3.1. An example of illustration

Imagine a student in a physics class who expressed the following misunderstanding: “Explanations in quantum mechanics appeal to observers, so quantum mechanical effects rely on conscious observation. Therefore [some woo].” Though this is wrong, observation is a difficult and contested notion in physics, and it might be impossible to point the student to a passage in the textbook with a clear definition of the term. A better tactic would be to show the student that quantum mechanical (QM) explanations work, predict, and explain experimental results even if any potential conscious observers are, say, looking away from the experiments. If the student’s misunderstanding were to persist (“But we set up the device that does the observation, and we’re conscious”), you might respond with a more extreme hypothetical: a universe without conscious observers at all. We can imagine a physics lab popping into existence in that universe, arranged so that it sets into motion a classic QM experiment. Our QM explanations, models, and predictions, including everything they say about observation, would still succeed there, and in all the same ways they succeed in the actual world.

All we’re doing here is paring away an irrelevant feature (consciousness) to show that it is, in fact, irrelevant to our explanations. We’re hoping the student follows us in an inference that goes something like the following. If QM explanations go through just as well, and in all the same ways, whether consciousness does or doesn’t exist, then those explanations must not rely, for their success, on consciousness. And if observation is defined by the features it must have to serve those explanations, consciousness won’t be part of its definition.

Now, does it matter that consciousness actually exists? That the universe we’ve described, with a physics lab but no consciousness, is imaginary and fantastical? Such a universe is surely not an instance of the same worldly pattern as our own, which developed according to dynamical laws—not by the random coming-into-existence of complex and apparently goal-directed things like physics laboratories. A world with physics labs but no consciousness might even be impossible, on certain (niche) views of consciousness (Goff, Seager, and Allen-Hermanson Reference Goff, Seager, Allen-Hermanson and Edward2022; Tononi et al. Reference Tononi, Boly, Massimini and Koch2016). But none of this matters, because you’re not using the thought experiment, in the first place, as an example of the kind observation, then deriving the nature of the kind from that example. In the terminology I use, you’re using the thought experiment to illustrate a type of scientific explanation by paring away a feature to show that it is irrelevant to the success of that type of explanation.

To return to the main lesson of Millikan’s response to Swampman, note that in the QM case, we’re starting with actual scientific explanations, and we’re asking how they work in their actual contexts. It’s just that we answer this question by paring away features of those contexts and checking the consequences for the explanations’ success. That does involve a hypothetical nonconscious world, but we’re not starting from a claim that the nonconscious world is a member of some kind, nor are we saying that, because we would explain it in terms of some kind, it must be a member of that kind. Instead, we’re showing that our actual explanations would be just as successful, in all the same ways, if consciousness were absent. And we’re drawing the natural conclusion: The success of those explanations does not depend on the presence of consciousness. Kinds don’t even come into the picture until a further step. If we agree with the teleosemanticist that the nature of a scientific kind should be determined by its scientific role, the kind observation should not be defined in terms of consciousness, because consciousness is irrelevant to the kind’s role and the explanations it serves.

3.2. The logic of illustration

The logic of illustration is fundamentally no different than the logic you might use to understand how any system performs any task, from how primates develop emotional capacities (Harlow Reference Harlow1959) to how cooks make a good cacio e pepe (Bartolucci et al. Reference Bartolucci, Busiello, Ciarchi, Corticelli, Terlizzi, Olmeda, Revignas and Maria Schimmenti2025). Pare away parts, and ask whether the system is still able to perform the task to the same degree and in the same way. If it is, then the part you removed likely wasn’t contributing to the task in the first place. If you want to understand how a can opener works, you can determine that its color properties are irrelevant by paring them away (either really or imaginatively) and seeing whether you have impeded its ability to open cans. The same goes when the system is science rather than a kitchen utensil and the task is explanation rather than opening cans.

But I’ve aligned myself with the side of the debate that thinks we should take scientific explanation more seriously, and I don’t think anyone will be satisfied by a comparison between explanations and can openers. So I want to flesh out the logic of illustration by comparing it to two experimental paradigms in cognitive science. These comparisons will reinforce the point that illustrations don’t make problematic assumptions about kind membership. And they will support the legitimacy of illustration (at least insofar as the paradigms I compare it to are legitimate) as well as introducing some challenges that illustrations must face.

