Mind | DeepDyve

journal article

LitStream Collection

Why Free Will is Real, by Christian List

2009 Mind

1 Introduction Christian List’s crisp and confident new book, Why Free Will is Real, begins with its target clearly lined up in its sights: two epigraphs, from neuroscientist Sam Harris and biologist Jerry Coyne, each proclaiming that contemporary science leaves us with no option but to conclude that free will is an illusion. List’s aim in his book is to develop a naturalistic antidote to this persistent strain of scientistic free-will scepticism. The mistake the sceptics make, according to List, is to suppose that since the conditions on free action aren’t satisfied at the fundamental level, they aren’t satisfied at all. There is more to the world than just what is described by fundamental physics; insofar as the sceptics have failed to find the properties which ground free will, it’s only because they’re looking for them in the wrong place. ‘Free will and its prerequisites’, List concludes, ‘are higher-level phenomena’, and ‘no less real for that’ (p. 5, emphasis in original). We agree with List that free will is real, and that its constituents are not located at the microphysical ground floor. But we remain unconvinced, both by the account of free will presupposed in this book, and by List’s arguments that the relevant conditions are satisfied at, but only at, some ‘higher’ level. We take up these points in §§2-5, before concluding in §6 with some broader methodological reflections on the project in which List is engaged. 2 Defining Free Will List claims that his ‘primary interest in this book is free will as a capacity of an agent’ (p. 27), which he defines as ‘the capacity to act intentionally… to choose between alternative possibilities… [and] to control one’s action’ (p. 16). But a capacity is of little use without opportunities to exercise it; and on the face of it, a free-will sceptic could happily grant that we have the capacity to choose between alternative possibilities while insisting that we never actually have alternative possibilities of the relevant kind to choose between. For this reason, we think it’s better to regard List’s book as a defence of the reality of free action, which List defines in chapter 1 as follows: [A]n action is ‘freely performed’ if and only if it is intentional – that is, appropriately supported by the agent’s intentions; it is possible for the agent to do otherwise; and the action is under the agent’s causal control. (p. 28) List claims – without argument – that these conditions capture what it is to have free will ‘conventionally understood’ (p. 26, emphasis in original), in the sense presupposed by attributions of agency, deliberation and moral responsibility (p. 27). But there are large literatures challenging the necessity of each of these conditions. For example, subconscious actions, like drumming one’s fingers on the table, can seemingly be free without being intentional (Anscombe 1957); and non-causalists in the philosophy of action would presumably deny (on the basis of the possibility of deviant causal chains, for example) that acting freely requires any kind of causal connection between an agent’s intentions and their action. But most seriously, there is an extremely well-trodden path, since Frankfurt (1969), of denying that free actions are always such that the agent could have failed to perform them. Suppose that Lizzy has a chip in her brain which will intervene to ensure that she votes Democrat should she show any signs of wanting to do otherwise; but as it happens Lizzy decides to vote Democrat, does so, and all the while the chip lies dormant. Intuitively, Lizzy could not have done otherwise (the chip would have intervened); nevertheless, she acted freely. Many philosophers have concluded from such cases that acting freely is not even partially grounded in whether the agent could have acted otherwise, but instead depends on, for example, the degree of fit between the agent’s first- and second-order attitudes (Frankfurt 1971; Watson 1975), the agent’s degree of sensitivity to reasons (Fischer and Ravizza 1998; Sartorio 2016), the presence or absence of manipulation (either of the agent or of their environment) (Mele 1995), or some other property of the actual sequence of events leading up to the action. List does acknowledge the existence of such alternative, (as he puts it) ‘watered-down’ conceptions of free will. He nevertheless sees pragmatic merit in sticking with the so-called ‘maximalist’ definition outlined above, the suggestion apparently being that showing that free will exists on his definition would suffice to show that free will exists on any plausible alternative definition (p. 27). If that is the suggestion, however, it is mistaken – that an agent could, in some sense, have acted otherwise doesn’t imply that they satisfy any of the alternative compatibilist conditions on free will outlined above. Indeed there are several, potentially more troubling challenges to the reality of free will, which a different book of the same title might well have focused on, but which don’t feature at all in List’s book. Do not expect to see, for instance, discussion of how far it matters to freedom that an agent is not ultimately responsible for coming to be the way she is (Sartorio 2016, §5; Kment 2017), or of how far it matters that we might be less sensitive to reasons than we would perhaps like to think (Nelkin 2005; Brink 2013; Sartorio 2018). Perhaps it’s too much to expect of a defence of the reality of free will that it also answer the question of what free will is. But it’s worth pointing out that the two questions are intimately connected. By defining free will as he does, List risks defending a straw man. 3 Intentional States In chapter 3, List argues that human agents have intentional states, which are not ‘reducible’ to physical states. Together these conclusions are taken to vindicate List’s claim that his first condition on free will is satisfied, but only at some ‘higher’ level. List’s argument that humans have intentional states takes the following, familiar form: Behaviour (including human behaviour) is best explained in intentional terms, rather than in physical terms; If a system’s behaviour is best explained in intentional terms, then it has intentional states; Therefore, humans have intentional states. In defence of premise 1, List asks us to imagine what it would take to explain in microphysical terms why a dog ran into the kitchen when a bowl of water was put down. Not only would this be exceptionally difficult, such an explanation would also be profoundly unilluminating, because ‘[a] good scientific explanation should not only describe the things that actually happen but also give us some indication as to what would happen in different circumstances’ (p. 61). There are various different microphysical states of the system of dog+environment that would have led to the dog running into the kitchen; simply describing the actual such state fails to capture what they have in common, in virtue of which they all lead to the effect we want to explain. By contrast, explaining the effect in intentional terms, as the most rational response given the dog’s beliefs and desires (in particular, its desire for water), allows us to ‘identify the most salient regularities underlying the dog’s behaviour’ (p. 61), thus making for a better explanation overall. Not all physical explanations are microphysical explanations, however. To use one of List’s own examples, we can explain why a glass broke either by citing the microphysical state of the liquid inside it, or by citing its temperature; both are physical explanations, but the latter is clearly better, in List’s sense of identifying a broader range of regularities with the effect we want to explain. Consider, then, an explanation of the dog’s behaviour in terms of some coarsely-specified pattern of neural activation. This state is also realisable by many microphysical states of the dog, all of which lead to the dog running into the kitchen. Which explanation is better: the explanation citing the dog’s intentional states, or the explanation citing this coarse-grained state of the dog’s brain? A familiar argument for the latter turns on the fact that intentional states are apparently broad – whether the dog desires water, for example, depends on whether the watery stuff in its environment is in fact water. Thus, the (coarse-grained) brain state is arguably better correlated across possible worlds with the behaviour to be explained than the intentional state: the dog would have rushed forward even if the watery stuff had been XYZ (and so it had desired XYZ, not water). There have been various attempts to respond to this argument (see, for example, Yablo 1997; Williamson 2006); it nevertheless serves to illustrate that one cannot infer from the mere fact that intentional explanations of behaviour are better than microphysical explanations of behaviour that behaviour is best explained in intentional terms. List’s argument that psychological states are not ‘reducible’ to physical states begins with the observation that, since psychological properties are ‘multiply realizable’, any physical predicate that is co-intensional with some psychological predicate would have to be, at best, highly disjunctive. Moreover, this disjunctively specified physical property cannot be identical to the psychological property, List argues, because intentional predicates have explanatory functions that their disjunctive physical counterparts lack (in particular, they can be used in rationalizing explanations of behaviour). But it’s not clear why it should follow from this difference between psychological and physical predicates that they cannot express the very same property. This point is somewhat obscured by List’s discussion, which often conflates predicates with the properties they express. Consider, for example, his claim that a higher-level property is ‘reducible to’ a lower-level property only if ‘the lower-level property can serve as a substitute for the higher-level property in scientific discourse’ (p. 66). Well, what scientific discourse contains are predicates, not properties; we can substitute a higher-level predicate in scientific discourse for a lower-level one, but whether that amounts to talking about a different property is precisely the question at issue. List’s argument thus seems to presuppose a rather fine-grained theory of property individuation, which the identity theorist is not given much reason to accept. 