R2. Knowledge and the actual world, or, the truth

As we were at pains to point out in the target article, knowledge, unlike belief, is factive. While there are different ways to spell out the factivity condition on knowledge (see the commentary by Nagel), what is not controversial is that it ensures that you cannot represent others as knowing anything you take the actual world to preclude. Thus, there is some uncontroversial correspondence between your own understanding of the actual world and your attribution of knowledge to others. But what does this correspondence amount to?

A number of the commentaries propose that this basic form of theory of mind essentially involves your own understanding of the world (among others, see the commentaries by Durdevic & Krupenye, Tomasello, and Westra). On this view, when you represent others as knowing something, you represent their mind as being related to the actual world as you understand it. You yourself take the actual world to be some way, and representing someone as knowing something involves representing them as having the right kind of relation to that part of the world. The things that they know, then, are things involved in your own understanding of the world. As emphasized in the commentary by Tomasello, a form of theory of mind with this structure might be as simple as tracking whether other agents are acquainted with physical parts of the world, such that they know where the ball is, or know the woman who knitted mittens for your niece.

One alternative proposal takes a step back from the actual world and proposes that this basic form of theory of mind essentially involves monitoring whether others' understanding conflicts with your own (see, e.g., the commentary by Deschrijver). The actual world might happen to inform one's own understanding, but does not play a direct role, because knowledge attribution takes place at the level of tracking the relationship between two minds: your own mind and another's.

A third approach, which we might think of as the farthest from actuality, argues that this basic form of theory of mind, if it is a genuine form of theory of mind, must be meta-representational in the same way that belief is (Leslie, Reference Leslie1987). That is, it must essentially involve you representing someone else's independent representation of the world, and accordingly is not essentially concerned with the actual world as you understand it. With varying degrees of commitment, versions of this approach are discussed by Dudley & Kovács; Gordon; Kampis & Csibra, and Binmore.

So why do we think knowledge involves one's own understanding of the actual world? One reason is that if you don't think this, you don't have a natural way of accounting for why knowledge but not belief is factive. If attributing knowledge (but not belief) involves understanding others' minds in relation to the actual world, then the factivity of knowledge comes for free. Obviously, you cannot understand someone's mind as being related to some part of the actual world when you think the actual world contains no such part. If knowledge attribution is instead meta-representational in the same way belief is, then some extra explanation must be given for why this representation just happens to be limited by precisely the bounds of your own understanding of the actual world. It's not that you couldn't give such an account, but we can't see why you would want to. There's a much simpler explanation on offer.

This way of explaining the factivity of knowledge is importantly different from the version proposed by Nagel. Nagel proposes that the factivity constraint on knowledge is modal: Knowledge necessarily only binds agents to truths. The natural way of understanding Nagel's suggestion is that when people represent someone as knowing something, they understand the person's mental states as having this property (necessarily being true). From our perspective, the trouble with this approach is that the data suggest that knowledge attribution is unlikely to involve representing any such modal property. Nonhuman primates, for instance, seem to have a remarkable capacity to attribute knowledge (see sect. 4.1), but we'd be pretty shocked if they have the capacity to represent anything as being necessarily true. And if they don't have the capacity to represent anything as necessarily being true, then a fortiori they don't have the capacity to represent others' mental states as necessarily being true. So we can be pretty sure that if nonhuman primates do attribute knowledge, knowledge does not involve reasoning about which truths hold across possible worlds. And there's evidence that human adults aren't all that different (Turri, Reference Turri, Mizumoto, Stich and McCready2018). In fact, humans seem relatively happy to attribute knowledge in exactly the cases that philosophers designed to illustrate that knowledge cannot be attributed when such modal properties are violated (see, e.g., Colaço, Buckwalter, Stich, and Machery, Reference Colaço, Buckwalter, Stich and Machery2014, on fake barn intuitions and Turri, Reference Turri2016a on reliabilism). Such attributions are to be expected if knowledge concerns the actual world, not merely possible ones. So the first and perhaps most obvious reason to think knowledge involves the actual world is that this gives you a simple explanation of both why and how knowledge is factive.Footnote ¹

A second reason to find this approach promising is that it also explains why the capacity for knowledge representation is more basic than the capacity for belief representation (see the commentary by Westra for a related line of reasoning). If knowledge attribution involves understanding others' minds in relation to the world, then one can maintain a single representation of the world, parts of which others also know about. Belief, unlike knowledge, cannot involve one's own understanding of the world in the same way because beliefs can be false. Thus, belief, unlike knowledge, requires an independent representation of the world – the world merely as understood by another – which must be maintained separately from one's own understanding, and thus can be false.

