1. Many factual findings at trial concern past events, but the disputed facts at trial may also involve current or future events (for example, current injuries and future earnings).

2. Within epistemology, issues involving law typically fall under the domain of “social epistemology,” which concerns, among other issues, transmission of knowledge and the use of evidence within social institutions. See generally Goldman, Alvin I., A Guide to Social Epistemology, in Social Epistemology: Essential Readings (Goldman, Alvin & Whitcomb, Dennis eds., 2011), at 11Google Scholar; Alvin Goldman & Thomas Blanchard, Social Epistemology, in Stanford Encyclopedia of Philosophy (2015) (https://plato.stanford.edu/entries/epistemology-social/ ) (“Since evidence, confidence, reasonable doubt and so on are epistemological notions, these [legal] rules are of interest to the social epistemologist.”).

3. See, e.g., Susan Haack, Evidence Matters: Science, Proof, and Truth in the Law (2014); Ho Hock Lai, A Philosophy of Evidence Law (2008); Larry Laudan, Truth, Error, and Criminal Law: An Essay in Legal Epistemology (2006); Alex Stein, Foundations of Evidence Law (2005); Pardo, Michael S., The Gettier Problem and Legal Proof, 16 Legal Theory 37 (2010)CrossRefGoogle Scholar; Pardo, Michael S. & Allen, Ronald J., Juridical Proof and the Best Explanation, 27 Law & Phil. 223 (2008)CrossRefGoogle Scholar; Redmayne, Mike, Exploring the Proof Paradoxes, 14 Legal Theory 281, 299 (2008)CrossRefGoogle Scholar; Schauer, Frederick, In Defense of Rule-Based Evidence Law—And Epistemology Too, 5 Episteme 295 (2008)CrossRefGoogle Scholar; Amaya, Amalia, Justification, Coherence, and Epistemic Responsibility in Fact-Finding, 5 Episteme 306 (2008)CrossRefGoogle Scholar; Goldman, Alvin I., Quasi-Objective Bayesianism and Legal Evidence, 42 Jurimetrics J. 237 (2002)Google Scholar; Allen, Ronald J. & Leiter, Brian, Naturalized Epistemology and the Law of Evidence, 87 Va. L. Rev. 1491 (2001)CrossRefGoogle Scholar; Brewer, Scott, Scientific Expert Testimony and Intellectual Due Process, 107 Yale L.J. 1535 (1998)CrossRefGoogle ScholarPubMed. For an overview of the literature, see Hock Lai Ho, The Legal Concept of Evidence, in Stanford Encyclopedia of Philosophy (2015) (http://plato.stanford.edu/entries/evidence-legal/).

4. On safety, see Sosa, Ernest, How to Defeat Opposition to Moore, 13 Phil. Persp. 141, 142 (1999)Google Scholar (“S would believe that p only if it were so that p.”). On sensitivity, see Robert Nozick, Philosophical Explanations (1981), at 172 (“If p weren't true, S wouldn't believe that p.”). See also Pritchard, Duncan, Safety, Sensitivity, and Anti-Luck Epistemology, in The Oxford Companion to Scepticism (Greco, John ed., 2008)Google Scholar. Consider my (true) belief that there is unread email in my inbox. If I could easily be mistaken about this, then my belief is unsafe. If I would believe this even if there were no unread email, then my belief is insensitive.

5. Although similar epistemological issues arise in the legal context, it is important to note that the law's interests and concerns diverge from some of those at issue in the related philosophical literature. For example, safety may play an important role for legal evidence regardless of whether it is a necessary condition for knowledge. And issues regarding skepticism are less of a concern in the legal context.

6. Greco, John, Better Safe Than Sensitive, in The Sensitivity Principle in Epistemology (Becker, Kelly & Black, Tim eds., 2012), at 194Google Scholar (“The spirit of a safety condition is that, in cases of knowledge, S would not easily go wrong by believing as she does.”).

7. On the prevalence of false confessions, see generally Garrett, Brandon L., Contaminated Confessions Revisited, 101 Va. L. Rev. 395, 395–398 (2015)Google Scholar; Griffin, Lisa Kern, Silence, Confessions, and the New Accuracy Imperative, 65 Duke L.J. 697 (2016)Google Scholar.

8. Greco, supra note 6, at 194 (“The spirit of a sensitivity condition is that, in cases of knowledge, one would notice if things were different.”).

9. Id. at 195 (“[A] belief can be safe without being sensitive, and sensitive without being safe.”).

10. And, of course, findings may also be safe and sensitive as well as unsafe and insensitive. Both safety and sensitivity are distinct from probability. Two items of evidence may have the same probabilities associated with them, but findings based on them may differ in terms of safety or sensitivity.

11. Probative value plays a central role in the admissibility and sufficiency of legal evidence. See, e.g., Fed. R. Evid. 403 (authorizing courts to exclude evidence when its “probative value” is substantially outweighed by countervailing considerations).