To set the stage for these comparisons, we can break down the logic of illustration into three components:

1. 1. There is a broad explanatory target. We’re trying to understand how a form of explanation (e.g., QM explanation) successfully explains events and patterns.

2. 2. We narrow this target to ask about a feature of the explanations (e.g., the consciousness of observing devices). Does that resource contribute to the explanations’ success?

3. 3. And we probe this question by paring that resource away (e.g., imagining an unconscious lab). What would happen to the success of the explanations if that resource were missing?

How does this compare to typical cases of cognitive scientific reasoning? One characterization of cognitive science is “the study of how agents perform tasks” (Mekik and Galang Reference Mekik and Michael Galang2022, 2), specifically how they use different resources to perform them—both internal resources (such as neural structures or activity patterns) and external ones (such as features of the environment).Footnote ⁸ To answer these questions, cognitive science routinely removes resources and examines the effect on task performance—just what I’m claiming illustrations do. In the rest of this section, I compare illustration to one cognitive scientific paradigm that removes internal resources and one that removes external resources.

Let’s start with the former. Ablation studies investigate the role of a brain area in some task by either ablating that area or finding organisms in whom it has been ablated naturally, for example, by a stroke or a railroad spike (Damásio et al. Reference Damásio, Grabowski, Frank, Galaburda and Damasio1994; Salvalaggio et al. Reference Salvalaggio, Grazia, Zorzi, de Schotten and Corbetta2020). Classical work found, among other things, that “bilateral lesions to lateral occipital–temporal cortex could lead to impairments in recognizing objects but no difficulty performing grasping and reaching movements to the same objects,” prompting the inference that the lesioned area was used in the former task, but not the latter (Mahon and Hickok Reference Mahon and Hickok2016, 942). Current work creates more carefully targeted ablations (Liu et al. Reference Liu, Dai, Feng, Peng, Qin, Kearns and Huang2019; Zhang et al. Reference Zhang, Kaye, Cai, Wang, Prescott and Liberles2021) or temporarily disrupts activity in a brain area, assuming that the area is incapacitated by the disruption (Weissman-Fogel and Granovsky Reference Weissman-Fogel and Granovsky2019). But the inferences have the same form as always. If an area or its activity can be eliminated without affecting task performance, it must not have been used to perform the task in the first place (von Eckardt Klein Reference Klein and Walker1977; Bickle, Mandik, and Landreth Reference Bickle, Mandik, Landreth and Edward2019, 34).Footnote ⁹ That logic comes with important caveats, which I’ll discuss shortly, but it bears comparison to the logic of illustration:

1. 1. There is a broad explanatory target. We’re trying to understand how an organism (e.g., a human being) successfully performs certain tasks.

2. 2. We narrow this target to ask about a feature of the organism (e.g., a particular brain area). Does that feature contribute to the organism’s performance?

3. 3. And we probe this question by paring that feature away (literally ablating it). What happens to the organism’s performance when that feature is missing?

I wanted to do two things with these comparisons. First, I wanted to make it clear that this logic does not require any illicit assumptions about the kinds our target organisms belong to. To see this, note what the experiments conclude. Just as an illustration probes the resources real scientific explanations use (not just the resources the imaginary one uses), ablation studies conclude that intact organisms of a certain species do or don’t use a brain area to perform a task (not just that the ablated organism does). So the conclusions aren’t based on an assumption that the ablated organisms are members of the kind we’re trying to draw conclusions about—ablated organisms are, by definition, not intact organisms. In fact, they often aren’t even the same species as the organisms we’re ultimately trying to understand (e.g., Weiskrantz et al. Reference Weiskrantz, Warrington, Sanders and Marshall1974, 709), and there are even attempts to draw conclusions about human brains from ablations in artificial neural networks (Lillian, Meyes, and Meisen Reference Lillian, Meyes and Meisen2018). Clearly the conclusions about intact organisms of a particular species do not derive from any assumption that the ablated system is also an intact organism of that species.Footnote ¹⁰ They derive, instead, from an assumption that the two organisms are performing the task the same way, that is, using the same brain areas or structures. So it would miss the point entirely to apply the real kinds response here and say, “The ablated organisms do not belong to the kind intact organisms, so we cannot draw any conclusions about intact organisms from them.”