Once again, another feature of knowledge – its comparative basicness – falls naturally out of our way of understanding knowledge attribution, while other approaches leave this feature unexplained. Consider, for example, the suggestion by Lassiter that the basic theory of mind capacity we presented may be better explained by representations of true belief rather than knowledge, or the suggestion from Sobel that the evidence may be better explained with the notion of prelief (i.e., representations that are understood to not be real, but also are not understood to be false, as in pretense). What remains perplexing is why attributions of true belief or prelief, which require the same resources as genuine belief representation, would show all of the signature features of a more basic cognitive capacity: emerging early in phylogeny, ontogeny, processing time, and may be processed automatically and persist in the face of other cognitive impairments.

Of course, one could go on to give some further explanation of why some additional difficulty emerges in cases of false beliefs in particular, for example, Deschrijver proposes a difficulty with conflict monitoring. However, such explanations face the challenging task of carving apart the cases that are genuinely difficult from cases that seem similarly complex but are not as difficult. Consider, for example, a recent piece of empirical evidence that demonstrates a nuanced capacity for attributing knowledge in monkeys (Horschler, Santos, & MacLean, Reference Horschler, Santos and MacLean2019). In this experiment, monkeys watched an experimenter who saw a piece of fruit move into one of two containers in a display in front of them. A screen then blocked the view of the experimenter and one of two things occurred. In half of the conditions, the fruit itself briefly moved out of the container and then back inside. In the other half of the conditions, the fruit remained where it was, but the container briefly moved off of the fruit and then back on it. In both cases, all objects had returned to the position where the experimenter had last seen them, and using looking time, researchers investigated whether the monkeys expected the experimenter to reach for the fruit where it was last seen. What Horschler and colleagues found was that when the box moved, monkeys continued to expect the experimenter to reach for the fruit where they had last seen it. However, when the fruit moved instead, monkeys no longer expected the experimenter to reach for the fruit where they had last seen it.

If your account proposes a difficulty with conflict monitoring (Deschrijver) or that one can only represent true beliefs (Lassiter), these results are worryingly hard to explain. Such accounts focus on representations that are independent from the actual world. Thus, when the experimenter doesn't know about them, things that happen in the actual world shouldn't change what the experimenter believes. Accordingly, the most natural prediction for such accounts is that in both conditions the experimenter will be represented as having a belief about the location of the fruit, and this belief will happen to be true – it will match the monkey's own ideas about the location of the fruit. So, in both conditions, monkeys should expect the experimenter to reach for the fruit where it actually is. But, of course, monkeys don't do that. Instead Horschler and colleagues found that monkeys only expect the experimenter to reach for the fruit when it was the container, rather than the fruit, that moved.

The difference between the conditions is easy enough to explain, however, if monkeys represent knowledge rather than belief. Knowledge requires more than having a justified true belief (sect. R2; Gettier, Reference Gettier1963). And so when the fruit moves but the experimenter doesn't see it (but then happens to return to the original location), the experimenter might end up with a true belief about the location of the fruit by coincidence, but they do not share the monkey's understanding of the location of the fruit. By contrast, when only the container moves, this should not affect the experimenter's knowledge of the fruit, and monkeys should continue expecting the experimenter to act in accordance with this knowledge. This is exactly what they do.

This is just one of a growing number of studies that demonstrate clear failures to represent others' true beliefs while simultaneously demonstrating clear success in representing their knowledge (see Horschler, Santos, & MacLean, Reference Horschler, Santos and MacLean2021; Krachun, Carpenter, Call, & Tomasello, Reference Krachun, Carpenter, Call and Tomasello2009). The key difference is that knowledge tasks can be passed by simply keep track of whether the agent understands the relevant part of the actual world, whereas the true belief tasks require you to construct a separate representation of the world as the agent understands it, which just happens to align with your own understanding, and thus is true.

Therefore, in short, if you can accept that knowledge concerns the actual world, you get a surprisingly simple explanation for why knowledge is basic, why it is factive, and how it differs from belief.

R3. But what do we know anyway?

