters causal reality. Nothing about the physical process changes when we redescribe it. The claim is instead that explanation is hyper-intentional: substitution of co-referring expressions need not preserve explanatory value.

PROPERTY: And I can accept a version of that at the level of communicative success. My resistance concerns what follows for theories of representation. If a reductive theory tells us which natural property representational content is, and if that property plays a genuine explanatory role, why should the theory also be required to solve every problem about which vocabulary scientists should use to communicate that role?

DESCRIPTION: Because the scientific claim under dispute is not merely that a certain physical property matters. The claim is that representation explains. If representational vocabulary makes no difference to the explanation, it becomes unclear what that claim amounts to.

5. Labels, Descriptions, and Properties

PROPERTY: I worry that we are slipping between expressions, concepts, and properties. Suppose scientists discover that a toad responds to objects satisfying a complicated conjunction: roughly, dark objects elongated along the direction of movement. We introduce a new predicate, “worm-like,” for that complex property. Now imagine that the property is indeed causally relevant to snapping behaviour. It would be strange to say: The complex property is causally relevant, but being worm-like is not. We introduced “worm-like” precisely to denote that property.

DESCRIPTION: That case helps, but perhaps in the opposite way. If “worm-like” is merely a shorthand label for the longer description, then nothing interesting follows about genuinely alternative explanatory descriptions. The two expressions may simply package the same descriptive information. The more difficult case occurs when two descriptions differ in content while referring to the same worldly property.

PROPERTY: Fine. Let us grant different descriptive content. My argument does not require synonymy. Suppose predicates (F) and (G) have different meanings but genuinely refer to the same property (P). Then I want to say: (P) is explanatorily relevant. Whether we introduce (P) by (F) or by (G) does not change the truth of that claim.

DESCRIPTION: But notice what you have now established. You have established the invariance of a property-level claim. You have not yet established the invariance of explanations containing (F) and (G). That distinction may be the central one.

6. The Argument from Arbitrary Redescription

DESCRIPTION: Here is my main reason for thinking descriptions cannot be treated as explanatorily inert. For almost any scientifically relevant physical system, indefinitely many true descriptions can be generated. Some will correspond to familiar mechanistic variables. Some will describe topological organization. Some will describe dynamical properties. Some will describe information-processing or representational structure. And some will be bizarre. We could classify neural states according to the distance of the active neuron from the sun. We could construct predicates that combine scientifically relevant features with arbitrary features. We could invent enormously gerrymandered descriptions that happen to pick out exactly the same entities or events that participate in a genuine mechanism. Scientists do not treat all such descriptions alike. They repeatedly select certain kinds of descriptions as explanatory and ignore others. Why?

PROPERTY: One answer is that many of the strange descriptions fail to pick out the causally or constitutively relevant property in the first place.

DESCRIPTION: Sometimes. But stipulate that we have achieved co-reference. Otherwise, we never reach the issue. Let (F) and (G) differ in descriptive content while referring to exactly the same property. Suppose scientists systematically regard explanations formulated using (F) as illuminating and explanations formulated using (G) as defective. Your account tells us that the underlying property is equally relevant in both cases. I agree. But that does not explain the scientific selectivity.

PROPERTY: Why must a theory of explanatory relevance explain every such form of selectivity? Some of it may concern comprehensibility, salience, background knowledge, explanatory interests, or conventions of presentation.

DESCRIPTION: Because once all of those considerations are placed on the “merely pragmatic” side, we need to ask what remains of scientific explanation as scientists actually practice it. Mechanistic, dynamical, topological, computational, and representational explanations are not merely different strings of words attached to a common hidden relation. Their different forms appear to organize explanatory understanding differently. An account of scientific explanation should make sense of that fact.

7. What Does “Representation” Add?

DESCRIPTION: This becomes especially pressing in the case that actually matters to us. Suppose a reductionist account identifies representational property (R) with natural property (N). Then ask: What does describing (N) as representational add? There are two possibilities.

First, “representation” may contribute no explanatory content beyond whatever is already stated by the reductive description of (N). Second, representational description may contribute some further conceptual organization—for example, by characterizing a state as standing in for something, carrying content, being correct or incorrect, or relating a system to environmental conditions in a distinctive way. If the first is true, then representational vocabulary looks explanatorily dispensable. If the second is true, then we must ask whether that additional descriptive content does genuine explanatory work.