4 Alternative Possibilities In chapter 4, List argues that there’s a sense in which it is possible for free agents to act otherwise than they actually do, even if determinism is true. Consider the following argument for incompatibilism: Free will requires that, at least in relevant situations, more than one course of action be possible; Determinism implies that, in any situation, only one course of action is ever possible; Therefore, free will and determinism are incompatible. (p. 86) List claims that this argument trades on an equivocation. Let an action be physically possible if its occurrence is compossible with the physical laws and physical state of the universe at some earlier time, and let an action be agentially possible if its occurrence is compossible with the psychological laws and the psychological state of every agent at some earlier time. According to List, if ‘possible’ means physically possible, then (2) is true, but (1) is false; whereas, if it means agentially possible, then (1) is true, but (2) is false. This is because ‘[e]ven the totality of psychological facts at a given point in time may be consistent with a variety of different physical states of the world’ (p. 91), and different such states could result, given the deterministic laws, in different actions being performed. Together these conclusions are taken to vindicate List’s claim that his second condition on free will is satisfied, but only at a ‘higher’ level. List is surely right that it is often compossible with the psychological facts and laws that we act otherwise than we actually do. But this doesn’t by itself tell us anything whatsoever about whether our actions were free – simply showing that there’s a sense in which an agent could have acted otherwise isn’t sufficient to show that they could have acted otherwise in the sense, if any, that matters for free will. As List himself notes (pp. 103-7), ‘could have acted otherwise’ is notoriously context-sensitive; the interesting question is not whether there’s a sense in which we could have acted otherwise but rather which of those senses, if any, we should care about. So List needs to explain why compossibility with the psychological facts and laws is the kind of possibility relevant to whether some agent acted freely. On closer inspection, moreover, it’s clear that List’s ‘agential possibility’ cannot be the kind of possibility in which freedom consists. Consider a paradigm case of unfree action – for example, the compulsive behaviour of a drug addict. Presumably List wants to say that this action wasn’t free because, in the relevant sense, the addict couldn’t have acted otherwise. (After all, her action was caused by an intention to take drugs, so List’s other conditions on free action are satisfied.) Yet it’s perfectly compatible with the psychological facts about the addict at time t that, for example, a neuroscientist intervenes by stimulating the relevant neurons in the addict’s brain to ensure that she refrains from taking drugs at time t+1 (for any unit of time). So the addict satisfies List’s conditions on free will, despite being clearly unfree. At some points, List seems to suggest that certain facts about the agent’s ‘macroscopic environment’ – such as, perhaps, the absence of any neuroscientists in the vicinity – should also be included among the facts to be held fixed when talking about what was agentially possible for them (p. 91). But it’s not clear that this really solves the problem. Just as psychological states can be realized by multiple different physical states, so too macrophysical states can be realized by multiple different microphysical states, each of which may result in very different things happening. It is compossible with the macrophysical state of a cracked egg (and the fundamental laws of nature), for example, that all the pieces spontaneously come back together again. (To see this, just take an actual, recently cracked egg, and reverse all of the momenta of all the particles in the system; the result is a microphysical state which realizes the same macrophysical state the egg is actually in, but which, when evolved forwards in time under the actual laws, develops into a realizer of the macrophysical state the egg was in before it was cracked.) So even if we hold fixed various macrophysical facts about an agent’s environment, List’s condition is likely to be trivially satisfied in most cases. Even setting aside this problem, though, there are many cases where we don’t want certain facts about the agent’s environment to be held fixed, such as the case of Lizzy voting Democrat in §2. If we include among the facts we are holding fixed that the neuroscientist will intervene if Lizzy shows any signs of wanting to vote Democrat, then it was not agentially possible for Lizzy to do otherwise. Yet only a sceptic would deny that Lizzy’s action was free. Indeed, this very same problem arises when thinking about which psychological facts to hold fixed. As Cohen and Handfield (2006) point out, there could be Frankfurt-style cases in which the ‘back-up’ mechanism is a psychological property of the agent. Suppose that Ann – willingly, and motivated entirely by her own reasons – decides to steal Cecilia’s bag; but deciding otherwise would have triggered an irresistible compulsive kleptomaniac desire causing her to steal Cecilia’s bag anyway. Intuitively, Ann’s action was free in this case. Yet List’s definition would have us hold fixed the facts about Ann’s latent kleptomania when determining which actions were possible for her to perform. List does acknowledge the challenge posed by Frankfurt cases, in a footnote (pp. 168-9). There he appears, peculiarly, simply to grant that it is agentially impossible for Lizzy not to vote Democrat, or for Anne not to steal Cecilia’s bag, from which it follows, on List’s view, that these actions aren’t free. He insists that it was up to Lizzy whether to make the decision to vote Democrat herself or as a result of the neuroscientist’s intervention, so that there is still ‘a kind of fork in the road’ for her. But even setting aside the question of whether such alternatives are sufficiently ‘robust’ to ground free will (Fischer 1994), they are not possibilities in which Lizzy acts otherwise, so it’s not clear how this strategy could vindicate List’s claim to have found a sense in which agents act freely only if they could have acted otherwise. One final worry for List’s definition of agential possibility stems from certain externalist commitments in the philosophy of mind. Suppose that knowledge is a mental state (Williamson 2000, §§1–3). If it is, facts about what one knows should be held fixed in assessing what is agentially possible. But since knowledge is factive, knowing that one will φ is inconsistent with not φ-ing. So List’s view apparently implies that agents who know in advance what they will do cannot act freely. But this is completely implausible. (The usual refrain is that the truth will set you free, not render you unfree.) Perhaps knowing what one will do is incompatible with properly deliberating over what to do; but an agent who, having already decided what to do, thereby knows that she will do it, seems perfectly able to act freely. Of course, one could reply by insisting that there are no factive psychological states, or that the relevant psychological states to be held fixed when determining what is agentially possible do not include the factive ones. But neither of these options would sit well with List’s other naturalistic commitments. First, as Nagel (2013) notes, it’s much less controversial in psychology that knowledge is a mental state than it is in philosophy. And second, there’s an obvious tension between the externalist thought that we should hold fixed features of the agent’s environment when evaluating what she could have done, and the internalist thought that we should hold fixed only those psychological states which don’t depend constitutively on the agent’s environment. In personal communication, List suggests that the relevant kind of possibility is ultimately to be read off the psychological sciences. It’s not clear to us how exactly this reading-off process is supposed to work, however. It’s certainly true, as List (pp. 97-103) points out, that the social sciences often represent agents as facing a choice between multiple courses of action. But even granting List his premise that we should be realists about any entities that scientists cite in their explanations of behaviour, it’s again not clear why the right way to be a realist about special scientists’ alternative possibilities need imply anything about what we could have done in the sense of relevance to free will. Merely having multiple options to choose from doesn’t imply that the choice one eventually makes is free, since there may be possible worlds at which one acts otherwise without one’s in fact being free to choose which possibility eventuates. Deferring to the sciences, in other words, can only get you so far: how their results ought to be interpreted is often itself an issue for philosophical investigation. 