Instead of locating the difference between knowledge and belief in the role of one's own understanding of the world, a number of commentaries argued that the essential difference between them concerns the kind of content they allow you to attribute. After all, as Tomasello and Starmans point out, human languages typically encode an intriguing difference between knowledge and belief. In English (as in many other languages), one can know ways of doing things and know the smell of summer rain, but one cannot believe ways of doing things or believe the smell of summer rain. Tomasello and Starmans argue that the content of belief attributions seems to be propositional, while the content of knowledge attributions can be both actual things in the world and abstract matters of fact.

Following their line of thought further, it wouldn't be surprising if there were different mechanisms for understanding, on the one hand, the kind of acquaintance other agents have to physical parts of the world and, on the other hand, their acquaintance with things like abstract propositional truths. Moreover, it is plausible enough that the mechanisms for figuring out what physical parts of the world another is acquainted with may be simpler than the mechanisms for figuring out which propositions another is acquainted with. And as Tomasello and Starmans point out, much of the evidence for basic knowledge ascriptions in nonhuman primates and human infants suggest that these populations represent others as knowing about physical objects or having certain skills. So perhaps all of this points to a key distinction in kinds of knowledge, with a basic form of knowledge attribution that amounts to little more than knowledge-by-acquaintance or know-how and differs sharply from belief attribution, and a separate more complex form of propositional knowledge ascription that is not more basic than belief ascription but is rather quite similar to it. On this view, the difference in basicness we illustrated in the target article is a matter of the basicness of the content attributed (propositional vs. non-propositional), and not truly a matter of the basicness of the attitude itself (knowledge vs. belief). This all seems quite convincing.

The trouble is that there actually seems to be a simpler explanation for why there isn't great evidence that nonhuman primates and human infants represent others as having knowledge of abstract propositions. Namely, there isn't great evidence that nonhuman primates and human infants represent abstract propositions, in general. If knowledge attributions essentially involve your own understanding of the world, then the kind of content one can represent others as knowing will depend on what kind of content your own representation of the actual world involves.

Moreover, the similarity between propositional and non-propositional content shouldn't be hard to see here. If you do not understand the actual world to involve any extraterrestrial aliens, you could not represent anyone as knowing them (“knowledge-by-acquaintance”). And if you do not think there are ways of turning water into gold bullion, you can't represent anyone as knowing how to do that (knowledge-how). And in just the same way, if you do not understand the actual world to involve abstract propositions, like “2 + 7 = 10” then you certainly will not be able to represent others as knowing this sort of thing either (“propositional knowledge”).

Nonhuman primates (and perhaps very young human infants) may not have the capacity to represent propositions, and thus their knowledge representations will necessarily be restricted to simpler forms of content, whether knowledge-by-acquaintance (Tomasello) or even just visual perspective (Asaba, Chuey, & Gweon [Asaba et al.]). And if this is right, then such knowledge representations are also likely to be guided by specific attention to cues such as eye gaze or direct perception (Dudley & Kovács; Grossmann & Dela Cruz; Kano & Call). However, for human adults who clearly can and do represent the world in something closer to propositional terms, the same capacity may be used to represent others as having knowledge of abstract truths. For example, as emphasized beautifully by Mikhail, human adults represent others as having moral and legal knowledge. Although unquestionably abstract, these rules make up part of our understanding of the world, and given that, we have no trouble representing others as sharing our understanding of them. Note that in the latter kinds of cases, we agree with Westra that there is reason to think the content of knowledge is propositional and with Farina & Lavazza who argue that knowledge is content-involving and representational.

Importantly although, even for unambiguously propositional content, attributing knowledge seems easier than attributing belief. One completely uncontroversial piece of evidence is that young children succeed at attributing propositional knowledge (e.g., “Sally does not know her marble is in the basket”) before they succeed in attributing similarly propositional beliefs (e.g., “Sally believes her marble is in the box”). Similarly, adults are faster to correctly attribute or deny knowledge claims than they are to correctly attribute or deny corresponding belief claims, even when the term used for knowledge is explicitly propositional, for example, “savoir” in French, which only takes propositional complements (Phillips, Knobe, Strickland, Armary, & Cushman, Reference Phillips, Knobe, Strickland, Armary and Cushman2018). A third piece of evidence comes from the commentary by Bricker, who used electroencephalogram (EEG) to show that propositional knowledge representation elicited a weaker P3b amplitude than belief representation (Bricker, Reference Bricker2020). Thus, even when knowledge attributions unambiguously involve propositional content, they continue to show signs of emerging earlier, being simpler, and requiring less processing than matched belief attributions.

