Safety and Unawareness of Error-Possibility

Abstract

In this paper, I first seek a relatively plausible formulation of the safety principle. To this end, I refute a recent form of safety by Duncan Pritchard and then defend another weaker form of safety as a necessary condition for knowledge. Second, and more importantly, I point out that this weaker safety is still insufficient, in that it neglects one’s belief regarding nearby error-possibilities—a factor that is largely omitted in the literature but could determine whether or not one knows. I then develop a safety-based account of knowledge that incorporates this element of belief concerning nearby error-possibilities. It is argued that such an account addresses various problematic cases and delivers useful resources for accommodating the phenomenon of knowledge-defeat, while preserving the important anti-skeptical power behind a safety principle.

Notes

1 See, among others, Beddor and Pavese (2018), Broncano-Berrocal (2014), Carter (2010, 2013), Greco (2016), Grundmann (2020), Manley (2007), Peet and Pitcovski (2018), Pritchard (2005, 2012), Sainsbury (1997), Sosa (1999), and Williamson (2000).

2 Among safety theorists, this is most explicitly endorsed by Pritchard (2005). But see Baumann (2008), Kvanvig (2008), and Xu (forthcoming) for criticisms of Pritchard’s endorsement of possible world semantics. See also Whiting (2020) for an alternative version of safety that appeals to epistemic, as opposed to metaphysical, possibilities.

3 The idea of relativizing a modal account of knowledge to the basis/method of a belief originates with Nozick (1981, 179). See also Sosa (2007, 25–27) and Pritchard (2005, 2007a) for discussions of basis-relativized safety.

4 See, e.g., Pritchard (2005, 73–75).

5 See Goldman (1976).

6 Cf. Zhao (2020, 854–855; forthcoming) and Zhao and Baumann (2021).

7 These three versions of safety are differentiated primarily according to the quantifier. Another way to differentiate versions of safety is in terms of beliefs. That is, there can be different safety principles depending on whether safety locally requires truths of the same belief as formed in the actual world, or it globally requires truths of a range of different beliefs. This paper focuses on the version of safety that only takes the same belief into account, but the points I will make below can be carried over to a safety principle that abandons this ‘same belief’ restriction. See, e.g., Bernecker (2020), Hirvelä (2019), Melchior (2021), and Zhao (2019) for discussions of the globalized version of safety.

8 Safety in general has come under attack recently. A number of authors have presented counterexamples to the effect that unsafe knowledge is possible (Baumann 2008; Bogardus 2014; Comesaña 2005; Freitag 2014; Hiller & Neta 2007; Kelp 2009; McEvoy 2014, Neta & Rohrbaugh 2004). Since in these counterexamples subjects’ relevant beliefs appear to be false in many nearby possible worlds, these authors are targeting all three versions of safety listed here.

However, whether these counterexamples work or not is controversial. (See, e.g., Beddor & Pavese 2018; Broncano-Berrocal 2014, 2018; Grundmann 2020; Pritchard 2009 for defenses of safety against some of these counterexamples.) Also, different from the above critics of safety, part of my project in this paper is to identify a version of safety that is more defensible than others, as opposed to arguing against the safety principle simpliciter.

9 But see Baumann (2016, chapter 4) and Sosa (2015, 119–123).

10 See Sosa (1999, 145).

11 Cf. Neil (2019). I will reconsider Safety (Weak) and Safety (Strong) in Section 4.

12 See, e.g., Pritchard (2005). Some forms of luck (such as ‘evidential luck’) are compatible with knowledge. See Engel (1992) and Pritchard (2005) for relevant discussions.

13 In judging the luckiness of a belief, Pritchard (2007a, 292; 2015, 100–102) claims that a full-fledged Modal VEL not only takes into account the number(/proportion) of not-p worlds, but also takes into account the closeness of those not-p worlds, such that the closer a not-p world is to the actual world, the more luckily one forms a true belief in p. As such, one might argue that Ava’s belief is of a high degree of (Modal) VEL, given that the scenario in which Ava sees the fake barn is a very close possible world to the actual world. So, Ava’s belief is too (veritically) lucky to be counted as knowledge. However, I find this objection unconvincing: Given that in almost all close possible worlds p is true, taking into account the degree of closeness of (a small number of) not-p worlds hardly alters anything. Thanks to an anonymous referee.