12. Enoch, David & Fisher, Talia, Sense and “Sensitivity”: Epistemic and Instrumental Approaches to Statistical Evidence, 67 Stan. L. Rev. 557 (2015)Google Scholar.

13. Enoch, David, Spectre, Levi & Fisher, Talia, Statistical Evidence, Sensitivity, and the Value of Legal Knowledge, 40 Phil. & Public Affairs 197 (2012)CrossRefGoogle Scholar. See also Blome-Tillmann, Michael, Sensitivity, Causality, and Statistical Evidence in Courts of Law, 4 Thought 102 (2015)Google Scholar (referring to the sensitivity account as a “recently influential” account of legal evidence).

14. To be clear, I do not claim that safety provides a complete account of legal evidence or that safety explains all aspects of the law of evidence. Rather, I argue for the more modest claims that safety is an important consideration for legal evidence and that it plays a greater explanatory role than sensitivity.

15. See David Lewis, On the Plurality of Worlds (1986). For overviews, see Jon Divers, Possible Worlds (2002); Christopher Menzel, Possible Worlds, in Stanford Encyclopedia of Philosophy (2016) (http://plato.stanford.edu/entries/possible-worlds/).

16. See Duncan Pritchard, In Defense of Moderate Anti-luck Epistemology, in The Sensitivity Principle, supra note 6, at 177; Greco, supra note 6, at 194–195.

17. For other examples of legal analyses using possible worlds, see Michael S. Moore, Causation and Responsibility: An Essay in Law, Morals, and Metaphysics (2009), at 390 (“causation is best worked out using the ‘possible worlds’ conception”); Solum, Lawrence B., Constitutional Possibilities, 83 Ind. L.J 307, 316–320 (2008)Google Scholar.

18. “Sensitivity” throughout this article refers to the epistemological concept and not to other uses of the term. In particular, and most importantly, it should not be confused with the common use of “sensitivity” in diagnostic testing to refer to the “true positive rate”—e.g., the probability that a test will yield a correct positive result among those who in fact have a disease. See David H. Kaye & David A. Freedman, Reference Guide on Statistics, in Reference Manual on Scientific Evidence (3d ed. 2011), at 296. The diagnostic use of “sensitivity” is contrasted with a test's “specificity” (or true negative rate)—e.g., the probability that a test will yield a correct negative result among those who do not in fact have a disease. Id. at 298. Although there are some connections between these uses of “sensitivity,” in order to avoid confusion, readers familiar with the diagnostic use of the term should put that meaning to the side.

19. “Safety” throughout this article refers to the epistemological concept and not to other meanings of the term. In particular, epistemic safety is related to but distinct from the broader “unsafe verdict” standard for reviewing criminal appeals in the UK. See Criminal Appeal Act §2 (1995). A reliance on epistemically unsafe evidence may be one reason a verdict is “unsafe” in the latter sense. See Risinger, D. Michael, Unsafe Verdicts: The Need for Reformed Standards of the Trial and Review of Factual Innocence Claims, 41 Hous. L. Rev. 1281, 1332 (2004)Google Scholar.

20. These include rules regulating experts, witnesses, exhibits, impeachment, hearsay, and character. See infra notes 110–117 and accompanying text.

21. Some scholars have questioned the relevance of epistemology (or aspects of it) for legal proof. See, e.g., Stein, supra note 3, at 59 (“Adjudicators do not even purport to satisfy the ‘justified true belief’ standard or similar criteria for knowledge.”); Enoch et al., supra note 13, at 211 (“[T]he law should not care about knowledge, or indeed about epistemology in general.”).

22. See Fed. R. Evid 102 (“These rules should be construed . . . to the end of ascertaining the truth and securing a just determination.”); Tehan v. United States, 382 U.S. 406, 416 (1966) (“The basic purpose of a trial is the determination of truth.”).

23. Indeed, legal rules and procedures that regulate the “sufficiency of evidence” require judges to determine whether particular findings are justified based on the evidence and the burden of proof. See, e.g., Fed. R. Civ. P. 56 (summary judgment); Fed. R. Civ. P. 50 (judgment as a matter of law); Fed. R. Crim. P. 29 (motion for a judgment of acquittal based on insufficient evidence); Jackson v. Virginia, 443 U.S. 307 (1979). Thus, with the exception of acquittals in criminal cases, the fact that a jury finds the evidence persuasive is neither sufficient, nor is it necessary, to constitute a legal judgment. Criminal acquittals present a special case—in order to preserve the power of jury nullification, acquittals are not reviewable on sufficiency grounds.

24. See Anderson v. Liberty Lobby, Inc., 477 U.S. 242, 248 (1986) (explaining that sufficiency of evidence in civil cases depends on what a “reasonable jury” could find based on the evidence); Jackson v. Virginia, 443 U.S. 307, 318 (1979) (explaining that sufficiency of evidence in criminal cases depends on what a “rational trier of fact” could find).