So it turns out the surprisingly simple solution is that the mechanism for representing knowledge is just the same across all of these different kinds of cases – you are just figuring out what parts of the world someone else understands – and seeming differences in the complexity of knowledge attributions across species or development arise simply from the complexity of representing different parts of the actual world. (See Rosenbaum, Halilova, & Pathman for related commentary on the difference in complexity between episodic and semantic content in knowledge attribution.)

The upshot of our view is that knowledge attributions won't be limited to any particular type of content (propositional, knowledge-how, etc.). Knowledge attributions can be as rich as your own understanding of the world. It is for this reason that we suspect that the capacity for knowledge attribution we provided evidence for in the target article will be not be fully captured by approaches that place limits on the content of basic theory of mind attributions, for example, reducing it to representations of visual perspective (Asaba et al.), uninterrupted perceptual access (Dudley & Kovács), skill (Carpendale & Lewis), goals (Schlicht, Brandl, Esken, Glock, Newen, Perner, Poprawe, Schmidt, Strasser, & Wolf [Schlicht et al.]), episodic experience (Kampis & Csibra), or knowledge-by-acquaintance (Tomasello). While each of these commentaries does an excellent job of pointing to specific aspects of knowledge we can attribute to others, it would be quite surprising on each of these views if knowledge attribution just happened to work in much the same way in all the other cases as well. That is, each of these cases shares the signature features of knowledge attribution (sect. 2 of the target article). We don't think this is surprising though. Each of these cases involve various aspects of the actual world, and representing others as knowing that part of the world will work similarly in each case.

If you are wondering at this point whether we are really proposing that knowledge attribution may function in essentially the same way in nonhuman primates as it does in human infants and adults, let us be clear. We are. In fact, the commentary by Moss suggests that it might even extend to philosophers. Moss argues that the history of philosophy suggests that explicit theories of knowledge preceded those of belief in the Presocratics. As she argues, this suggests that explicitly theorizing about knowledge may be easier for creatures like us than explicit theorizing about belief. It may then be no coincidence that this empirical fact aligns with the other ones we reviewed in our target article and may provide yet another indicator that knowledge is more basic than belief for creatures like us – even those of us who are philosophers.

R4. Knowing what you don't know

A third objection that was touched on by a number of the commentaries was that theory of mind is for predicting and explaining behavior (see, e.g., Binmore, Dudley, Gordon & Kovács). This perspective makes sense if one is committed to belief being the most basic theory of mind representation. But, if we are right that knowledge is more basic than belief, then the trouble faced by this approach is that the more basic form of theory of mind seems oddly ill-designed for action prediction and explanation, in particular (see Bazhydai & Harris for a similar line of reasoning). One notable feature of knowledge representation is that it seems to require more than justified true belief. But of course, justified true belief should be more than sufficient if your goal is just to predict someone's actions. Even unjustified false beliefs will do. A second notable feature of knowledge representation is that it allows you to represent others as knowing more than you yourself know. Others know all sorts of things you don't. But just knowing that others know more than you doesn't do you much good if your primary goal is predicting what they are going to do or explaining why they did what they did. So it's odd that our theory of mind capacity would have these particular features if it primarily evolved for the purpose of predicting and explaining behavior.

In contrast, if knowledge attributions involve representing others as understanding the actual world, then the ability to represent others as knowing more than you isn't particularly puzzling. In fact, it's precisely what you'd expect. When you represent someone as knowing more than you, you represent them as knowing something about the actual world you do not. You probably do not know how to play the zither, but you do think that there are, in fact, ways of playing the zither. And if you didn't think there was a fact of the matter, you couldn't represent someone as knowing that fact. For example, those of us who don't think each person's soul weighs a certain amount can't represent others as knowing the amount each soul weighs. Further, in cases where you yourself don't exactly know something, but you have a pretty good idea about it, you have a correspondingly good idea of what it is that the other person knows. And when you yourself have a great idea about the relevant part of the world, you'll have a correspondingly great idea about the content of someone else's mind. If you know why you randomly assign participants to conditions in a controlled experiment, and you represent someone else as knowing why too, then you'll have a great idea of exactly what it is they know. Not only do you know the precise content of their mental states, but you'll be able to predict what they'll do, and explain why they did what they did.