14 In addition to the above modal and probabilistic versions of VEL, is there any other way to formulate VEL? Focusing on the literature on luck, the major remaining candidate for an account of luck is the ‘lack of control account’. according to which, roughly, an event is lucky if the subject has little control over it. Such an account, however, can hardly be adopted to formulate VEL, since it cannot accommodate the fact that we (typically) form beliefs involuntarily. For instance, I am clearly not lucky to have such beliefs as ‘I am now seeing purple’ when this is so, although I have no control over the formation of this belief (Pritchard 2005, 127).

15 Specifically, the objectors have to appeal to a relatively strong form of safety—e.g., Safety (Strong), Safety (Moderate)—which requires that a safe belief be true in all the very close worlds.

16 Cf. Baumann (2008).

17 To be clear, one cannot argue that Ava lacks knowledge on the grounds that One Fake Barn, specified in the above way, is analogous to the lottery case. Again, it should be stressed that in the lottery case the subject is aware of the possibility of winning the lottery—the possibility in which her belief (that she will lose the lottery) is false. Besides, the subject’s belief about her losing the lottery is based on probabilistic considerations. By contrast, in the present example, Ava is not only unaware of the not-p possibility (that is, she is unaware that there is a fake barn), but her barn-belief is based on perception. In fact, as I will argue in the next section, once we contrast One Fake Barn with its lottery variety, safety theorists will be faced with more challenges.

18 One may argue that One Fake Barn* and the lottery case are not analogous, in that the subject’s belief in the former case is at least partly perceptual but in the lottery case Lottie’s belief is non-perceptual.

Two responses. First, consider the distinction between ‘factors that trigger belief-formation’ and ‘factors that constitute the epistemic basis of a belief’. Enlightened Ava’s perceptual experience of the barn triggers her barn-belief, in the sense that her perceptual experience allows her to be in a position to form the belief that she is seeing a barn. However, as stipulated, it is her probabilistic evidence that constitutes the very epistemic basis of her belief—that is, it is such evidence that makes her adopt and sustain her barn-belief. But if enlightened Ava’s perceptual experience merely triggers her belief without being the basis of her belief, then such experience is not something epistemically relevant. As such, enlightened Ava’s perceptual experience does not mark an epistemic difference between her belief and Lottie’s belief.

Second, there is a sense in which even Lottie’s belief is ‘perceptual’: Just as enlightened Ava sees the barn and forms a barn-belief, we may stipulate that Lottie also sees the numbers on the ticket and forms a belief that she will lose. Again, these two cases go hand in hand.

19 It might be tempting to think that correct verdicts can be given once the method (/way) of belief-formation is properly described. However, this does not do the trick. Notice that in One Fake Barn the method is perceptual, whereas the method in One Fake Barn* is probabilistic evidence. In spite of these different methods, the modal profiles of the two cases are the same: both of the subjects’ beliefs formed with these different methods are true in the majority of nearby worlds where they see one of the many real barns, and false in a small number of nearby worlds where they encounter the fake barn.

20 cf. Pritchard (2008, 445).

21 There is inevitable vagueness associated with Safety (Weak): exactly how many close p-worlds should there be for a belief to be true in ‘most’ close possible worlds? (Is 80% enough, or 90%?) It seems there is no clear answer here. But I do not find such vagueness to be particularly problematic. First, this kind of vagueness affects other theories as well. Consider process reliabilism—the view that, roughly, if one’s belief that p is justified then it must be produced by a reliable cognitive process or method. No matter how one understands the notion of ‘reliability’ here (probabilistically or modally), there is unavoidable vagueness with regard to the degree of reliability. Second, notice that the concept of ‘safety’ itself involves vagueness. For example, it is unclear how low the crime rate in a neighborhood should be for the neighborhood to be counted as a safe one. But if the notion of safety is already vague, advocates of Safety (Weak) may claim that the vagueness with ‘most’ could indeed be a virtue, since such vagueness may well match the vagueness in the term ‘safety’.