25. In discussing legal proof, Jeremy Bentham famously declared that “the field of evidence is no other than the field of knowledge.” An Introductory View of the Rationale of Judicial Evidence, Works VI(5) (1843) (John Bowring ed., 2002).

26. See supra note 3.

27. See Fed. R. Evid. 602 (“A witness may testify only if evidence is introduced sufficient to support a finding that the witness has personal knowledge of the matter.”), 702 (regulating testimony based on an expert's “scientific, technical, or other specialized knowledge”).

28. See Redmayne, supra note 3, at 299 (rejecting knowledge as the aim of verdicts); Stein, supra note 3, at 59 (same). Moreover, knowledge requires belief but jurors may not necessarily believe their findings. See Beltran, Jordi Ferrer, Legal Proof and Fact Finders’ Belief, 12 Legal Theory 293 (2008)Google Scholar. But see Pardo, supra note 3 (discussing different ways in which knowledge may relate to verdicts); Thomson, Judith Jarvis, Liability and Individualized Evidence, 49 Law & Contemp. Probs. 199, 206–215 (1986)CrossRefGoogle Scholar (discussing similarities between knowledge and verdicts); Anthony Duff et al., The Trial on Trial: Towards a Normative Theory of the Criminal Trial (2007), at 87–91 (arguing that criminal verdicts require knowledge); Sarah Moss, Probabilistic Knowledge (forthcoming 2018) (arguing that legal verdicts require knowledge of probabilistic contents).

29. Support for this idea may follow from the structural similarities between knowledge and verdicts. In particular, legal verdicts have analogs to the three components in the traditional conception of knowledge as “justified true beliefs.” First, both verdicts and knowledge require some level of cognitive endorsement of a proposition (a belief in the case of knowledge or a finding in the case of a verdict). Second, both require some level of epistemic support, justification, or warrant. Third, both aim at truth (factual accuracy).

30. In other words, even though sensitivity and safety each arise in the context of philosophical debates about knowledge, we can prescind from any debates about how each relates to knowledge.

31. See supra notes 15–16. Philosophical debates about the metaphysical status of possible worlds (among other issues) are outside the scope of this article and nothing in the analysis will turn on them. See Menzel, supra note 15. Similarly, philosophical debates about how best to analyze or account for subjunctive conditionals are also outside the scope of this article. See Jonathan Bennett, A Philosophical Guide to Conditionals (2003); Stalnaker, Robert, A Theory of Conditionals, in Studies in Logical Theory (American Philosophical Quarterly Supplementary Monograph Series) (Rescher, Nicholas ed., 1968), at 98Google Scholar; David Lewis, Counterfactuals (1973); Goodman, Nelson, The Problem of Counterfactual Conditionals, 44 J. Phil. 113 (1947)CrossRefGoogle Scholar.

32. Greco, supra note 6, at 194 (“A more straightforward method is to state the safety and sensitivity conditions directly, by means of a possible-worlds heuristic.”); Pritchard, supra note 16, at 174 (“Sensitivity is usually cashed out in modal terms as demanding that in the closest possible world in which what the agent actually believes is false, the agent no longer believes it on the same basis as in the actual world.”).

33. See, e.g., Greco, supra note 6, at 195 (“Think of a space of possible worlds centered on the actual world and branching out according to some appropriate similarity ordering.”); Martin Smith, Between Probability and Certainty: What Justifies Belief (2016), at 106–108.

34. Solum, supra note 17, at 318.

35. Id. Also outside the scope of this article are the complex issues about how best to measure similarity between possible worlds and whether there is an objective metric for doing so when evaluating conditionals. See Ippolito, Michela, How Similar Is Similar Enough?, 9 Semantics & Pragmatics 1 (2016)Google Scholar. In the legal-proof context (as in other contexts), this issue will depend heavily on the context of the case and will be influenced by the disputed facts, the available evidence, and the alternative arguments being made by the parties. The process of legal proof relies to a large extent on the parties to present the relevant alternative possibilities. See Pardo & Allen, supra note 3; Allen, Ronald J., The Nature of Juridical Proof, 13 Cardozo L. Rev. 373 (1991)Google Scholar.

36. Indeed, referring to possible worlds in the process of analyzing legal evidence has a natural fit with the law of evidence. Two foundational concepts in evidence doctrine—relevance and probative value—are each defined in terms of comparisons with alternative possibilities. The federal rules of evidence, for example, define “relevant evidence” as evidence having “any tendency to make a fact [of consequence] more or less probable than it would be without the evidence.” Fed. R. Evid. 401. This definition requires judges to compare the world in which this item of evidence is admitted with an alternative possible world in which the evidence does not exist and then to assess whether the likelihood of the fact being true changes. See, e.g., United States v. Stever, 603 F.3d 747 (9th Cir. 2010). Similarly, the United States Supreme Court has explained that the probative value of evidence depends not only on the connection between the evidence and the fact for which it is offered—it also requires comparisons with possible evidentiary alternatives. See Old Chief v. United States, 519 U.S. 172 (1997) (explaining that probative value depends on “comparing evidentiary alternatives”); Fed. R. Evid. 403. This necessarily requires judges to assess the value of evidence by considering different possible ways in which parties may prove disputed facts.