5 Mental Causation In Chapter 5, List argues that ‘a person’s actions are really caused by his or her intentional mental states rather than by some sub-intentional physical states of the brain and body’ (p. 118). Together these conclusions are taken to vindicate List’s claim that his third condition on free will is satisfied, but only at some ‘higher’ level. List’s argument relies on a broadly counterfactual account of causation according to which, to a first approximation, C caused E if and only if (i) if C had occurred, E would have occurred, and (ii) if C hadn’t occurred, E wouldn’t have occurred. List takes these counterfactuals to be true just in case (i*) E occurs in all the closest worlds in which C occurs, and (ii*) E doesn’t occur in all the closest worlds in which C doesn’t occur. Now suppose S forms an intention to φ, and φ-s. In many such cases, List argues, it’s true both that S’s φ-ing occurs in all the closest worlds where her intention occurs, and that S’s φ-ing doesn’t occur in all the closest worlds where her intention doesn’t occur. By contrast, it’s not the case that, had S not been in the exact microstate she was actually in, she wouldn’t have φ-ed; this, according to List, is because the closest worlds where S is not in the exact microstate she’s actually in are ones in which she’s in a slightly different microstate, one which still realizes the state of intending to φ. Therefore, List concludes, S’s φ-ing was caused by her intention to φ, and not by the exact microstate of her brain. One can raise a similar objection here to List’s focus on microphysical states, rather than other more coarse-grained, but still sub-personal, states (see §3, above). But even setting aside that worry, there is a further problem with List’s argument here. He seems to be assuming that if an event E is multiply realizable, then E would still have occurred had the event which actually realizes E not occurred. But this assumption simply isn’t plausible in general. Imagine for example that S’s brain is being controlled by a malevolent neuroscientist. The neuroscientist has assigned to every mental state a corresponding exact microphysical realizer, and has built a device that can manipulate S’s brain into each of those microphysical states (but can’t do anything else). The neuroscientist was weighing up whether to manipulate S into intending to φ or intending to ψ, and eventually decided on the former. Suppose also that the device is extremely reliable (it couldn’t easily have gone wrong). Now ask yourself: had S been in a different microphysical state, would it have been because the device went wrong and put S’s brain into a slightly different state to the one it was designed to put it into, or would it have been because the neuroscientist decided to make S intend to ψ instead (and the device worked as designed)? Arguably the latter. But then it follows, on List’s view, that both S’s intention to φ and S’s exact microphysical state caused S's action, since in both cases conditions (i) and (ii) are satisfied. Of course, List might reply that such cases are rare, and his conclusion still holds in most everyday cases. But this would miss the point – insofar as intentional states seem like more ‘fitting’ causes of actions than brain states in normal cases, the same seems true in the case above, notwithstanding its odd counterfactual structure. This was clearly appreciated over twenty-five years ago now by Stephen Yablo (1992). He noted that it would seem right to say that Socrates' death was caused by his drinking the hemlock, not his guzzling it, even if it turned out that ‘Socrates, always a sloppy eater, had difficulty drinking without guzzling, to such a degree that if the guzzling hadn’t occurred, the drinking wouldn’t have either’ (Yablo 1992, p. 276). For this reason, Yablo added two further conditions, on top of (i) and (ii), to his definition of causation: (iii) If a different determinate of C had occurred, E would still have occurred; (iv) If a determinable of C had occurred without C occurring, E wouldn’t have occurred. This definition arguably does imply that the microphysical state of S’s brain doesn’t cause her action in the case above (assuming, what is controversial, that intentions are determinables of the brain states which realize them), because condition (iv) isn’t satisfied – if the brain state hadn’t obtained but the intention still had, S would still have φ-ed. But the important point is that (i) and (ii) alone do not deliver the results List advertises. 6 Conclusion As we see it, Why Free Will is Real makes three main claims: (1) an agent acts freely if and only if her intentions appropriately caused her action and it was agentially possible that she not perform that action; (2) on the assumption that an action is free if and only if it satisfies these two conditions, many human actions are in fact free; and (3) free will and its grounds are ‘higher-level’ phenomena. In the preceding sections, we have raised reasons to doubt List’s arguments for all three of these claims. But we also have a deeper, more methodological concern about the project List is attempting here. At different points, List appeals to what he calls the ‘naturalistic ontological attitude’, according to which ‘[o]ur best guide to any questions about which entities, properties, or phenomena exist in any given domain is to be found in our best scientific theories of that domain’ (p. 74). This methodological principle is billed as the hero of the day, rescuing the phenomena of intentionality, mental causation, and ultimately free will, from a less ecumenical naturalism that threatens to reveal them as illusions. In highlighting these theoretical successes, List presumably hopes to stage an abductive case for the naturalistic attitude itself. Indeed, we suspect that the best way of understanding Why Free Will is Real is as a kind of manifesto: an advertisement for List’s particular brand of non-reductive naturalism. In the end, though, we worry that List’s approach ends up granting too much to the scientistic sceptics with whose doubts he opens this book. Compatibilist approaches to free will have always been motivated by the conviction that the question of our status as free and responsible agents is not the sort of thing to which fundamental physics alone could provide an answer. Perhaps the most intriguing aspect of List’s book, then, is the suggestion that while free will might be compatible with physical determinism, it is nevertheless the sciences – specifically, the higher-level sciences – to which we ought to look to determine whether free will exists. There’s a certain dialectical appropriateness to this strategy, given List’s target audience (and if it succeeds in waking a few philosophically-minded scientists from their dogmatic slumbers, then so much the better). But while we don’t doubt the relevance of the higher-level sciences to the question of who has free will and to what extent, we are less optimistic than List seems to be that its existence can be established merely by taking the higher-level sciences seriously. There is a difference between maintaining that philosophy does and should operate on a methodology similar to the disciplines classically considered sciences, and maintaining that, if there are answers to philosophical questions, they will be discoverable by those disciplines. We find the first of these methodological principles attractive; the latter much less so. For all that List has shown, whether we have free will is – and remains – a distinctively philosophical question.* Footnotes " * The authors are grateful to Christian List for helpful comments on an earlier draft of this review. References Anscombe G.E.M. 1957 , Intention ( Cambridge, MA : Harvard University Press ). Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Brink David 2013 , ‘Situationism, Responsibility, and Fair Opportunity’ , Social Philosophy and Policy , 30 ( 1-2 ): 121 – 149 . Google Scholar Crossref Search ADS WorldCat Cohen Daniel , Handfield Toby 2007 , ‘Finking Frankfurt’ , Philosophical Studies , 135 ( 3 ): 363 – 374 . Google Scholar Crossref Search ADS WorldCat Fischer John Martin 1994 , The Metaphysics of Free Will: An Essay on Control ( Cambridge : Blackwell ). Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Fischer John Martin , Ravizza Mark 1998 , Responsibility and Control: A Theory of Moral Responsibility ( Cambridge : Cambridge University Press ). Google Scholar Crossref Search ADS Google Scholar Google Preview WorldCat COPAC Frankfurt Harry 1969 , ‘Alternative Possibilities and Moral Responsibility’ , Journal of Philosophy , 66 ( 23 ): 829 – 839 . Google Scholar Crossref Search ADS WorldCat Frankfurt Harry 1971 , ‘Freedom of the Will and the Concept of a Person’ , Journal of Philosophy , 68 ( 1 ): 5 – 20 . Google Scholar Crossref Search ADS WorldCat Kment Boris 2017 , ‘Free Will and Ultimate Explanation’ , Philosophical Issues , 27 ( 1 ): 114 – 130 . Google Scholar Crossref Search ADS WorldCat Mele Alfred 1995 , Autonomous Agents: From Self Control to Autonomy ( Oxford : Oxford University Press ). Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Nagel Jennifer 2013 , ‘Knowledge as a Mental state’ , Oxford Studies in Epistemology , 4 : 275 – 310 . OpenURL Placeholder Text WorldCat Nelkin Dana 2005 , ‘Freedom, Responsibility and the Challenge of Situationism’ , Midwest Studies in Philosophy , 29 ( 1 ): 181 – 206 . Google Scholar Crossref Search ADS WorldCat Sartorio Carolina 2016 , Causation and Free Will ( Oxford : Oxford University Press ). Google Scholar Crossref Search ADS Google Scholar Google Preview WorldCat COPAC Sartorio Carolina 2018 , ‘Situations and Responsiveness to Reasons’ , Noûs , 52 ( 4 ): 796 – 807 . Google Scholar Crossref Search ADS WorldCat Watson Gary 1975 , ‘Free Agency’ , Journal of Philosophy , 72 : 205 – 20 . Google Scholar Crossref Search ADS WorldCat Williamson Timothy 2000 , Knowledge and its Limits ( Oxford : Oxford University Press ). Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Williamson Timothy 2006 , ‘Can Cognition be Factorized into Internal and External Components?’, in Stainton R.J., (ed.), Contemporary Debates in Cognitive Science ( Malden, MA : Blackwell Publishing ). Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Yablo Stephen 1992 , ‘Mental Causation’ , The Philosophical Review , 101 ( 2 ): 245 – 280 . Google Scholar Crossref Search ADS WorldCat Yablo Stephen 1997 , ‘Wide Causation’ , Philosophical Perspectives , 31 ( 11 ): 251 – 281 . OpenURL Placeholder Text WorldCat © Mind Association 2020 This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/open_access/funder_policies/chorus/standard_publication_model)

journal article

LitStream Collection

What the Future ‘Might’ Brings

Boylan,, David

2009 Mind

doi: 10.1093/mind/fzz037pmid: N/A

Abstract This paper concerns a puzzle about the interaction of epistemic modals and future tense. In cases of predictable forgetfulness, speakers cannot describe their future states of mind with epistemic modals under future tense, but promising theories of epistemic modals do not predict this. In §1, I outline the puzzle. In §2, I argue that it undermines a very general approach to epistemic modals that draws a tight connection between epistemic modality and evidence. In §3, I defend the assumption that tense can indeed scope over epistemic modals. In §4, I outline a new way of determining the domain of quantification of epistemic modals: epistemic modals quantify over the worlds compatible with the information accumulated within a certain interval. Information loss can change which interval is relevant for determining the domain. In §5, I defend the view from some objections. In §6, I explore the connections between my view of epistemic modality and circumstantial modality. It is easy to talk about what we know right now, what we knew yesterday, and what we will come to know tomorrow. We can also talk about what might be now and what might have been the case yesterday. But what about what might be tomorrow? I argue that we can indeed talk coherently about future epistemic possibilities, but not as we might have expected. In cases where information is lost, epistemic modals cannot simply be used to talk about our future evidence. In light of this, I argue that we need a new theory of the domains of quantification for epistemic modals. Epistemic modals are often thought of as tools for describing a given body of evidence. Specifically, a sentence like (1) It might be raining. means roughly (2) The current evidence doesn't rule out that it is raining. Most theories of epistemic modals formalize this thought as follows:1 Evidence: ⦑might φ if and only if φ is compatible with the evidence of the relevant agent(s) in c in w at t. Theories differ over whose evidence is relevant. But they agree on the central point: what is epistemically possible at a given time is whatever is compatible with the evidence of a certain set of agents at that time. I test the predictions of Evidence against cases of predictable forgetting. In such cases, speakers cannot describe their future states of mind with epistemic modals under future tense, even though they know that they will be able to do so. Moreover, no similar barrier exists for past epistemic modal claims. Evidence can't predict these data. I propose replacing Evidence with: History: ⦑might φ⦒ if and only if φ is compatible with the evidence in the relevant interval up to and including t. According to History, epistemic modals collect information that agents gather over time, not just their evidence at a particular time. This makes epistemic modals sensitive to evidence agents used to have, as well as evidence they have now. Epistemic modals never lose information: once something becomes necessary, it remains necessary. They are tools to describe information obtained over time in abstraction from the deficiencies of the agents who gather it. 1. Predictable forgetting An agent predictably forgets some proposition p just in case she knows p at some time t and knows that she will forget p by some later time ⁠. Predictable forgetting is not uncommon. I can predict that I will forget that the person I just met is called John, that John lives in Apartment 211, that his phone number is such-and-such. The structure of my cases will be somewhat more involved. But we do well to remember that the basic phenomenon is pervasive. Here's an example. Keys on Monday. On Monday night, Alice leaves her keys in the kitchen. But she always forgets where she has left the keys overnight. She foresees that tomorrow morning she will look for them in her bedroom. What kinds of epistemic modal claims can Alice make? Three observations. First, Alice cannot report her future ignorance by saying: (3) # Tomorrow my bedroom will be one of the places that the keys might be in. (4) # Tomorrow my keys might be in the bedroom. (5) # Tomorrow it will be possible that the keys are on the nightstand. (6) # Tomorrow I'll look in my bedroom because the keys might be on the nightstand. (7) # Tomorrow I'll look in my bedroom because it will be possible that the keys are on the nightstand. (8) # Tomorrow I will search every drawer in my room which might contain the keys. These sentences should have true readings where the tense scopes over the modal. Nevertheless, they sound marked. Observation 1 sums this up:2 Observation 1. Whenever a speaker knows that φ, they cannot assert ⌜⌝, even if they know they will no longer know φ at some later time. What can Alice's future, less informed self say? Consider Keys on Tuesday. Alice wakes up on Tuesday and has forgotten where the keys are. She looks for them upstairs. Alice says (9) The keys might be on the nightstand. What Alice says is perfectly assertable; she's not making some mistake in uttering this sentence. In Alice's situation, you might say something like (9) too. This pattern is general and gives us our second observation: Observation 2. At the later time when the speaker has forgotten φ, she may assert ‘’. Finally, consider what Alice can say about her past self. As von Fintel and Gillies (2008) note, agents can use past ‘might’-claims to talk about times when they had less knowledge. Continue the case: Keys on Monday Again. On Monday night, Alice's partner asks her what she was doing rummaging around in their bedroom on Sunday morning. Alice was in fact looking for the keys, which her partner had placed in the living room. Imagine Alice replies (10) Yesterday I was searching our bedroom because the keys might have been in there. (It turned out they were in the living room.) (11) Yesterday I was searching every drawer that the keys might have been in. Strikingly, this is perfectly natural, even though Alice knows that the keys were not in the bedroom. Thus: Observation 3. If they know they previously did not know φ, even a speaker who currently knows that φ can assert ⌜⌝. There are two surprises here. First is the combination of Observations 1 and 2. In these cases, why is there a difference between what I will be able to say about myself using an epistemic modal and what I can now say about my future self? This is not how things usually go. If I foresee that tomorrow I will correctly say ‘It is raining in Topeka’, then it should be correct for me now to say ‘It will rain in Topeka’. Why are epistemic modals different? The second surprise is the combination of Observations 1 and 3. They show that there is an asymmetry between past- and future-tensed epistemic modals. Present information that φ interferes with our assessment of ⌜⌝ but not ⌜⌝. Why does the direction in time matter? 