22 See also Neil (2019).

23 It is worth mentioning here that (i) ‘nearby possible worlds in which S forms a false belief that p’ is different from (ii) ‘nearby possible worlds in which p is false’. The latter also takes into account the nearby possibilities in which p is false but S does not believe that p. Indeed, formulating Unawareness in terms of (i) is more accurate. To illustrate, consider the following case (which is a variant of Nozick’s grandmother example): Suppose that a grandmother sees her grandson doing well and thus believes that he is well. However, the grandson luckily survived a critical and risky surgery recently. In other words, there are close possible worlds in which the grandson does not survive the surgery. Furthermore, suppose that the grandmother knows about all of this. In this case, intuitively the grandmother knows that p—that her grandson is well—by seeing him doing well in the actual world. This is so even though she believes that there are close possibilities in which p is false—that is, possibilities in which her grandson does not survive. However, since the grandmother does not believe that there are close possibilities in which she forms a false belief that p, Unawareness correctly predicts that she knows that p. For this reason, Unawareness is more plausible than the account proposed in my recent article (Zhao forthcoming), which is formulated in terms of (ii).

24 Thus, in the present context, when I speak of ‘a belief that p is being defeated’ I mean that the belief that p is prevented from constituting knowledge because one has a doxastic defeater—i.e., another belief regarding close error-possibilities concerning p. I will leave it as an open question whether such a doxastic defeater defeats one’s justification/warrant. cf. Pritchard (2018, 3069).

25 See also Pritchard (2018), Zhao (2020; forthcoming), and Zhao and Baumann (2021, 124–125) for related discussions on defeat.

26 See Sosa (1999) and Pritchard (2005).

27 Is the present account of knowledge internalist or externalist? If internalism is defined as the claim that whatever turns true belief into knowledge must be exhausted by factors that the subject has access to, then our account is clearly not internalist, for it incorporates Safety (Weak), which is externalist in spirit. But to the extent that according to the present account one’s belief regarding nearby error-possibilities matters for knowledge and that beliefs are items that one has access to, the account is more internalist than an account of knowledge that merely incorporates a safety condition. Thanks to an anonymous referee for pressing the need to clarify this.

28 One may ask why Unawareness is couched in terms of belief, not credence. The answer is that credence is too fine-grained in the present context to serve as a no-defeater condition. Specifically, suppose that any credence in close error-possibilities concerning p prevents one from knowing that p; then our proposal is too demanding. One may have, say, 0.01 credence that some nearby error-possibilities obtain, but this could hardly be an obstacle for one’s knowledge that p. (An example: suppose a highly reliable speaker tells me a true proposition that p. I thereby believe that p; but I have a 0.01 credence that the speaker is wrong on this occasion. Clearly, entertaining such a low credence does not prevent me from knowing that p from the reliable speaker’s true testimony.)

Thus, a potentially more plausible candidate is ‘sufficiently high credence of nearby error-possibility concerning p’. But this immediately invites unpalatable vagueness: how high should a credence be in order to be counted as sufficiently high? Is it 0.5, or 0.6, etc.? Or is ‘sufficiently high’ context-dependent, so that in different contexts, different credences are regarded as high enough to have the defeating force? A plausible answer could hardly be given to these questions. Fortunately, formulating Unawareness in terms of ‘belief’ could avoid the complications here.

29 For an argument to the effect that the subject retains knowledge in such a case, see Lasonen-Aarnio (2010). cf. Janvid (2017).

30 Cf. Carter and Pritchard (2016).

31 Cf. Greco (1994).

32 Thanks to an anonymous referee, Kent Staley, and Katherine Sweet for this objection. Thanks to Yiling Zhou for discussions.

33 Or, if you are a reductionist about testimonial knowledge, add that the patient has good reasons to believe that the doctor is trustworthy, sincere, and competent.

34 See also Zhao (forthcoming, 14–15), where I argue that the present observation also applies to the faculty of memory.

35 For helpful comments and discussions on earlier drafts of this paper, I am grateful to Ken Akiba, Peter Baumann, Anthony Bolos, John Greco, Joe Salerno, Tomoji Shogenji, Donald Smith, Kent Staley, Catherine Sutton, Katherine Sweet, Yingjin Xu, Yiling Zhou, participants of a Virtue Epistemology seminar at St. Louis University, and two anonymous referees for this journal. This research was supported by Fundamental Research Funds for the Central Universities (No. 2072021100).