37. See Enoch et al., supra note 13; Enoch & Fisher, supra note 12.

38. Kelly Becker & Tim Black, The Resilience of Sensitivity, in The Sensitivity Principle, supra note 6, at 1 (referring to sensitivity as “a very simple and intuitively compelling idea in epistemology”). But see Ernest Sosa, Judgment & Agency (2015), at 123 (referring to sensitivity as a “siren call” with “execrable consequences”).

39. See Nozick, supra note 4, at 172; Dretske, Fred, Conclusive Reasons, 49 Australasian J. Phil. 1 (1971)CrossRefGoogle Scholar; Goldman, Alvin, Discrimination and Perceptual Knowledge, 78 J. Phil. 771 (1976)CrossRefGoogle Scholar; DeRose, Keith, Solving the Sceptical Puzzle, 104 Phil. Rev. 1 (1995)CrossRefGoogle Scholar. For recent discussions, see The Sensitivity Principle, supra note 6.

40. Beliefs may also fail to be true for reasons other than being false—for example, they may have no truth value or no content. See Manley, David, Safety, Content, Apriority, Self-Knowledge, 104 J. Phil. 403 (2007)CrossRefGoogle Scholar.

41. Sensitivity also features prominently in philosophical discussions of general issues such as skepticism, see Pritchard, supra note 4, closure, see Sherrilyn Rousch, Sensitivity and Closure, in The Sensitivity Principle, supra note 6, at 242, and luck, see Duncan Pritchard, Epistemic Luck (2005), at 152–173. Sensitivity also potentially explains some “Gettier cases.” See Gettier, Edmund L., Is Justified True Belief Knowledge?, 23 Analysis 121 (1963)CrossRefGoogle Scholar.

42. These examples are part of a family of puzzling cases involving lotteries. See John Hawthorne, Knowledge and Lotteries (2004).

43. See Pritchard, supra note 16, at 177.

44. Id. (“the sensitivity principle offers us a very attractive way of dealing with the lottery problem”).

45. Id. at 174.

46. Sensitivity is also typically limited to beliefs formed via the same method. See Nozick, supra note 4, at 179. This, however, also raises additional difficulties in delineating what counts as the same method. For discussions, see Kelley Becker, Methods and How to Individuate Them, in The Sensitivity Principle, supra note 6, at 81; Peter Baumann, Nozick's Defense of Closure, in id. at 17 (“[I]t has proved notoriously difficult to identify the method used by the subject in a systematic, principled, and non-arbitrary way . . . [this] so-called generality problem probably has no solution.”).

47. The context for these objections is whether sensitivity is necessary for knowledge.

48. In this context, “reliability” means a tendency to produce true beliefs. Alvin I. Goldman, Epistemology and Cognition (1986), at 26. This meaning is distinct from the use of “reliable” in some contexts to refer to merely consistent or similar results (even if false or invalid). For a discussion of these different meanings, see Haack, supra note 3, at 200–201. The use of “reliable” in evidence law also typically refers to the former (i.e., a tendency to produce true beliefs). See id. (discussing the meaning of “reliability” in the context of scientific expert testimony under Fed. R. Evid. 702). Reliability may refer to a process, method, or type of evidence in general or it may refer to close possible worlds. Reliability in general and in close possible worlds may diverge. See Manley, supra note 40, at 409 (“My ability to discriminate larks from other birds may be so reliable that there are only five token birds in the world that I mistake for larks. But if all five happen to be in my yard along with a real lark, that is enough to undermine my knowledge.”). Sensitivity fails to track the reliability of evidence both generally and in close possible worlds.

49. Sosa, supra note 4, at 145–146.

50. Id.; Pritchard, supra note 16, at 176 (discussing this example: “if Ernie doesn't have knowledge, then it would appear that inductive knowledge is very hard to come by, since Ernie's inductive basis for his true belief is about as good as an inductive basis can be.”). Readers who reject knowledge in Trash Bag may find the following example more persuasive:

Sosa, supra note 38, at 119. Nothing in the analysis to follow, however, will depend on whether the beliefs in the examples constitute knowledge.

51. At trial, my testimony would likely be permissible as a lay opinion. See Fed. R. Evid. 701.

52. This example is based on Timothy Williamson, Knowledge and Its Limits (2000), at 159–160.

53. There is an analog to this point in the distinction between so-called “hard” and “easy” legal cases. See Frederick Schauer, Thinking Like a Lawyer (2009), at 20–24. The fact that some legal questions (typically at the appellate level) permit more than one plausible answer based on the legal materials does not mean that there are not right or correct answers in easy cases.