2. What Evidence predicts I now make precise why predictable forgetting is a challenge for Evidence. To make Evidence precise, I use a broadly Kratzerian account, which posits a modal base, ⁠, a contextually supplied function from a world and a time to a set of propositions. The modal quantifies over the intersection of ⁠.3 (12) if and only if ⁠. I treat tenses as operators, shifting the time of evaluation: (13) if and only if ⁠. (14) if and only if ⁠. where just in case t is earlier than ⁠.4 Combining these assumptions yields this schematic entry for future ‘might’-claims: (15) if and only if ⁠. ⦑⦒ is true at time t if and only if there is a φ-world in the modal base as evaluated at some point later than t. Similarly, for past ‘might’-claims: (16) if and only if ⁠. ⦑⦒ is true at a time t just in case there is a φ-world in the modal base as evaluated at some point earlier than t. Start with solipsistic contextualism, according to which the modal talks about the speaker's evidence. More specifically, when given a world w and a time t, outputs the set of propositions that make up the speaker's evidence in w at t. In Keys, Alice's only evidence about the keys is what she remembers about their location, so Alice's evidence and her knowledge are interchangeable here. So is the set of propositions that Alice knows in w at t. This semantics fails to predict Observation 1. We said Alice will forget the location of the keys. So when (3)–(8) are read as future ‘might’-claims, they are predicted true. And Alice knows this. She should be able to tell that (3)–(8) are true. Evidence predicts that (3)–(8) should be sensible things for Alice to say. I have focused on solipsistic contextualism, a theory thoroughly undermined by MacFarlane's (2011) problem of disagreement. Nevertheless, proposals for treating disagreement do not help. Take Dowell's (2011) proposal: the speaker can select whatever modal base she likes, so long as she can reasonably expect her audience to figure out what she means. In predictable forgetting cases, it would be reasonable to take the relevant knowledge to be the speaker's. This interpretation makes (3)–(8) true and relevant. This view won't rule out the problematic readings of (3)–(8); it just adds possible readings. This problem generalizes to other sophisticated contextualisms.5 What about relativism? Relativists add a parameter to the index—a context of assessment parameter. It supplies the relevant knowledge. Where is the set of propositions known by the assessor in w at t, (17) if and only if ⁠.6 That is, ⌜⌝ is true at just in case the the assessor's evidence in a at t is consistent with φ. What does relativism predict about ⌜⌝ in Keys? That depends on what the assessors will know the next day. Plausibly, we expect our current knowledge to persist through Tuesday, so relativism predicts that the ⌜⌝ examples are false in our context of assessment: since we won't forget ⌜might φ⌝ will be false when fed times from Tuesday. But then ⌜⌝ is false as said on Tuesday too: the same information is used to assess it. Relativism wrongly predicts ⌜⌝ as said on Tuesday should sound just as bad to us as ⌜⌝ as said on Monday.7 The problem is clear. Evidence can't predict both Observations 1 and 2. It seems to make the ‘might’ claim made in the future true only if it makes the future tensed ‘might’ claim made in the present true too. Likewise, Evidence can't predict both Observations 1 and 3. According to Evidence, modals simply track the evolution of some body of information: present information is irrelevant in the past and the future. 3. Tensing modals We've assumed that tense can scope over epistemic modals. Some, like Groenendijk and Stokhof (1975), Iatridou (1990) and Hacquard (2006), reject this assumption. I argue that they are mistaken. There are examples of past- and future-shifted epistemic modals. 3.1 Evidence of shifting To be sure, epistemic modals at least seem to have tensed readings. As noted in Observation 3, epistemic modals have backwards-shifted readings: past ‘mights’ there are felicitous. We also find forward-shifted readings. Consider this case:8 Shortlist. You and I are trying to figure out who got the job at our university. We don't know very much about who the candidates might be, but we know we will see the shortlist tomorrow. We will contact everyone on the shortlist to see if they got the job. You say to me: (18) Let's drop this matter for now. Tomorrow we will contact everyone who might have been hired. (19) Tomorrow we will contact every possible hire. These sentences aren't about who might have been hired given our current evidence, but about who might have been hired given our future evidence. Suppose today we think John might have been hired, but tomorrow we discover he wasn't. If we didn't contact John, it would be bizarre for me to claim that you are not following up on what you said in (18) or (19). This is because you were talking about your future evidence, not your evidence at the time of utterance. This is an important point. We can't explain the badness of future ‘might’-sentences (3)–(8) by denying that forward-shifting is possible. This makes our situation much more puzzling. If future-shifted readings exist, why are they not available in Keys? 3.2 Explaining away shifting Some claim that past-shifting is merely apparent. Can we say the same about future-shifting? One thought is that the examples from the last section involve an elided attitude ascription. When we hear sentences like (20) The keys might have been in the drawer. (21) We will interview everyone who might have gotten the job. we are really evaluating sentences like (22) I thought that the keys might have been in the drawer. (23) We will interview everyone who we will think might have gotten the job. Then we could say, as Abusch (1997) and Hacquard (2006) do, that it is the attitude verb that shifts the time of the modal, not tense. So perhaps we can explain apparent tense-shifting, while still denying the key assumption in my statement of the puzzle.9 But this account makes exactly the wrong predictions in predictable forgetting: allowing tacit attitude verbs overgenerates felicitous future ‘mights’. In Keys on Monday, Alice cannot say (3) # Tomorrow (I'll look in my bedroom because) the keys might be on the nightstand. But she can say (24) Tomorrow I will think the keys might be on the nightstand. If we countenance elided attitudes, this contrast is puzzling. Why is there no elided attitude in (3)? After all, if it were interpreted this way, it would not be marked, but perfectly intelligible. And there is no obvious reason why elision should happen in von Fintel and Gillies's cases as well as Shortlist but not in cases like Keys. In both kinds of cases, speakers are heard as trying to explain past or future actions in terms of what they know at the time; but in only one of them are they successful. The elision strategy gives no insight as to why this should be. The elision idea doesn't dissolve the puzzle; indeed, the puzzle is evidence against the elision approach in the first place.10 Before moving to my proposal, I want to address the reasons that are taken to motivate scepticism about tense shifting. In Hacquard (2006) and Hacquard (2011) it is noted that certain sentences seem to lack the expected tensed readings. Take: (25) Mary must have taken the train. (26) Tomorrow Marikos might be dead. The only epistemic reading of (25), when uttered out of the blue, is one which talks about our present knowledge; and it is hard to get any epistemic reading of (26). Likewise, it is hard to get epistemic readings when modals are embedded under temporal adverbials: (27) # Usually Mary might take the train. Hacquard and others think that the absence of such readings is readily explained if we say tense is barred from ever shifting the relevant knowledge.11 But what about a pragmatic explanation? If the syntax and semantics of modals are supposed to explain the data, we should not expect to see similar data for knowledge attributions. But note that even tensed knowledge attributions can sound odd out of the blue. Consider, for instance: (28) Q. How did Mary leave town? A. ?? I knew she took the train. (29) ?? Tomorrow I won't know whether Marikos is dead. Without context, both sentences sound odd. We notice the same pattern with temporal adverbials: (30) ?? Usually I don't know whether Mary will take the train. There is a natural explanation of these data: without context, it's hard to see why anything other than our current attitudes would be relevant. When we are wondering how Mary left town or whether Marikos is dead, why care about what I knew in the past or will know in the future? This explanation naturally extends to modals. We search for readings of (25) and (26) that aren't just true, but also relevant. Since the missing readings are irrelevant, we do not select them. And when context makes clear why other knowledge states would be relevant, those shifted readings should appear. This is exactly what we observe: our prima facie cases of shifting are precisely cases that make clear why other knowledge states would be relevant. Our original observations do not derive from some general problem about combining epistemic modals and tense. Tense can scope above epistemic modals. It just doesn't always produce the results we would expect. 4. An interval-based solution I now give a semantic solution to our puzzle. I give a new kind of modal base to predict our three observations. 