The second thing I wanted to do was highlight challenges that apply to ablation experiments and maybe, by extension, illustrations. The main challenge to ablation experiments concerns the possibility that an organism learned to perform its task in a new way or that its brain somehow compensated for the missing area. That is, our assumption that the organisms perform their tasks the same way might be mistaken. Maybe the ablated area is used in the task, but it’s redundant. Or maybe another brain area took over its role in the ablated organism. The point is that similar task performance does not strictly entail similar resources used (e.g., Barsalou Reference Barsalou2016, 1128). I won’t rehearse the many examples of this here. Suffice it to say, ablation studies have a difficult problem. They have to make the case that the organism under study really is doing the task the same way as the target (intact) organisms. In practice, this means ruling out plausible learning or compensatory mechanisms, for example, by eliminating the necessary time for learning or plasticity or showing that brain areas that might be expected to compensate don’t actually change their activity patterns after the ablation. To make a plausible argument from an ablation study, you must do this convincingly, and the same goes for illustrations. If you use the illustration I described in the QM case, you need to rule out plausible ways the explanations might have succeeded even if they originally did rely on the consciousness of the observer. Perhaps some concept does extra explanatory work, or concepts change roles and contribute differently.

The other type of experiment I promised to discuss is one that pares away an organism’s external resources, namely, features of its environment. There is a particularly clear description of this experimental paradigm in a paper that teleosemanticists are familiar with, investigating prey capture in frogs (Lettvin et al. Reference Lettvin, Maturana, McCulloch and Pitts1959). The setup of that paper tells us that a frog “will starve to death surrounded by food if [the food] is not moving” (1940). Eliminate motion and prey capture is affected. So motion seems to be one of the environmental features prey capture depends on. But the frog “will leap to capture any object the size of an insect or worm, providing it moves like one. He can be fooled easily not only by a bit of dangled meat but by any moving small object” (1940). Eliminate any of the prey’s features aside from its rough physical dimensions and movement patterns and prey capture isn’t affected. So prey capture must not depend on those features. The logic is even more transparent here than it was with ablation:

1. 1. There is a broad explanatory target. We’re trying to understand how an organism (e.g., the frog) performs certain tasks.

2. 2. We narrow this target to ask whether the organism relies on a particular resource (e.g., a feature of the environment). Does that resource contribute to the organism’s performance?

3. 3. And we probe this question by paring that resource away (either stripping a feature from the environment or constructing an environment lacking the feature). What happens to the organism’s performance when that resource is missing?

Lettvin et al. used stimuli that had various features in common with flies, a typical prey item for the frog. And they recorded from a number of nerve fibers assumed to drive prey capture behavior to see which stimuli caused a change in their response and which didn’t. Eliminating all the physical features of a “fly” except its rough size and shape did not stop the fibers from responding, but eliminating its movement relative to a background did (1945). The point is that, assuming that these nerve fibers’ response is what drives prey capture, the physical features of a fly aside from its rough size, shape, and movement patterns do not contribute to the prey capture task.

To be fair, these experiments don’t exactly represent the state of the art. Neander (Reference Neander2017) brings out the increasing complexity of anuran prey capture research. But much of that research follows the same logic, with the essential finding being that “the relevant visual discrimination in an unconditioned toad is largely unaffected by features not captured by [a small number of] dummy stimuli”—the famous cardboard cutouts of worms and flies (104). That finding is so important because, by telling us which features can be eliminated with no effect on prey capture, it tells us which features are and aren’t used in that task.