So it's not that we don't think knowledge representations can be used for prediction and explanation or that these representations don't reflect the content of others thoughts, it's just that the traditional view gets it backward. Knowledge concerns the actual world and which parts of it others understand. In some cases, others' understanding of the actual world will align with yours, and in those cases, you will know the content of others' thoughts, and be able to predict and explain their behavior. But there are also cases in which others' know more about the actual world than you do. Our point is that your own representation of the actual world plays much the same role both when you attribute knowledge to another of some fact you do know and when you attribute knowledge of facts you do not know. In both cases, you are representing another as understanding some part of the actual world (the way that part of the world actually is). What is changing is simply your own understanding of that part of the world (see Durdevic & Krupenye for related discussion).

This proposal for how to understand others as knowing more than you (egocentric ignorance) differs in important ways from the suggestions raised in many of the commentaries. For comparison, consider the proposal by Tomasello that the basic form of knowledge involves only knowledge-by-acquaintance. On this view, nonhuman primates only represent others as having been acquainted (or not) with physical objects in the world. Following Kampis & Csibra, suppose that the mechanism here works by simply tagging which physical objects someone is acquainted with. As Kampis & Csibra point out, such a mechanism does not seem to allow for representations of egocentric ignorance. To make this concrete, consider the success apes have in representing conspecifics as knowing whether there is a piece of fruit in a given box even when they themselves do not know (e.g., Kaminski, Call, & Tomasello, Reference Kaminski, Call and Tomasello2008). In such cases, subjects don't actually represent there being a piece of fruit in the box, and thus it's hard to see how they could tag that object as having been acquainted with the relevant conspecific. What this example illustrates is the difficulty in accounting for egocentric ignorance faced by views that reduce knowledge representations to simple representations like acquaintance, tagging, visual perspective, or perceptual access. Even more perplexing is how this kind of knowledge representation could be extended to understanding others as knowing how to crack open a nut (as nonhuman primates do, Rapaport & Brown, Reference Rapaport and Brown2008) or knowing how to play the zither (as we humans do). What others know in such cases are not objects that can be tagged and clearly cannot be reduced to some particular visual perspective.

At the same time, our proposal for how to understand what happens when you represent others as knowing less than you (altercentric ignorance) also differs from those discussed in the commentaries. For example, Deschrijver suggests that altercentric ignorance may amount to simply attributing no representation whatsoever to another agent – much like the representation you attributed to the Prince of Liechtenstein before reading this sentence. But, just as it is possible to represent someone as sharing your knowledge of some particular part of the world (e.g., knowing a person) without representing them as sharing all your knowledge (e.g., knowing all the people you know of), it is possible to represent someone as not sharing your knowledge of a particular part of the world, without representing them as not sharing any of your knowledge. That is, the basic capacity for knowledge attribution allows for representations of knowledge and ignorance about specific parts of the world (this point provides a helpful contrast with the suggestion from Gordon that we may simply attribute all of our knowledge to others by default). In fact, much of the evidence we reviewed demonstrates precisely this kind of specificity in attributions of knowledge and ignorance. Consider simple studies in which nonhuman primates will selectively steal the piece of food that a dominant competitor does not know about (Hare, Call, Agnetta, & Tomasello, Reference Hare, Call, Agnetta and Tomasello2000). Success on these tasks requires that chimpanzees selectively represent the dominant competitor as ignorant of the existence of one piece of food while knowledgeable about the other. If they simply attributed no representation whatsoever to the other chimpanzee, they should be equally likely to take either piece of food (because the other chimpanzee would be equally unaware of both). When one represents others as ignorant, there must be specific parts of the world you do not represent them as knowing.

Importantly, the kind of ignorance we have been discussing does not involve representing someone else as being aware of their own ignorance (this would require a separate capacity involving meta-representation, see Durdevic & Krupenye). The difference is that when you represent another agent as being selectively ignorant about some part of the actual world, the predictions you'll make concern only the other parts of the actual world they do know about (you'd predict that they'd be upset about you eating the food they do know about, but not the food that they don't know about). But, if you are able to represent other agents as having some awareness of their own ignorance (knowing that they don't know), then the predictions you'll make may also concern the ignorance itself, and what other agent's might do to alter their ignorance. As emphasized by Royka & Jara-Ettinger, one possibility is that such a meta-representation requires a mind that can employ some kind of symbolic negation operator, allowing you to represent the agent as knowing that they do not know. Future work may want to explore this possibility.