4.1 The modal base I introduce two innovations to how modal bases work. First, I replace Evidence with: History. ⌜⌝ is true at a world w and time t if and only if φ is compatible with all the evidence acquired in w in the relevant interval up to and including t. Instead of simply collecting together the propositions that make up the agent's evidence at a given time, as Evidence does, epistemic modal bases will also collect those propositions at times within a certain interval before the input time. Thus, while epistemic modals still have something to do with the evidence of the relevant agents, they can also come apart from that evidence as we go forward in time. (Again, in our cases the only relevant evidence is what the agent knows, so I will treat evidence and knowledge as interchangeable.) Second, I give a theory about which intervals are relevant. I say the relevant interval starts with the most recent episode of information loss. Given a time t, define ⁠, which I'll call a partial history, as follows: (31) either or and is not earlier than the most recent episode of information loss relative to t}. Intuitively, to construct a partial history for a given t, take an agent's entire history and remove the part before the most recent (relative to t) episode of information loss.12 The modal base is defined relative to a partial history. Where is the partial history corresponding to the time of the context c: Partial History. = ⁠: the relevant agent knows p at some where and ⁠. When fed a time t (and world w), an epistemic modal base delivers the set of propositions known at times up to t (in w) in the partial history ⁠. In other words, modal bases collect the facts known within a given partial history up until t. To find the domain of quantification ⁠, we find all the propositions known at any time in before t and intersect that set. Note two important features of this account. First, the modal base is monotonic: when is later than t, if p is in the modal base at t, then it must also be in the modal base at ⁠. So the set of epistemic possibilities cannot grow as time passes. Second, moving between partial histories introduces a context-shift. Whenever information is lost, the relevant partial history changes; and since modal bases are anchored to partial histories, the modal base changes, too. And this seems right. Were the context not to shift, information would continue to accumulate in the modal base across contexts of utterance, even though the relevant agents do not have access to it. This is undesirable: the more time passes, the fewer epistemic modal claims an agent would be justified in making. (I would not, for instance, be justified in telling you that on this day a year ago, I might have eaten cornflakes for breakfast, for I know my past knowledge settles the matter one way or the other!) If the modal base collects what is known from some point closer to the context of utterance, context-shifting solves the problem. 4.2 Predictions Now recall our case: Keys. On Monday night, Alice leaves her keys in the kitchen. But she always forgets where she has left the keys overnight. In fact, she can foresee that tomorrow morning she will go looking for them in her bedroom. Alice says (3) # Tomorrow I'll look in my bedroom because the keys might be on the nightstand. Keeping the semantics from earlier, we still have: (15) if and only if ⁠. Let be the modal base supplied by Monday's context of utterance. To evaluate (3), we need to see whether contains worlds where the keys are on the nightstand. Partial History tells us that contains all propositions known by Alice at any time in Monday's partial history up until Tuesday. Since she knows on Monday that the keys are in the kitchen, contains the proposition that the keys are in the kitchen. In Monday's context, ⁠, the domain for the modal evaluated at Tuesday will not contain worlds where the keys are on the nightstand. So (3) is false. Moreover, Alice's semantic knowledge, together with her knowledge of the case, guarantees that she knows (3) is false. Hence (3) is marked. This explains Observation 1. We also rightly predict that Alice's knowledge only gets carried forward. Consider again: Keys on Monday Again. On Monday night, Alice's partner asks her what she was doing rummaging around in their bedroom the previous morning. Alice was in fact looking for the keys, which her partner had placed in the living room. Alice says (10) Yesterday I looked in our bedroom because the keys might have been there. Again, we give past ‘might’-claims the following semantics: (16) if and only if ⁠. To see whether (10) is true, we need to see whether for some time t on Sunday will contain worlds where the keys are in the bedroom. Partial History predicts that it will. will contain just the propositions known by Alice up until Sunday (within Monday's partial history). Until the point at which she found them, Alice did not know where the keys were. So will contain worlds where the keys were in the bedroom. Hence (10) is true. This explains Observation 3. Partial History predicts the asymmetry brought out by Observations 1 and 3 precisely because it is monotonic: past information is carried forwards, but future information is not carried backwards. (This feature will be shared by any History-based view.) Finally, we want to explain the contrast between what Alice can say on Monday and what she can say on Tuesday. Consider the continuation of the case: Keys on Tuesday. Alice wakes up on Tuesday morning and has indeed forgotten where the keys are. She goes looking for them upstairs. We noted that Alice can now say (9) The keys might be on the nightstand. I predict this by appeal to partial histories. Since something is forgotten overnight, different partial histories correspond to Monday and Tuesday, so will be different from ⁠. Moreover, since on Tuesday Alice does not know where the keys are, will not contain the proposition that the keys are in the kitchen. So will contain worlds where the keys are on the nightstand. Thus (9) as said by Alice on Tuesday is true and assertable. This explains Observation 2. On my proposal, modal bases are monotonic within a context, but not across contexts. This is why I predict the contrast between what Alice can say on Monday and what she can say on Tuesday, but also the asymmetry between past and future ‘might’-claims.13 5. Problems of context Now for two objections, one about the judgements in Keys on Monday and one about monotonicity. My answers to both exploit the connection between time and the context-sensitivity of ‘might’. 5.1 Felicitous future ‘mights’? There is a wrinkle in the data around Observation 1. Some report that we can improve future ‘mights’ by tying them more explicitly to an explanation of our future behaviour. For example, as a referee points out, sentences like (6) and (7) may sound better than those without the ‘because’.14 Now, this is not a reason to think future ‘mights’ are fine in cases of predictable forgetting. Even if (6) and (7) are better, sentences like (3)–(5) and (8) are still marked in Keys. But there certainly is something to be explained. I argue that these judgements track a second, less prominent reading of a future ‘might’, a new kind of exocentric reading. Exocentric readings are ones where ‘might’ does not have its default interpretation, but rather latches onto the information of some third party. Take a case from Egan, Hawthorne and Weatherson (2005): The Bus. My friend Sally is avoiding Tara. Sally is on a bus and, worried that Tara might board it, goes to hide at the back. I'm watching all this from some nearby bushes, but can see that Tara is not in fact on the bus. On seeing Sally's behaviour, I explain to a companion: (32) Sally is hiding because Tara might be on that bus. (32) is true, even though I clearly know that Tara isn't on the bus. Rather I am using ‘might’ to express what it normally does in Sally's mouth. We get evidence for thinking the good readings of sentences like (6) and (7) are exocentrics by noting that they appear and disappear in some of the same contexts as exocentrics. Exocentrics tend to require explanatory contexts. In The Bus, I cannot simply turn my attention to Sally and say (33) # Tara might be on that bus.15 The ‘because’ is necessary to elicit the exocentric reading. This is one point of similarity: the improved (6) and (7) explicitly involve some explanation; the marked (3)–(5) and (8) do not. But even in explanatory contexts, certain continuations block exocentrics. Suppose instead of just (32) I say: (34) Sally is hiding because she thinks Tara might be on that bus; # and indeed Tara might be. The continuation, ‘and indeed Tara might be’, is marked here. But then the second ‘might’ cannot be exocentric. I would be saying that Sally thinks that Tara's being on the bus is compatible with her evidence, and that her belief is correct. But this is true and relevant. Presumably instead the ‘might’ here has its default reading, which makes the continuation false. Now consider what happens when future ‘mights’ occur in these environments. Suppose that Alice says: (35) Tomorrow I'll look in my bedroom because I'll think the keys might be on the nightstand. #And indeed that will be one of the places they might be. As in (34), the material beginning ‘And indeed’ is marked. Again we see a parallel with exocentrics. My theory explains how (6) and (7) could be exocentric. Exocentrics