PROPERTY: I object to the move from “dispensable” to “eliminable.” Identity does not entail elimination. Once we discovered that water is H₂O, we did not need to stop talking about water. Indeed, accepting the identity gives us a reason to say that claims about water are claims about H₂O, not that water does not exist. Likewise, if representational property (R) is natural property (N), then showing (N) to be explanatory vindicates (R). It does not eliminate representation.

DESCRIPTION: Ontologically, agreed. But explanatory vindication is stronger than ontological permission to retain a word. The challenge is not: Are we still allowed to say “representation”? It is: Why is representational talk itself scientifically explanatory? If replacing representational vocabulary with its reductive naturalistic vocabulary loses nothing explanatorily, then representational explanation has been vindicated only in a relatively weak sense.

8. Water and H₂O

PROPERTY: Then water gives us exactly the case we need. Suppose some material reacts violently when water is poured onto it. If water is H₂O, then: “It reacted because water was poured onto it” and “It reacted because H₂O was poured onto it” identify the same relevant substance. Surely the identity transfers explanatory relevance.

DESCRIPTION: It transfers relevance of the substance. Whether it transfers the explanation unchanged is exactly what is disputed. Consider another case: “Why was the house destroyed? Because water flooded it.” Now substitute: “Because H₂O flooded it.” Something appears to have changed.

PROPERTY: Only because the second formulation sounds unnatural. If what matters is a sufficiently large quantity of H₂O striking the house with a certain velocity and pressure, we can formulate that perfectly well. The explanation remains true.

DESCRIPTION: But perhaps “sounds unnatural” is not an accidental feature here. “H₂O” directs attention toward chemical composition. “Water” in the flooding explanation directs attention toward a macroscopic fluid behaving in a mechanically destructive way. The contrast suggests that descriptions can select different explanatory features of the same stuff.

PROPERTY: Or it suggests only that conversational expectations differ. Once we formulate both statements with enough precision, their truth conditions converge.

DESCRIPTION: But explanatory equivalence is not simply sameness of truth conditions. That is the heart of my objection. A sentence can be true and still fail badly as an answer to a why-question.

PROPERTY: And I grant that possibility. I grant, more generally, that replacing one expression with a co-referring expression can make an explanation worse in various respects. What I resist is the inference: Therefore, the property is explanatorily relevant under one description but not under the other. That formulation conflates the explanation with the explanatory relevance of the property.

9. A Better Statement of the Disagreement

DESCRIPTION: Good. Then perhaps we can formulate our disagreement without forcing either side into an implausible position. You say: If (F) and (G) pick out the same property (P), then the claim that (P) is explanatorily relevant cannot be true when expressed using (F) and false when expressed using (G), merely because of the difference in description.

PROPERTY: Exactly.

DESCRIPTION: And I can concede that. I can say: (P) is explanatorily relevant. At the same time, I can hold: Explanations that characterize (P) as (F) may be good explanations, whereas explanations characterizing (P) as (G) may be bad explanations.

PROPERTY: That seems consistent.

DESCRIPTION: Then we should not ask vaguely whether “explanation” is description-sensitive. We should distinguish at least two claims:

Property Invariance: If (F) and (G) co-refer to property (P), then (P)'s causal, constitutive, or other objective relevance does not vary merely because (P) is called “F” rather than “G”.

Explanatory Hyper-Intensionality: Explanations containing (F) and (G) can nevertheless differ in explanatory quality even when (F) and (G) co-refer, because the descriptions may organize information differently and answer different explanatory demands.

PROPERTY: I am willing to accept both if the second is interpreted largely in pragmatic terms.

DESCRIPTION: And I resist calling all of the second merely pragmatic. So perhaps our deepest remaining disagreement concerns the location of the boundary between the constitutive conditions of explanation and the pragmatics of explaining.

10. Is This “Only Pragmatics”?

PROPERTY: Let me explain my hesitation. Imagine that somebody invents a new predicate (G) tomorrow, stipulating that it refers to exactly the same natural property currently denoted by scientifically respectable predicate (F). Nobody knows the new word. It is cumbersome. Its use confuses everyone. Of course explanations using (G) will be terrible. But surely we should not conclude that the property has stopped being explanatory when denoted by (G). That is why the distinction between property relevance and pragmatic presentation matters.

DESCRIPTION: Agreed again. But my argument does not depend on mere unfamiliarity. Scientific descriptions are systematically selective. Scientists do not simply prefer familiar words to unfamiliar ones. They construct models that abstract from some features and highlight others. They use mechanistic descriptions for some purposes, topological descriptions for others, computational descriptions for still others. Those choices appear responsive to explanatory structure. If your account says only: The underlying properties are objectively relevant; everything concerning which scientifically meaningful description reveals that relevance is pragmatics, then the account risks locating much of what philosophers of science want to understand about explanation outside explanation proper.