54. Enoch et al., supra note 13; Enoch & Fisher, supra note 12.

55. Enoch et al., supra note 13, at 199 (“[T]he distinction between statistical and individual evidence is a general one, and it seems to call for a general solution.”); Id. at 209 (“Sensitivity-like counterfactuals capture—often enough, in sufficiently central cases—an epistemically relevant feature of the distinction between statistical and individual evidence.”). They also contend that sensitivity explains “the prevailing legal doctrine” on evidentiary requirements. Enoch & Fisher, supra note 12, at 558.

56. Enoch et al., supra note 13, at 197; Enoch & Fisher, supra note 12, at 573–576. This famous example is based on dicta in Smith v. Rapid Transit, 58 N.E.2d 754 (Mass. 1945).

57. Enoch et al., supra note 13, at 197; Enoch & Fisher, supra note 12, at 573–576. Findings in experimental psychology support the idea that people tend to treat the two examples differently. See Wells, Gary L., Naked Statistical Evidence of Liability: Is Subjective Probability Enough?, 62 J. Personality & Soc. Psychol. 739 (1992)CrossRefGoogle Scholar.

58. Enoch et al., supra note 13, at 197.

59. Id. Three doctrinal issues are embedded in this quotation and should be distinguished: whether evidence is (1) sufficient to prove a fact (to a standard of proof), (2) admissible, or (3) relevant. Evidence may be relevant without being admissible or sufficient. And it may be admissible without being sufficient.

60. Id. at 209 (“Sensitivity-like counterfactuals capture—often enough, in sufficiently central cases—an epistemically relevant feature of the distinction between statistical and individual evidence.”); Enoch & Fisher, supra note 12, at 557 (claiming that sensitivity provides “a comprehensive answer to the statistical evidence debate”).

61. Enoch et al., supra note 13, at 206, 197.

62. Id. at 206–207. Smith challenges the conclusion that the eyewitness evidence is sensitive, Smith, supra note 33, at 62, and Blome-Tillman presents a variation in which the statistical evidence is sensitive, Blome-Tillman, supra note 13, at 106. I will accept for the analysis to follow that the eyewitness evidence is sensitive and the statistical evidence is insensitive.

63. Enoch et al., supra note 13, at 212. Based on their sensitivity analysis, they conclude that knowledge and epistemology also do not matter for law. Id. at 211 (“[T]he law should not care about knowledge, or indeed about epistemology in general.”). This claim is problematic for several reasons. First, the fact that sensitivity is not important for law does not mean that other epistemic notions are likewise unimportant. Second, epistemology may contribute to law even if verdicts do not require knowledge. See Section II. Finally, the manner in which they challenge the relevance of knowledge is itself problematic. They ask us to compare two different standards for criminal convictions: World A in which jurors “only convict when they know,” and World B in which “the chances of System B convicting an innocent are lower.” Enoch et al., supra note 13, at 212. They contend that the law should prefer System B. Id. The choice presented, however, is a false one. Knowledge is “factive”—if something is known, then it is true. Therefore, every conviction in World A will be true and World B cannot produce fewer false convictions.

64. Although they reject epistemic reasons, they argue that “instrumental” reasons may justify a preference for sensitive evidence. Enoch et al., supra note 13, at 201; Enoch & Fisher, supra note 12, at 583. These reasons include the ex ante incentives for actors deciding whether to comply with the law. Cf. Sanchirico, Chris William, Character Evidence and the Object of Trial, 101 Colum. L. Rev. 1227 (2001)CrossRefGoogle Scholar (arguing that ex ante incentives justify the character rules). They illustrate their claim with the following example from the evidence literature:

Enoch et al., supra note 13, at 217. The evidence is insensitive: it would be the same regardless of whether John crashed the gate. By contrast, a piece of individualized evidence (“a videotape”) that is “probabilistically equivalent . . . seems perfectly fine.” Id. at 207. Instrumental reasons, they argue, justify this distinction. Using the statistical evidence “almost entirely annihilates” John's incentive not to break the law because he will be held liable “regardless of whether or not he buys a ticket.” Id. at 217–218. When evidence is sensitive, they argue, “there is no similar incentive-corrupting effect.” Id. This “instrumental” rationale is generally outside the scope of this article; however, to the extent that this rationale depends on aligning the statistical-individual distinction with the sensitive-insensitive distinction, the critique below applies to the instrumental account as well.

65. The critique thus challenges their descriptive claims about the relationships between sensitivity, evidence, and legal doctrine.