arise when we interpret ‘might’ as it would normally be used in other contexts. Alice's future self is not literally a third party, but that future self does occupy a different context: Alice will occupy a different partial history, and so her ‘might’ will mean something different then. This will be a meaning for exocentrics to target; and since that reading will be true, a present exocentric ‘might’ would be true now. 5.2 Monotonicity violations in the past? My account is fully monotonic: modal bases are monotonic from any time forward. But one might wonder whether we really want monotonicity in the past. Return to Keys on Monday. Suppose that Alice has woken up and remembers that she knew the keys were either in the living room or in the kitchen. She says (36) (a) Either the keys had to be in the kitchen or the keys had to be in the living room. (b) But now they could be in either place. These utterances sound true. But my proposal says that if (36a) is true, then (36b) should be false. For what was necessary remains necessary. Even worse, I predict the modal base for (36a) is empty: since yesterday falls outside Alice's current partial history, no proposition known then will be in the modal base. So the domain of the modal will include all possible worlds and each disjunct of (36a) is false. A revised fully monotonic account can predict these data. First, define our modal bases as follows: where is a partial history, (37) If t is in or later than ⁠, ⁠: there's a time not later than t s.t. and p is known at ⁠; is undefined otherwise. Now, when the modal base is supplied a time before the relevant partial history, it will be undefined. Next, tweak the account of how the modal base is supplied. Instead of just selecting the modal base corresponding to the smallest partial history the context is in, we select the smallest partial history that also makes the utterance defined. More formally: Defined Partial History. Where is the smallest partial history containing that makes the utterances in c defined, ⁠. We take the partial history starting with the most recent episode of information loss and if the utterance is defined against the corresponding modal base, we use that one. If not, we move on to the next episode of information loss and try again. And so on. Now return to (36a). The smallest partial history that makes Alice's utterance defined is the smallest one that includes yesterday. And yesterday she knew where the keys were. So when fed a time from yesterday, the corresponding modal base will contain either the proposition that the keys are in the kitchen or the proposition that they are in the living room. Since that is the interval we are quantifying over, (36a) is true. To predict (36b), I say (36b) is interpreted in a different context from (36a). This is because (36b) is uttered after (36a) and contexts include their times. In (36b)'s utterance context, the partial history which stops at Alice's most recent forgetting makes (36b) defined. But this does not include yesterday's forgotten information. So (36b) is true. This kind of temporal context-shift is not ad hoc, but independently required. Consider a discourse like this: (38) Alice is sitting down. Now she's standing up. Now she's sitting down again. Here ‘now’ does not pick out the same time throughout. Here too the natural thought is that, because each utterance takes place in its own temporally distinct context, each ‘now’ picks out a different time.16 The example can be explained away; but why not instead opt for a partially monotonic solution, where modal bases are only monotonic from the context-time onwards? After all, this would predict (36a) and (36b) to be true without any subtle manoeuvring. The trouble is that I do not know how to generate this kind of monotonicity in a principled way. We could write it in by hand into our modal bases, but we should ask for more than this. My account gives a principled explanation of the full monotonicity property. And so I think adopting Defined Partial History is worth the cost to simplicity. 6. A new picture I have solved my puzzle with monotonicity and context-shifting. I now conclude by considering some bigger-picture upshots for the distinction between epistemic and historical modality, noting that, on my view, the two are more similar than typically thought.17 Consider (39) It hasn't been decided yet who will get the job. Alex might get it and so might Billy. The ‘might’ here has a historical (or circumstantial) reading, where it talks about what is possible given the relevant facts. On this reading, the truth of (39) depends not on what we know, but on what holds in the world more generally. These readings are generated by historical modal bases like the following:18 (40) is a proposition about an interval that does not extend past t and p is true at ⁠. Notice that, as stated in (40), historical modality will be monotonic. This is no accident. Among other things, Thomason (1984) and Condoravdi (2002) note that this captures an important asymmetry between past and future: once something becomes settled, it remains settled; but what lies in the future can remain open. For example, suppose that it rains at time t. Before t, it was historically possible that it would not rain at t. But once we get to t and it in fact does rain, it is no longer historically possible for it to have not rained at t. What was possible in the past becomes historically impossible from t onwards. We now have two asymmetries in time, one epistemic and one historical. It is surprising that both kinds of modal base should have this property, especially if there is a sharp distinction here. Why would both core flavours of modality have an open future but a settled past? One avenue to explore is that epistemic modality is actually a circumscribed form of historical modality. φ is historically possible in a world at a time t if and only if φ is compatible with the relevant facts in that world which hold up to and including t. We might say that φ is epistemically possible at t just in case it is compatible with the facts about the evidence in a given partial history up to and including t. Thus, instead of being a distinct form of modality, epistemic modality would be a subspecies of historical modality, one that limits itself to facts about particular partial histories. Spelled out, we would have Historical: if and only if φ is compatible with the facts in w about the evidence in the relevant interval up to t. This reduces the two asymmetries to one. The monotonicity property of historical modals accounts for the difference between past and future in both the epistemic and the historical cases. The epistemic asymmetry that this paper is about would just be a special case of the more familiar asymmetry between past and future. Clearly, much more would need to be said before Historical could be accepted; epistemic modals are thought to be special in a variety of ways. But it would give a satisfying explanation of why epistemic modals have the monotonicity properties needed to solve my puzzle; and this, I think, should make it worthy of future investigation.19 Footnotes 1 Something like this view originates in Moore (2013), and can be traced through Hacking, (1967), Teller (1972), DeRose (1991), and Kratzer (1981). 2 This markedness also projects under supposition. Imagine Alice supposes the scenario above held. The following is then true: (i) Alice supposes that she knows the keys are not in her bedroom, but that she will forget this. Can she suppose that it will in the future be possible that they are somewhere else? It seems not. Consider: (ii) Alice supposes that she will search every drawer in her room which might contain the keys. Now Alice seems to be contradicting her initial supposition. 3 This is von Fintel and Gillies's (2008) formulation, which adds a time-parameter to Kratzer (1977, 1981, 1991). 4 I could equally well use a referential theory of tense, as in Partee (1973) or Abusch (1997), or a modal theory of ‘will’, as in Enç (1985), Abusch (1997), and Cariani and Santorio (2017). (See Ogihara 2007 for more on these issues.) 5 For instance, von Fintel and Gillies's (2011) cloudy contextualism faces the same issue. 6 I add a time parameter to MacFarlane (2011). 7Stephenson (2007) and Egan (2007) face the same problem. 8Dorr and Hawthorne (2013) discuss a similar case. 9Hacquard (2011) suggests apparent tense-shifting is due to a tacit, modal base-shifting ‘because’. But in (18) and (19), there is nowhere to put the ‘because’ that gets right reading. 10 Another issue is the following contrast: (i) (a) I thought the keys might have been in the drawer and I was right: they might have been. (b) I thought the keys might have been in the drawer and I was right: # I thought they might have been. The elision approach predicts no contrast here. 11 Note also that these data hold only for the auxiliaries ‘might’ and ‘must’ and not ‘it is possible’. But my data include epistemic ‘possible’ too. 12 Really this should be the most recent episode of relevant forgetting. We don't want, say, Alice's forgetting what she had for breakfast to trigger a context-shift. At a first pass, a relevant proposition is a (partial) answer to a question under discussion (in the sense of Roberts (2012)). 