How would the real kinds response fare against this body of work? About as well as it did against ablation studies. Here we’re using nonnaturalistic environments, including some wildly unrealistic ones—cardboard cutouts standing in for worms and flies—to probe the way that anuran prey capture works in real, natural environments. We might object that prey capture works differently in the two environments, but these objections would appeal to facts about how prey capture works in those environments, not about kind membership. They would not say, “Well, these are very nice experiments, but unfortunately, you’ve made a terrible error. The nonnaturalistic environment isn’t of a kind with the naturalistic one! Shame you spent so much time on this study; it’s worthless.” The reason that sounds so ridiculous is that the work in question puts no evidential or epistemic weight on the experimental stimuli (cardboard cutouts) being examples of the kind of stimuli we’re trying to draw conclusions about (real prey) but only on the fact that the frog is using the same resources in the nonnaturalistic environment as in the naturalistic one. I’m not trying to mock Millikan here; she didn’t give the real kinds response to this sort of experiment. The point is that she was right not to. It would have been a mistake to think that the logic of these studies—and, by extension, the logic of illustration—relied on the experimental situation (nonnaturalistic prey capture) being an example of the kind (naturalistic prey capture) that we’re ultimately drawing conclusions about.

This comparison raises complications similar to the ones we saw with ablation studies. It’s possible to argue that prey capture works differently in naturalistic and nonnaturalistic environments. Ecological psychologists have been making points like this for decades. In impoverished (experimental) environments, organisms may perform tasks differently than they do normally, when they have more environmental information to use (Baggs and Sanches de Oliveira Reference Baggs and Sanches de Oliveira2024; Gibson Reference Gibson, Jospeh and William1972; but see Shepard Reference Shepard1984). The task for the frog researcher is to show that the potential differences in how prey capture works are not actual.

For a different example, consider Hartle and Wilcox’s (Reference Hartle and Wilcox2016) study of stereopsis. They were investigating binocular cues for depth perception, but, chasing down some unexpected patterns in the data, they realized that their participants had discovered artifactual monocular cues in the experimental stimuli. So when Hartle and Wilcox found that depth perception was not affected by the removal of binocular cues, it was only because they hadn’t just removed binocular cues; they had added monocular cues for their subjects to use instead. That undermines any conclusion you might draw to the effect that binocular cues are not used in depth perception. This is just to say that if you pare away features of the environment to check the effects on task performance, and if you want to draw conclusions about how the task is performed in normal, nonexperimental settings, you need a plausible argument that the “paring” didn’t induce the organism to perform its task in a new way, for example, by accidentally introducing new resources or causing the organism to use different strategies or processing than it does in naturalistic environments. Likewise, illustrations will need to show that their manipulations don’t have unintended effects on the way scientific explanations work or introduce new resources for the explanations to use.

4.1. Swampman as an illustration

This argument will sound familiar. It’s supposed to. I’m trying to capture the argument that philosophers ought to be making when they talk about Swampman, which is different in small but essential ways from the argument that they tend to make and that teleosemanticists then respond to. (Hacohen’s Reference Hacohen2022 rendition of the argument is similar to the one that follows but does not draw out the structure of the argument as I do here.) Let Swampman be generated in a swamp just as normal. We can imagine him wandering into a university building and seeing posters advertising calls for participants. We can imagine that he signs up for a study and lands in a psychology or neuroscience lab. Because he is a molecule-for-molecule copy of Donald Davidson, he will display the same cognitive capacities as Davidson. So the first major step of the Swampman illustration is a simple disjunction: Swampman’s capacities are either explicable or inexplicable.

I need to rule out some easy responses on behalf of the teleosemanticist here. Swampman will display the same capacities as Davidson only insofar as those capacities are described without reference to history or to properties that rely on being the “real” Davidson. But this is not a significant limitation. Cognitive scientists might study the accuracy of Swampman’s and Davidson’s memories, but I doubt anyone would argue that what they mean by accurate is such that Davidson has the capacity to accurately recognize his mother but Swampman doesn’t because she didn’t actually give birth to him. Those kinds of historically inflected capacities might be interesting, but they are outside the remit of cognitive science and of typical representational explanations—my target here. So I take it as given that, as far as representational explanation is concerned, Davidson’s and Swampman’s cognitive abilities are the same, and the interesting role for selection history is as a resource to explain those abilities.