66. See Enoch et al., supra note 13, at 221 n.38; Enoch & Fisher, supra note 12, at 591. Although many examples of DNA evidence are sensitive, the evidence will be insensitive when there are close possible worlds in which innocence and matching samples coexist. This may be the case, for example, because (1) the DNA test does not discriminate among relatives (e.g., Y-STR testing) and there is a close possible world in which a matching relative commits the crime, or (2) there is a close world in which DNA is present for reasons unrelated to the crime (e.g., the defendant previously worked in the location where the sample was collected). For discussion and examples, see Erin E. Murphy, Inside the Cell: The Dark Side of Forensic DNA (2015), at 33.

67. In Enoch & Fisher, supra note 12, at 585–586, they acknowledge these examples. They qualify their claims by noting that they are “not interested in all instances in which statistical assessments are used as evidence.” Id. at 586. Rather, their focus is on cases like Blue Bus in which the evidence is “the base rate for the defendants’ liability.” Id. This qualification, however, is inconsistent with the thesis that sensitivity provides a comprehensive account of the general distinction between statistical and individualized evidence. Id. at 557 (“The aim of this Article is to provide a comprehensive answer to the statistical evidence debate.”); Enoch et al., supra note 13, at 199 (“[T]he distinction between statistical and individual evidence is a general one, and it seems to call for a general solution.”). Cases like Blue Bus are a tiny portion of statistical-evidence cases (if they exist at all); statistical-evidence cases typically involve the integration of statistical evidence with other evidence on both sides (including the failure to provide other evidence), even when one side attempts to rely on base-rate evidence. See Koehler, Jonathan J., When Do Courts Think Base Rate Statistics Are Relevant?, 42 Jurimetrics J. 373 (2002) (discussing examples)Google Scholar.

68. Tyson Foods, Inc. v. Bouaphakeo, 136 S. Ct. 1036 (2016).

69. Tyson Foods, 136 S. Ct. at 1043–1046. The Court rejected the defendant's argument that such evidence was improper because of possible variations involving individual employees. Id. In doing so, the Court rejected a categorical rule regarding the use of statistical evidence in class actions. Id. (“A representative or statistical sample, like all evidence, is a means to establish or defend against liability. Its permissibility turns not on the form a proceeding takes—be it a class or individual action—but on the degree to which the evidence is reliable in proving or disproving the elements of the relevant cause of action.”).

70. The significance of these categories is that they are frequently admissible despite being insensitive. Therefore, sensitivity cannot explain individualized evidence. It might be objected that in most cases evidence in these categories will not be sufficient by itself to prove liability (or to support a conviction). But this is a red herring because no single item of evidence is typically sufficient. Even cases based on one prominent item of evidence—such as confessions or “cold hit” DNA matches—are assessed in conjunction with all other evidence, as well as the possible reasons for not producing additional evidence.

71. See, e.g., Fed. R. Evid. 404(a)(2) (character evidence in criminal cases), 413–415 (defendant's prior acts of sexual assault or child molestation).

72. Fed R. Evid. 404(b)(2).

73. The relevance of impeachment evidence concerns the credibility of witnesses and this evidence is typically going to be insensitive. Consider, for example, evidence of bias on the part of a witness because of a prior relationship with one of the parties. See also Fed. R. Evid. 608 (impeachment with character evidence), 609 (prior convictions), 613 (prior inconsistent statements).

74. Fed. R. Evid. 406.

75. Fed. R. Evid. 702–703.

76. See, e.g., Fed. R. Evid. 803(6) (records of regularly conducted activities, including business records), 803(8) (public records), 803(21) (reputation concerning character), 803(22) (judgments of previous convictions).

77. This is not to suggest that sensitivity is never a relevant consideration—it may play an instrumental role in some types of cases. See supra note 64. Rather, the key point is that neither the statistical-individual distinction nor the admissibility-inadmissibility distinction can be explained in terms of the sensitivity-insensitivity distinction. None of these three distinctions (sensitivity-insensitivity, statistical-individual, admissible-inadmissible) can be explained in terms of another. Indeed, there are four different types of admissible evidence: (1) sensitive statistical, (2) insensitive statistical, (3) sensitive individual, and (4) insensitive individual—and the same four possibilities for inadmissible evidence.

78. On reliability, see supra note 48.

79. Steven Luper, False Negatives, in The Sensitivity Principle, supra note 6, at 222 (“[T]he closest worlds in which p is false might be remote indeed. They might be worlds in which the laws of physics are very different or even nonexistent. Does it really matter that we get false positives in such worlds?”).

80. A brief word on methodology. The purpose of these simplified examples is to illustrate the conceptual relationship between sensitivity and legal evidence by presenting extremes in the different categories. The notes provide additional examples and cases that fit the categories. My thanks to an anonymous referee for suggesting examples along these lines.

81. See Fed. R. Evid. 702.

82. See United States v. Dent, 149 F.3d 180 (3d Cir. 1998).

83. Similar to the belief in Speed, see supra note 52 and accompanying text, a finding based on the evidence is within a range of reliable inference even though it lacks sensitivity.