PROPERTY: That may be a consequence rather than a reductio. There can be objective explanatory relevance plus a rich pragmatics governing successful explanatory communication and modelling.

DESCRIPTION: Perhaps. But then we should say that explicitly rather than treating the two questions as though they were one. The interesting dispute becomes whether scientific model selection, abstraction, contrast, salience, and description belong merely to the pragmatics of conveying an independently constituted explanation or partly constitute explanatory relations themselves.

11. Ontology Is Not the Disagreement

PROPERTY: Another clarification is important. Your view does not require anti-reductionism about representation.

DESCRIPTION: Correct. I can be an ontological reductionist while denying explanatory reductionism in the relevant sense. Suppose representational states are nothing over and above natural states. I can accept all of the following:

(1) there are no irreducibly representational entities;

(2) the causal powers of representational states are exhausted by natural properties;

(3) objective causal or constitutive relations constrain correct explanations.

And I can still maintain:

(4) describing those states representationally may contribute something to an explanation that is not captured merely by identifying their ontic realizers.

PROPERTY: Whereas my position is that, at least for the compatibility problem we began with, point 4 is not a further burden that a reductionist theory of representation needs to discharge.

DESCRIPTION: Exactly. The disagreement therefore cannot be summarized as “physicalism versus anti-physicalism,” “reductionism versus non-reductionism,” or even straightforwardly as “ontic versus epistemic explanation.” We agree on too much ontology for those labels to locate the dispute adequately. The narrower question is what follows explanatorily once the reduction has been achieved.

12. Two Compatibility Challenges

DESCRIPTION: This now changes how we should formulate the compatibility challenge for representational content. Suppose we ask: How can representational content be compatible with mechanistic explanation? That question admits at least two readings.

Compatibility Challenge I: Property Compatibility - Can properties that constitute representational content—informational relations, functional relations, structural relations, or whatever the correct reduction identifies—be causally, constitutively, or otherwise mechanistically relevant? This is a question about properties and the organization of mechanisms.

PROPERTY: And this challenge arises completely independently of our disagreement about description. Even on my view, one must still show that the relational property proposed as representational content actually plays the relevant role. A proposed reduction does not automatically establish that. If content is identified with some informational-functional relation, for instance, it remains a substantive question whether that relation figures in the mechanism rather than merely accompanying it.

DESCRIPTION: Agreed. But there is a second challenge.

Compatibility Challenge II: Explanatory Relevance - Supposing that the representational property has passed the first test, why is a representational description of that property explanatorily appropriate? What does describing the mechanism representationally reveal, organize, or contribute that another description does not?

PROPERTY: And here I would say that the second challenge should not be built into the first. If a representational property has been shown to be explanatorily relevant, the core compatibility problem has been solved. Questions concerning representational vocabulary may remain, but they are additional questions.

DESCRIPTION: Whereas I would say that for anyone interested in representational explanation, rather than merely in whether the natural property identified with content has causal or constitutive relevance, the second question is part of the original problem. And that is not a minor terminological difference. It means we have been asking two different questions under one name.

13. Why This Matters for Representation

PROPERTY: Let me state the attraction of my view. The reductionist wants to avoid a mysterious requirement that properties exert causal or explanatory influence “qua represented,” “qua mental,” or “under a description.” Once we know what the representational property really is, we investigate whether that property is causally or constitutively relevant. This keeps the problem scientifically tractable. Otherwise, we risk demanding that nature somehow recognize our vocabulary.

DESCRIPTION: And let me state the attraction of mine. Representational explanation is interesting precisely because scientists do not merely say: Some physical process occurred and produced the behaviour. They characterize internal states as being about distal features, carrying information, misrepresenting, encoding variables, or standing in content-sensitive relations to one another. If all explanatory work is already performed by a non-representational property description, we should ask what epistemic work those representational characterizations perform. Otherwise, the theory may establish that the entity called “representation” is causally relevant without establishing that its being characterized as a representation helps explain anything.

PROPERTY: But the same could be said about water. Once water is identified with H₂O, the term “water” does not have to contribute some extra causal ingredient to remain legitimate.

DESCRIPTION: Certainly. The issue is not legitimacy. It is explanatory contribution. The question “Does water exist?” is different from “Why is describing this process in terms of water explanatorily useful?” Likewise, “Do representations exist?” and “Do representational explanations explain in virtue of something distinctive about their representational description?” are different questions.