More specifically, when I consider the explicability of Swampman’s capacities, I am considering capacities that cognitive science paradigmatically explains in representational terms, in ways that teleosemanticists take to support the selectional notion of representation. This includes, for example, explanations of navigation (Shea Reference Shea2018, chap. 5), prey capture (Neander Reference Neander2017, chap. 5), and the communication of resource locations (Millikan Reference Millikan1984, chap. 5), in addition to memory. It’s those explanations, and their reliance on selection history, that stand to confirm or disconfirm teleosemantics. This will limit the generality of my conclusions. I will (thankfully) not be able to conclude that selection history is entirely irrelevant to cognitive science—just to representational explanation.Footnote ¹¹

This also means that it’s not Swampman’s history that’s up for explanation; we’re not asking whether the way Swampman came by his capacities is explicable. We’re asking whether cognitive science can or can’t show how Swampman’s physical organization supports capacities like navigation, memory, and prey capture—whether it can or can’t reverse engineer those capacities. What would it be to accept the “can’t” side of those disjunctions? If we were to sit Swampman down in the laboratory, would we be stymied-in-principle by his behavior? Would it be impossible to model the structure of his brain at the levels of grain that allow us to predict and explain his actions? Or would this project at least be less successful than it is with Davidson? There are no tricks up Swampman’s sleeve—he’s just another physical system. I don’t see any way of denying that his various capacities would be as explicable as the capacities of any system. For anything he can do, there must be an explanation of how he does it.

The next step is to ask how we would explain Swampman’s capacities. Would we have to use explanatory methods, strategies, or resources different than the ones we use to explain Davidson? I don’t see any way of supporting that view either. We would observe Swampman the same way we observe organisms whose evolutionary history is unknown to us. We would see patterns in his behavior: a tendency to forage in his environment in a certain way, an ability to learn patterns in sets of stimuli, and so on. And we would investigate those behavioral patterns with questionnaires and response-time measurements, black-box models and eye-tracking experiments, computer simulations and fMRI data, circuit diagrams and electrode recordings, and so on. That is, we would apply the same explanatory resources that we do in cognitive science more generally. And we would take the same approach to building models of Swampman—including the use of representational notions.

To demonstrate this, imagine a meta-experiment: a single-blind trial in which the participants are two cognitive neuroscience labs. We send Davidson off to one lab and Swampman to the other. But we’ve given Swampman a shower and some clothes that aren’t covered in swamp goo, so the scientists can’t tell who is Swampman and who is Davidson. If we think that the success of cognitive scientific explanations relies on their targets’ selection histories, we must think they would fail in some respect when applied to Swampman but would be successful in that respect when applied to Davidson. What respect could this be? The models we would construct would be just as predictive of Swampman’s and Davidson’s behavior, just as revealing of its neural basis, just as useful in medical interventions. Nothing about Swampman, his swamp-brain, or his swamp-engagement-with-his-environment would seem to impede those projects, any more than our typical ignorance of an organism’s evolutionary history impedes those projects (cf. Hacohen Reference Hacohen2022).

Again, I’m not talking about all projects cognitive science might have. A cognitive scientist might explain how an organism evolved a certain brain organization, and selection history is clearly relevant to that question. This is just to reiterate that what’s at issue here are representational explanations and the capacities they typically explain. There must be some goal the Davidson lab would achieve and the Swampman lab wouldn’t (or would only to a lesser degree) when it tried to explain, in representational terms, how Swampman’s physical organization supports tasks like navigation or memory or perceptual discrimination. To come to the point of the Swampman illustration, I’ve suggested that there is no reason to think that scientists in the Swampman lab would fail in any way that those in the Davidson lab wouldn’t. If that’s the case, representational explanations must not rely, for their success, on selection history, because with selection history “ablated,” they would work just as well, and in all the same ways, as they do normally. So if we agree with the teleosemanticist that our account of a scientific kind should be determined by its role in scientific explanation, we should not define representation in terms of selection history for the same reason that we should not define observation in terms of consciousness. Because there are no problematic assumptions here about kind membership (the logic of the argument is the logic described in section 3.2), the teleosemanticist’s stock objection, the real kinds response, doesn’t apply.