84. See United States v. Hamel, 551 F.2d 107 (6th Cir. 1977). On authentication, see Fed. R. Evid. 901. Similar to the belief in Trash Bag, see supra note 49 and accompanying text, the mere fact that the evidence could lead to a false inference (in the closest world in which the finding is false) does not undermine the reliability of the evidence.

85. See, e.g., Tyson Foods, Inc. v. Bouaphakeo, 136 S. Ct. 1036, 1048–1049 (2016) (“[T]he study here could have been sufficient to sustain a jury finding as to hours worked if it were introduced in each employee's individual action.”). Importantly, this point applies to both statistical and individualized evidence: regardless of its form, evidence may be probative, reliable, and admissible, and yet fail to satisfy sensitivity.

86. See Fed. R. Evid. 702. Suppose, for example, that the technician does not have proper training or any specialized knowledge; the technician failed to follow proper protocols in the instant case; the technique itself is not generally accepted or recognized in the forensic science community; and the technique is of unknown reliability. Cf. Manuel v. City of Joliet, 580 U.S. __ (2017), slip op. at 1 (civil claim based on allegations of fraudulent forensic evidence and lies by technicians).

87. The evidence meets the test articulated by Enoch & Fisher, supra note 12, at 591: if the defendant had not committed the crime, he would, in all likelihood, not have been convicted.

88. As with the previous examples, the analysis above applies to both statistical and individualized evidence. Consider the following example involving an eyewitness:

Although the testimony may be admissible—under Federal Rule of Evidence 602, the witness would likely be allowed to testify that the cab appeared green to the witness—it will not be sufficiently probative to establish that Mr. Green's cab caused the harm. This is so because, by the witness's own admission (let's assume) the witness cannot distinguish Mr. Green's cab from nine of the ten other cabs in the town. But notice the evidence appears to satisfy sensitivity: in the closest world in which a non-green cab causes harm, the witness no longer testifies to seeing a green cab (instead, she testifies to seeing a blue one). For a similar example, see Greco, supra note 6, at 201.

89. This is not to deny, however, that sensitivity may play an instrumental role for some types of cases. See supra note 64.

90. Sosa, supra note 4; Pritchard, supra note 16.

91. In particular, philosophers have argued that safety explains issues such as closure, skepticism, and luck. Sosa, supra note 4 (skepticism); Luper, supra note 79 (closure); Pritchard, supra note 41 (luck).

92. See, e.g., Greco, supra note 6, at 195 (“S's belief that p is safe just in case: there are no close worlds where both S believes that p, and p is false.”). See also Manley, supra note 40. In assessing closeness, Pritchard distinguishes between the probability of the world obtaining and how easy it would be for that world to obtain. Pritchard, Duncan, Risk, 46 Metaphilosophy 436, 452–457 (2015)CrossRefGoogle Scholar.

93. Pritchard, supra note 16, at 180.

94. Id. Nothing in the analysis on legal evidence depends on whether this explanation succeeds with regard to knowledge (i.e., why Newspaper counts as knowledge and Lottery does not). See supra note 30 and accompanying text.

95. See supra notes 49–53 and accompanying text.

96. Greco, supra note 6, at 193 (“[S]afety just is reliability throughout a space of close counterfactual situations.”). On reliability, see supra note 48.

97. See, e.g., Neta, Ran & Rohrbaugh, Guy, Luminosity and the Safety of Knowledge, 85 Pac. Phil. Q. 396 (2004)CrossRefGoogle Scholar; Juan Comesana, Unsafe Knowledge, 146 Synthese 395 (2005).

98. Even if safety is vague, we can draw distinctions along a spectrum. See, e.g., Greco, supra note 6, at 196 (comparing “weak” and “strong” safety).

99. Epistemic safety has not been explored in detail in the legal literature. Redmayne briefly considers but rejects safety as a legal requirement; however, he employs an extremely strong conception of safety (a belief must be true in every close possible world). See Redmayne, supra note 3, at 301–302. Enoch & Fisher, supra note 12, at 575 n.60, briefly mention but dismiss safety as a “less appropriate” consideration. Pritchard briefly discusses safety in the context of error rates in criminal trials and suggests a possible safety requirement for verdicts. Pritchard, supra note 92, at 452–457.

100. In close possibilities where Blue Bus does cause the accident, the items of evidence would likely produce the same number of correct decisions (100 percent). And in close possibilities where no accident occurs (and thus any lawsuits are false claims), the items of evidence would also likely produce the same number of errors (100 percent).