14. What Would Decide Between Us?

PROPERTY: We should say what evidence could actually move the dispute. For my side, a powerful argument would show that the purported differences between co-referring explanations are fully captured by independent pragmatic factors—audience knowledge, conversational relevance, inferential convenience, familiarity, interests, and so forth. If those factors can explain why “water” is better than “H₂O” in one context, and why one scientific model is more useful than another, there is no need to postulate description-sensitive explanatory relevance.

DESCRIPTION: For my side, a powerful argument would show that the selection of descriptions is constrained in ways that cannot be reduced to such pragmatics. One possible route would be to show that explanatory descriptions systematically expose patterns of dependence, abstraction, counterfactual structure, modal information, or answers to contrastive questions that other co-referring descriptions fail to expose.

Then description-sensitivity would not be an accidental feature of human communication. It would be part of what makes something an explanation.

PROPERTY: But even then, I might say that the relevant structure has simply identified a more fine-grained property.

DESCRIPTION: And that response exposes another fault line: the ontology of properties. Whenever an explanatory distinction appears, the ontic theorist may attempt to reproduce it by individuating properties more finely. What I call different “explanatory descriptions,” you may call “references to different properties”.

PROPERTY: Sometimes correctly.

DESCRIPTION: Sometimes. But if every stable difference in explanatory behavior is absorbed into property individuation, the property view risks becoming difficult to distinguish empirically from the description-sensitive view. We therefore need independent criteria for property identity. Otherwise “same property” cannot function as an informative premise in our dispute.

15. A Provisional Agreement

PROPERTY: Despite all of this, we have more agreement than our initial confrontation suggested. We agree that explanations must answer to objective structures in the world. We agree that causation and constitution are property-sensitive. We agree that representational properties cannot acquire causal powers merely because we describe them representationally. We agree that co-referring expressions may differ dramatically in communicative or pragmatic usefulness. We agree that a theory identifying representational content with a natural property still owes us an account of whether that property actually plays a causal, constitutive, or mechanistic role.

DESCRIPTION: And we agree that the difference between us begins after those points. I maintain that an adequate theory of scientific explanation must also account for why particular descriptions and models are explanatory. The fact that a description refers to an objectively relevant property is necessary but not automatically sufficient. You maintain that this further demand should not be imposed on a theory whose ambition is to establish the explanatory relevance of representational properties. Once the relevant identity and property-level relevance are secured, representational content has passed the central test. Further differences among ways of describing that property belong elsewhere—perhaps largely to pragmatics.

PROPERTY: That is fair.

DESCRIPTION: Then our disagreement is no longer: Does the world matter, or do descriptions matter? Both matter. The disagreement is: Which features belong to explanation itself, and which belong to the use, presentation, or communication of an explanation whose objective basis has already been fixed?

PROPERTY: Yes. And perhaps a second disagreement remains: What exactly do we mean when we claim that representational content is “explanatorily relevant”?

16. The Adversarial Result

DESCRIPTION: An adversarial collaboration need not end with conversion. Sometimes progress consists in discovering that a single apparent disagreement contained two questions. At the beginning, we seemed to disagree over whether representational content could be explanatorily relevant. We can now distinguish:

Question 1. Can the natural property identified with representational content be causally, constitutively, or mechanistically relevant?

Question 2. If so, what makes representing the system as representational explanatorily appropriate, and what—if anything—does that description contribute?

PROPERTY: I regard the first as the central compatibility challenge. The second is an interesting but importantly different problem.

DESCRIPTION: I regard both as necessary for a full vindication of representational explanation. And there lies a genuine disagreement rather than a verbal one. If the first question alone is answered, we know that the worldly property to which representational discourse refers is explanatory. We do not yet know whether representational discourse explains because of what is representational about it.

PROPERTY: And I remain unconvinced that there is a further fact of that kind to be discovered. Once the property has been identified and its explanatory role secured, asking whether it explains “under the representational description” may simply reinstate a distinction that the successful reduction dissolved.

DESCRIPTION: Whereas I remain unconvinced that reduction of reference entails reduction of explanation. Two descriptions can reach the same place in the world while taking us there by different explanatory routes.

Our core disagreement:

PROPERTY: Once we know what representation is physically, and show that this physical property plays the relevant causal-explanatory role, representation has been explanatorily vindicated.

DESCRIPTION: Once we know what representation is, the explanatory question has only begun: we must still know what describing the system as representational contributes.