13 A referee asks about cases where both the learning and the forgetting are predictable, like the following: I will learn this evening whether aliens landed at Roswell; but my memory will be wiped by the men in black overnight. Consider: (i) Tomorrow morning it will be possible that aliens landed at Roswell and it will be possible that they didn't. I predict that this is marked: either the proposition that they landed at Roswell or its negation will be in my future modal base. I think this is the right prediction, but I grant that the intuitions are less clear here. To bolster my intuition, note that the judgements are clearer if the discourse highlights the information loss. (ii) Either I'll forget that they landed at Roswell or I'll forget that they didn't land at Roswell. # On Tuesday morning, it'll be true that the aliens might have landed in Roswell and that they they might not have. Similarly, we are reluctant to describe ourselves as knowing either of these future ‘mights’: (iii) # Tomorrow morning, I'll know that the aliens might have landed in Roswell and I'll know that they might not have. This is puzzling, if they are indeed true. Empirical investigation could be illuminating here; but hopefully these data make my predictions seem plausible. The referee floats an alternative proposal, whereby the modal base just unions the information at the future time and the information at the context of utterance. This would correctly predict that future ‘mights’ are marked in cases like Keys, as well as having the (possible) benefit of predicting that (13) is true. The data above aside, I think it will be difficult to spell out this proposal in a way that both predicts and explains the past–future asymmetry. The simplest general formulation of the view is that unions together the information at w and t and the information at the context c. But then information in the context should also get carried into past ‘mights’, and so we do not predict Observation 3. A different view might say that if t is earlier than the time of the context, then contains just the information at w and t; otherwise, it also includes the information in c. This does predict Observation 3, but basically by stipulation. History derives this from more general properties of the modal base. 14 In my experience, this opinion is not unanimous, hence my original hash-mark. But for most, an improved reading is possible if enough contextual pressure is applied. 15PaceEgan, Hawthorne and Weatherson's (2005) own presentation. 16 This assumes neither a Kaplanian semantics (Kaplan (1989)) nor an anaphoric semantics for ‘now’ (Hunter 2012; Altshuler 2016). 17 See, for instance, Kratzer (1991) and Williamson (2016). This also seems implicit in recent expressivist theories, such as Yalcin (2011) and Moss (2018). 18 This is Khoo's (2015) formulation. 19 Thanks to audiences at St Andrews, Oxford, UCLA, MIT, the 2017 Pacific APA, the 2017 Joint Session, and the Vancouver Summer Philosophy Conference, and to Martín Abreu Zavaleta, Alex Byrne, Cleo Condoravdi, Kai von Fintel, Cosmo Grant, Lyndal Grant, Abby Jaques, Matthias Jenny, Matt Mandelkern, Kieran Setiya, Jack Spencer, Shane Steinert-Threlkeld, Eric Swanson, Will Starr, Timothy Williamson, Jonathan Wright, Yuna Won and Alper Yavuz. Special thanks to Justin Khoo, Milo Phillips-Brown, Bob Stalnaker, Ginger Schultheis and Steve Yablo for very many comments on too many drafts. References Abusch Dorit 1997 : ‘Sequence of Tense and Temporal De Re’ Linguistics and Philosophy , 20 ( 1 ), pp. 1 – 50 . Google Scholar Crossref Search ADS WorldCat Altshuler Daniel 2016 : Events, States and Times: An Essay on Narrative Discourse in English . Berlin : De Gruyter Open . Google Preview WorldCat COPAC Cariani Fabrizio , Santorio Paolo 2017 : ‘Will Done Better: Selection Semantics, Future Credence, and Indeterminacy’ . Mind , 127 , pp. 129 – 65 . Google Scholar Crossref Search ADS WorldCat Condoravdi Cleo 2002 : ‘Temporal Interpretations of Modals: Modals for the Present and for the Past’. In David I. Beaver, Luis D. Casillas Martínez, Brady Z. Clark, and Stefan Kaufmann (eds.), The Construction of Meaning , pp. 59 – 87 . Stanford, CA : CSLI Publications . Google Preview WorldCat COPAC DeRose Keith 1991 : ‘Epistemic Possibilities’ . Philosophical Review , 100 ( 4 ), pp. 581 – 605 . Google Scholar Crossref Search ADS WorldCat Dorr Cian , Hawthorne John 2013 : ‘Embedding Epistemic Modals’ . Mind , 122 , pp. 867 – 913 . Google Scholar Crossref Search ADS WorldCat Dowell J. L. 2011 : ‘A Flexible Contextualist Account of Epistemic Modals’ . Philosophers' Imprint , 11 ( 14 ), pp. 1 – 25 . WorldCat Egan Andy 2007 : ‘Epistemic Modals, Relativism and Assertion’ . Philosophical Studies , 133 ( 1 ), pp. 1 – 22 . Google Scholar Crossref Search ADS WorldCat Egan Andy John Hawthorne, and Brian Weatherson 2005 : ‘Epistemic Modals in Context’. In Gerhard Preyer and Georg Peter (eds.), Contextualism in Philosophy: Knowledge, Meaning, and Truth , pp. 131 – 70 . Oxford : Clarendon Press . Google Preview WorldCat COPAC Enç, Mürvet 1985: ‘Temporal Interpretation’. Unpublished manuscript, University of Southern California . Groenendijk Jeroen , Stokhof Martin 1975 : ‘Modality and Conversational Information’ . Theoretical Linguistics , 2 ( 1–3 ), pp. 61 – 112 . WorldCat Hacking Ian 1967 : ‘Possibility’ . Philosophical Review , 76 ( 2 ), pp. 143 – 68 . Google Scholar Crossref Search ADS WorldCat Hacquard Valentine 2006: ‘Aspects of Modality’. Ph.D. thesis, Massachusetts Institute of Technology . Hacquard Valentine 2011 : ‘Modality’. In Claudia Maienborn, Klaus von Heusinger, and Paul Portner (eds.), Semantics: An International Handbook of Natural Language Meaning , Vol. 2 , pp. 1484 – 1515 . Berlin and Boston, MA : De Gruyter Mouton . Google Preview WorldCat COPAC Hunter Julie 2012 : ‘Now: A Discourse-Based Theory’. In Maria Aloni, Vadim Kimmelman, Floris Roelofsen, Galit W. Sassoon, Katrin Schulz, and Matthijs Westera (eds.), Logic, Language and Meaning: 18th Amsterdam Colloquium, Revised Selected Papers , pp. 371 – 80 . Berlin and Heidelberg : Springer . Google Preview WorldCat COPAC Iatridou Sabine 1990 : ‘The Past, the Possible, and the Evident’ . Linguistic Inquiry , 21 ( 1 ), pp. 123 – 9 . WorldCat Kaplan David 1989 : ‘Demonstratives’ . In Almog Joseph , Perry John , Wettstein Howard (eds.), Themes From Kaplan , pp. 481 – 563 . Oxford : Oxford University Press . Google Preview WorldCat COPAC Khoo Justin 2015 : ‘On Indicative and Subjunctive Conditionals’ . Philosophers' Imprint , 15 ( 32 ), pp. 1 – 40 . WorldCat Kratzer Angelika 1977 : ‘What “Must” and “Can” Must and Can Mean’ . Linguistics and Philosophy , 1 ( 3 ), pp. 337 – 55 . Google Scholar Crossref Search ADS WorldCat —— 1981 : ‘The Notional Category of Modality’ . In Eikmeyer Hans-Jürgen , Reiser Hannes (eds.), Words, Worlds and Contexts: New Approaches in Word Semantics , p. 39 – 74 . Berlin : De Gruyter . Google Preview WorldCat COPAC —— 1991 : ‘Modality’ . In von Stechow Arnin , Wunderlich Dieter (eds.), Semantics: An International Handbook of Contemporary Research , pp. 639 – 50 . Berlin : De Gruyter . WorldCat MacFarlane John 2011 : ‘Epistemic Modals Are Assessment-Sensitive’ . In Egan Andy , Weatherson Brian (eds.), Epistemic Modality , pp. 144 – 78 . Oxford : Oxford University Press . Google Preview WorldCat COPAC Moore G. E. 2013 : Commonplace Book: 1919–1953. Edited by Casimir Lewy . London : Routledge . Google Preview WorldCat COPAC Moss Sarah 2018 : Probabilistic Knowledge . Oxford : Oxford University Press . Google Preview WorldCat COPAC Ogihara Toshiyuki 2007 : ‘Tense and Aspect in Truth-Conditional Semantics’ . Lingua , 117 ( 2 ), pp. 392 – 418 . Google Scholar Crossref Search ADS WorldCat Partee Barbara Hall 1973 : ‘Some Structural Analogies between Tenses and Pronouns in English’ . Journal of Philosophy , 70 ( 18 ), pp. 601 – 9 . Google Scholar Crossref Search ADS WorldCat Roberts Craige 2012 : ‘Information Structure in Discourse: Towards an Integrated Formal Theory of Pragmatics’ . Semantics and Pragmatics , 5 , pp. 1 – 69 . WorldCat Stephenson Tamina 2007 : ‘Judge Dependence, Epistemic Modals, and Predicates of Personal Taste’ . Linguistics and Philosophy , 30 ( 4 ), pp. 487 – 525 . Google Scholar Crossref Search ADS WorldCat Teller Paul 1972 : ‘Epistemic Possibility’ . Philosophia , 2 ( 4 ), pp. 303 – 20 . Google Scholar Crossref Search ADS WorldCat Thomason Richmond H. 1984 : ‘Combinations of Tense and Modality’ . In Gabbay D. , Guenthner F. (eds.), Handbook of Philosophical Logic, Volume II: Extensions of Classical Logic , pp. 135 – 65 . Dordrecht : D. Reidel . Google Preview WorldCat COPAC von Fintel Kai , Gillies Anthony S. 2008 : ‘CIA Leaks’ . Philosophical Review , 117 ( 1 ), pp. 77 – 98 . Google Scholar Crossref Search ADS WorldCat —— 2011 : ‘Might’ Made Right’ . In Egan Andy , Weatherson Brian (eds.), Epistemic Modality , pp. 108 – 30 . Oxford : Oxford University Press . Google Preview WorldCat COPAC Williamson Timothy 2016 : ‘Modal Science’ . Canadian Journal of Philosophy , 46 ( 4–5 ), pp. 453 – 92 . Google Scholar Crossref Search ADS WorldCat Yalcin Seth 2011 : ‘Nonfactualism about Epistemic Modality’ . In Egan Andy , Weatherson Brian (eds.), Epistemic Modality , pp. 295 – 332 . Oxford : Oxford University Press . Google Preview WorldCat COPAC © 2019 The Author(s) This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/open_access/funder_policies/chorus/standard_publication_model)