We have been arguing for an understanding of knowledge ascription that is both rich and flexible in some ways and notably limited in others. However, both the richness and the limits arise from a single unassuming commitment: When one represents others as knowing or not knowing something, one represents them as knowing or not knowing something about the actual world.

R5. Give learning a try

This way of understanding knowledge fits seamlessly with our claim that knowledge is for learning. When you represent your friend as knowing how to ride a bike, even though you don't know how to, you take them to understand something about the actual world: a way in which bikes can be ridden. You are not particularly interested in their ideas about how to ride a bike, independent of whether they actually work; what you take them to know and what you want them to teach you is how to actually ride a bike.

A number of commentaries pushed back on this basic idea, arguing that the capacity for belief representation is a better candidate for underwriting learning from others, especially given the success of cultural evolution in humans in particular (Dudley, Gordon & Kovács; Richardson; Salazar; Sobel).

One form of this objection was succinctly put by Richardson, who argues that knowledge cannot be both what humans share with nonhuman primates and what explains humans' unique capacity for cultural accumulation. Stated this way, we couldn't agree more. While we do think that the capacity for knowledge attribution is likely shared with nonhuman primates, we agree that knowledge is not what explains humans' unique capacity for cultural learning. Rather, we suspect that humans' unique cultural accumulation of knowledge is instead explained by our unique representational capacities – perhaps the capacity for representing abstract propositions, encoding information linguistically, and so on. We believe these sorts of capacities, not the capacity for knowledge attribution, are what differentiates humans from other species. But of course, none of this means that knowledge attribution doesn't play a central role in the process of accumulating cultural knowledge. If knowledge attribution works the way we've been arguing, then changes in domain-general representational capacities will result in changes in what we can and do represent others as knowing, which in turn will change what we can learn from them. And so, while nonhuman primates may accumulate knowledge of foraging techniques (e.g., Musgrave et al., Reference Musgrave, Lonsdorf, Morgan, Prestipino, Bernstein-Kurtycz, Mundry and Sanz2020), human infants may accumulate knowledge of the names of novel objects (Bazhydai & Harris), and human adults may accumulate knowledge of math, all while using the same basic capacity for representing others as knowing something about the actual world.

While this response may help to address the differences in the content of cultural learning in nonhuman primates, there are clearly differences not only in content but also in frequency and tendency. We suspect that our proposal has little to contribute in explaining these differences. There are myriad ways in which humans are both more social and more successful in communicating than our primate relatives (see Henrich, Reference Henrich2015 for a discussion).

A second form of the objection that knowledge is for learning, raised by Sobel among others, is that the processes for selectively determining who to learn from may be better accounted for by a form of belief representation. In a helpful response, the commentary by Bazhydai & Harris provides a beautiful accounting of the empirical evidence that knowledge rather than belief representations support selective learning in infants. The body of work they discuss demonstrates that infants selectively learn from others who are knowledgeable and selectively pass on information to those who are ignorant, all while not yet demonstrating any real capacity for the kind of meta-representation required by belief. This literature similarly helps to address the point raised by both Handley-Miner & Young and Kampis & Csibra that if knowledge representations are going to be useful for social learning, they need to be accompanied by mechanisms for determining who actually knows what you want to know. The literature on trust in testimony provides remarkably thorough evidence for how these mechanisms may function, and we hope that this literature will become increasingly integrated into theory of mind research (see Bazhydai & Harris, Salazar, and Harris, Koenig, Corriveau, & Jaswal, Reference Harris, Koenig, Corriveau and Jaswal2018). On a related note, one would expect that what we choose to teach others is also guided by knowledge, and here again, there is a growing body of evidence that knowledge plays a key role in guiding the information we provide to others (Turri, Reference Turri2016b).

Finally, it may be worth being explicit that none of this means that belief attribution cannot also support social learning. However, if we are correct about the essential difference between knowledge and belief, then the cases in which belief attribution plays an essential role will be ones in which what you need to learn is something specifically about how others think, not how the world actually is. When the emperor wears no clothes, successfully predicting and coordinating with others certainly will require belief-based social learning. Still, we suspect such cases make up the periphery rather than the core of learning from others, especially in the course of primate evolution.