I’ve gone quickly, and it’s possible that I’ve missed some explanatory goals that wouldn’t be met when cognitive scientists explained Swampman (I return to this shortly). But my goal has just been to cast Swampman in a more compelling role than he normally plays and to show that the usual teleosemantic response doesn’t apply to him in this role. If representational explanations have goals that I haven’t discussed, and that expose a difference in explanatory success between Swampman and Davidson, then the teleosemanticist would need to describe those goals, argue that they are goals of representational explanation as it is actually practiced, and show how representational explanations of Swampman’s capacities would fall short of them. For now, I’m satisfied if I’ve got Swampman back on his feet.

Before I move on, what about the challenges we saw earlier for experimental paradigms with the same logic as illustration? First, we have to rule out compensation, where some other feature of Swampman or the lab explaining him does extra work when his selection history is absent. It’s not clear what this would be. Does the lab that receives Swampman have to put extra emphasis on behavioral as opposed to brain data? Will that lab make additional modeling assumptions? There doesn’t seem to be a plausible compensatory mechanism, especially because, in the hypothetical experiment, the two labs don’t know who got Swampman and who got Davidson. Second, we have to rule out any unintended effects of removing Swampman’s selection history, and especially any resources that might introduce. But there don’t seem to be any plausible worries here. Swampman will have had a very different day than Davidson, who woke up that morning in a bed, not a swamp. But that sort of difference doesn’t seem to provide resources that could fill explanatory gaps left by Swampman’s missing selection history or allow the Swampman lab to reach the same conclusions about their subject as the Davidson lab does by different means.Footnote ¹²

4.2. Other objections to Swampman

I want to show that understanding Swampman as an illustration, rather than as an example, helps deal with some teleosemantic objections aside from the real kinds response. First, teleosemanticists might accept that our explanations of Swampman work just as well as our explanations of Davidson, in all the same ways, but argue that some broader explanatory goal would be undermined if we were to define representation in nonselectional terms. For example, it is sometimes suggested—or, more accurately, stipulated—that cognitive science is in the business of generalizing over species kinds (Millikan Reference Millikan1996, 109) and maximizing the reliability (Millikan Reference Millikan1996, 108) or breadth (Neander Reference Neander1996, 123) of its generalizations. This, supposedly, makes selectional kinds necessary. But these are dubious characterizations of cognitive science as a whole and representational explanation in particular, which aim not only to generalize but also to model, problem solve, explain, and so on, as I noted in section 2. And anyway, it is not at all obvious that nonselectional kinds would fail to support the necessary generalizations and inductions—something that is acknowledged even by teleosemanticists (Shea Reference Shea2018, 22). So this appeal to broader goals is not a plausible objection unless teleosemanticists can (as I described at the end of section 4.1) argue for some conception of those goals and show how representational explanation would fall short of them if some or all of its target systems lacked selection histories.

Another common objection to Swampman comes from teleosemanticists who define selection to include learning, differential survival, and other short-term processes (e.g., Garson and Papineau Reference Garson and Papineau2019; Millikan Reference Millikan1984; Neander Reference Neander2017; Shea Reference Shea2018). These teleosemanticists think even Swampman has a selection history (his morning will include at least some differential survival and probably a bit of learning), so the Swampman illustration doesn’t show that representational explanations work just as well in the absence of selection history. If we understand Swampman as an example of a representational system, the teleosemanticist’s assent means he’s no longer a counterexample to teleosemantics; we have to ask whether Swampman has representations when short-term selection processes haven’t had a chance to act, if such a situation is even possible.

But if we understand Swampman as an illustration, we have other options. First, we can change the thought experiment so Swampman just has less of a selection history. This is easier than going back into lateral occipital–temporal cortex to ablate parts you failed to get on the first try. Just go up a couple pages and send Swampman to the lab sooner. Nothing in the thought experiment relied on significant stretches of time, so this shouldn’t change the results. And second, note that teleosemanticists who are enthusiastic about short-term selection history tend to accept that evolutionary selection is also relevant (e.g., Shea Reference Shea2018). A teleosemanticist who appeals to only short-term selection processes is as rare a sighting as Swampman. The short-termist objection, then, would be that although we have ablated the bulk of Swampman’s selection history, the remaining scraps (short-term selection) can serve representational explanation just as well as before, without any visible differences. This puts them in the unenviable position of having to explain why so much of their view is unnecessary. What are long-term selection processes doing in an account of representational explanation if their absence makes no difference to representational explanation? This is especially damning when combined with the first response, so all that remain are arbitrarily brief selection processes.