R6. What to do with belief?

Throughout, we have been arguing for a central way of understanding the differences between knowledge and belief attribution. An important separate question, which was raised in a number of commentaries, instead asks how these two forms of attribution may be related to one another (Bender & Gatewood, Brakel, Durdevic & Krupenye, Kano & Call, Nagel, Ninan).

As pointed out in the commentary by Nagel, the account of knowledge attribution we've given can occur entirely independently from belief attribution. Thus, our view differs in an important way from standard philosophical views of knowledge, according to which knowledge entails belief. It is important here to keep in mind the difference between the philosophical claim about the concept of knowledge and the psychological claim we have made (see the commentary by Gerken for a similar concern). We've argued that the representation of knowledge does not entail the representation of belief. This point is directly supported by the empirical evidence. For example, we've argued that monkeys can represent knowledge but not belief. And if that's right, it clearly can't be the case that their representing knowledge entails their representing belief. This same point is also supported by the growing experimental philosophy evidence for cases in which people will attribute knowledge but not belief (Myers-Schulz & Schwitzgebel, Reference Myers-Schulz and Schwitzgebel2013; Yuan & Kim, Reference Yuan and Kimforthcoming). Such cases can naturally be described as ones in which we take the agent to have access to the relevant part of the world even though this access doesn't exhibit the normal impact on the agent's thoughts or behavior (see the commentary by Brakel for related ideas).

One way of thinking about this independence between knowledge and belief aligns with the proposal from Kano & Call, according to which the capacities for knowledge and belief attribution are entirely separate. Kano & Call agree that human infants seem to have an ability to attribute knowledge but not belief. However, they are moved by the studies providing evidence for false-belief representation in apes (Kano, Krupenye, Hirata, Tomonaga, & Call, Reference Kano, Krupenye, Hirata, Tomonaga and Call2019; Krupenye, Kano, Hirata, Call, & Tomasello, Reference Krupenye, Kano, Hirata, Call and Tomasello2016; see also the commentary by Durdevic & Krupenye). Thus, as they argue, given that the two capacities do not consistently appear together, perhaps they should simply be understood as arising from separate systems. While we are less convinced that the existing research provides sufficient evidence for a capacity for belief representation in nonhuman primates (and monkeys in particular, see sect. 4.1), we can set this question aside for now. Note that if Kano and Call turn out to be correct that some nonhuman primates have a capacity for belief representation, there would remain a remarkably consistent pattern across species: One never finds a capacity to represent beliefs in the absence of a capacity to represent knowledge.

This consistent pattern suggests an alternative way of understanding the relation between knowledge and belief that aligns instead with the proposal from Ninan. As argued by Ninan, the capacity for belief representation may depend on a prior ability to represent knowledge. Following an idea from Williamson (Reference Williamson2002), Ninan suggests that instances of representing others as believing something may essentially be instances of representing someone as acting as if they knew something. If this is right, then belief attribution (even in cases where the belief is true), would require a form of counterfactual conditional reasoning. In other words, it would require representing a merely possible way the actual world could have been, and then taking the agent to be related to that world in much the same way we take others to be related to the actual world when they know things about it. Three features make Ninan's proposal intriguing. The first is that it could explain the general pattern whereby belief attribution appears later in development than knowledge attribution. The second is that it fits well with the empirical correspondence one finds in human development between counterfactual conditional reasoning and belief attribution (Riggs & Peterson, Reference Riggs, Peterson, Mitchell and Riggs2000). And the third is that it provides one way of understanding why there are cases in which knowledge does not entail belief, since knowing something does not entail acting as if one knew that thing (Myers-Schulz & Schwitzgebel, Reference Myers-Schulz and Schwitzgebel2013; Radford, Reference Radford1966).

While we are not yet sure whether belief representation should be understood as depending in some way on knowledge representation, this is clearly an important area for future research.

R7. One thousand flowers

There remains a great deal we do not know about the basic theory of mind capacity we have been concerned with. At least partially, this is because knowledge attribution has received comparatively less attention than belief attribution in the history of theory of mind research. So, while we agree with Dudley & Kovács, Kano & Call, and Richardson that our paper should probably not incite a wholesale abandonment of the study of belief attribution, we want to emphasize the range of commentaries that pointed to important new questions and future directions for the study of knowledge. We hope these questions spur a new era in theory of mind research.