Another common response points out that Swampman is (I don’t know how to put this delicately) just a bit ridiculous. Leave aside the philosophical niceties; aren’t the scientists laughing at us (e.g., Dennett Reference Dennett1988, Reference Dennett1996; Millikan Reference Millikan1996)? It’s tempting to brush off things like Swampman with nothing more than, “Bah! We’re doing serious work here.” But, as the study of frog vision makes clear, legitimate scientific inquiries also put their target systems in ridiculous scenarios made up of fake organisms. (Or consider the virtual realities that laboratory mice and flies spend so much of their lives in.) Just as in regular scientific experimentation, what matters is not the realism of the environment but that the deviations from reality are motivated. As Douglas Mook (Reference Mook1983) puts it, one cannot just point out that an experimental setting is unrealistic and reject it as externally invalid.Footnote ¹³ To judge an experiment, there is simply “no alternative to thinking through, case by case, (a) what conclusion we want to draw and (b) whether the specifics of our sample or setting will prevent us from drawing it” (386). Cardboard cutouts are even less similar to flies than Swampman is to Davidson, but they are relevant to real prey detection because they are used in a chain of reasoning that is carefully designed to let us draw conclusions about real prey detection. For the same reasons, what matters is not Swampman’s realism but that his specifics allow us to draw the conclusions we’re drawing via the chain of reasoning described in section 3.

A more ecumenical response to the “Bah” objection might bring Swampman down to earth, using real organisms to make a similar point. We might invoke organisms with evolutionarily novel traits that haven’t had a chance to be selected for yet (Peacocke Reference Peacocke2014; Peters Reference Peters2014; Porter Reference Porter2020; Walsh and Ariew Reference Walsh and Ariew1996). The problem with these cases is that they can be nitpicked to death. Sure, the trait is evolutionarily new, but is there some broader selected mechanism that confers content on it? Might it have derived content? Could it have an evolutionary precursor of a similar enough kind for it to count as selected for? The advantage of Swampman is the same advantage cardboard cutouts have over real, immobilized insects. He affords us control over exactly the features we want to manipulate and keeps the investigation from being swamped by confounds.

A related worry is that the explanations of Swampman are successful only because we made him up to be as similar to Davidson as possible, allowing us to take advantage of preexisting explanations of Davidson. That seems like sleight of hand. And it may be, if we’re arguing first of all that Swampman’s states are examples of the kind representation. Then his surface similarity to Davidson seems to trick us into thinking of him or explaining him in representational terms, regardless of his status as a representational system. But if we are, first of all, illustrating representational explanation rather than defining the kind representation, the logic is entirely different. Consider the frogs again. For us to make inferences about real prey detection, the stimuli must be similar enough, except in the features of interest, that the frog applies the same prey capture mechanisms as it does for real prey. The whole point is to “trick” the frog into triggering those same mechanisms, using clever stimuli that isolate particular features and therefore allow us to make inferences about those features’ role in triggering the mechanisms. That’s not sleight of hand; it’s experimental design. So Swampman as an illustration is in no worse shape than a typical anuran vision experiment.

One last version of the “Bah” objection might dismiss Swampman because he’s a thought experiment, rather than a real experiment. Is it a problem that he’s imaginary? It would be, if that meant he only probed our intuitions or invoked some kind of “voodoo epistemology” (Sorensen Reference Sorensen1992, 27). But the reasoning that an illustration asks you to undertake is more like prediction than intuition. What would happen if Swampman walked into the lab? isn’t mere intuition-mongering any more than What would Mom do if she found out I got in a fight? or How would this organism behave in its Normal environment ( Reference Millikan Millikan 1984 )? As long as we can be reasonably confident in our predictions, these are unproblematic questions—even if Mom doesn’t find out and the situation remains imaginary.