101. Although I conclude that safety plausibly distinguishes the two examples, it is important to emphasize that I am not arguing that safety explains a general distinction between statistical and non-statistical (or “individualized”) evidence. Therefore, explaining these two examples is less important for my purposes than it is for anyone trying to vindicate such a distinction. For the reasons discussed in Section III, I am skeptical that there is a meaningful epistemic distinction to be drawn. Moreover, as with sensitivity, the safe-unsafe distinction also cuts across the distinction between statistical and individualized evidence (some statistical evidence will be safe and some individualized evidence will be unsafe). The central thesis of this article—that safety affects the quality and hence the probative value of legal evidence—does not necessarily depend on distinguishing these cases. Thus, any readers who reject the above analysis may still accept the general claims about safety and legal evidence.

102. See supra notes 86–88 and accompanying text.

103. Safety also explains Bad Witness, supra note 88. Although the witness's testimony would change in the closest possible world in which the verdict is false (i.e., the cab is blue), there are nine other close possible worlds in which the witness cannot distinguish Mr. Green's cab from the competition.

104. See Fed. R. Evid. 403.

105. See Old Chief v. United States, 519 U.S. 172, 182–183 (1997) (discussing the factors that affect probative value, including (1) the parties’ need for the evidence, (2) possible evidentiary alternatives available to the party, and (3) how the evidence fits with the narratives being presented).

106. This is not to suggest that safety captures all aspects of probative value. Safe evidence may have low probative value for other reasons. For example, when there are not close possibilities in which the findings are false, then evidence of otherwise poor quality will be safe. In addition, safe evidence may have low probative value because it is cumulative of other evidence.

107. Similar considerations apply to Bad Witness, supra note 88.

108. For similar reasons, the probative value in Eyewitness is greater than the probative value in Blue Bus. See supra note 101 and accompanying text. See also Cheng, Edward K., Reconceptualizing the Burden of Proof, 122 Yale L.J. 1254, 1270 (2013)Google Scholar (arguing that the market-share data “borders on irrelevancy”). Safety also fits with a countervailing reason to exclude evidence under Federal Rule of Evidence 403—namely, the tendency of evidence to be “misleading” (i.e., to produce unwarranted inferences). When inferences from evidence are unsafe, the evidence is more likely to mislead the jury.

109. This is not to suggest that safety provides an overarching theory of admissibility. Probative value is one consideration underlying rules of evidence (and probative value includes more than safety). Some rules are justified based on other reasons, including privileges and rules designed to encourage out-of-court conduct. See, e.g., Fed. R. Evid. 407 (subsequent remedial measures), 408 (compromise negotiations), 409 (medical expenses), 410 (pleas), 411 (insurance).

110. Fed. R. Evid. 702–706.

111. Fed. R. Evid. 702. For discussion of reliability in this context, see Haack, supra note 3.

112. The same relationship exists between safety and the factors announced in Daubert to assess expert testimony: (1) whether the basis for the testimony has been tested, (2) whether it has known error rates, (3) whether it has been subjected to peer view and published, (4) whether it possesses standards of control, and (5) whether it is accepted in the relevant community of experts. See Daubert v. Merrell Dow Pharm., Inc., 509 U.S. 579 (1993). These factors each reduce unsafe inferences from expert testimony. As one leading evidence treatise explains, “Daubert expects judges to decide the question whether the theories, techniques, and data as applied can be trusted. . . . what Daubert seeks is the best available assurance of reliability in the sense of accurate and correct outcomes.” Christopher B. Mueller & Laird C. Kirkpatrick, 3 Federal Evidence §7:10 (4th ed. 2016).

113. See, e.g., Fed. R. Evid. 803, 807. The Advisory Committee Notes to Rule 803 explain that the categorical exceptions are justified based on their “circumstantial guarantees of trustworthiness.”

114. See Fed. R. Evid. 902. The Advisory Committee Notes to Rule 902 explain that these categories possess characteristics that “reduce the possibility of unauthenticity to a very small dimension.”

115. Fed. R. Evid. 602 (requiring evidence sufficient to support a finding that witnesses have personal knowledge), 901 (requiring evidence sufficient to support a finding that exhibits are what their proponents claim them to be).

116. Fed. R. Evid. 404. The Advisory Committee Notes to Rule 404 mention that character evidence often has low probative value and high potential for unfair prejudice.

117. In addition, impeachment rules designed to reveal possible defects with testimony allow for evidence indicating reasons why inferences from testimony might be unsafe (i.e., reasons to think the witness is mistaken or lying). See Fed. R. Evid. 607–609, 613.

118. This will also include the probative value of any competing evidence. Standards of proof implement policy goals regarding accuracy and the risk of error. The “preponderance of the evidence” standard is designed to minimize errors and to allocate the risk of error roughly evenly between the parties. See Grogan v. Garner, 498 U.S. 279, 286 (1991). By contrast, the “beyond a reasonable doubt” standard is designed to shift the risk of error away from false convictions. See In re Winship, 397 U.S. 358, 364 (1970).

119. See Anderson v. Griffin, 397 F.3d 515, 521 (7th Cir. 2005) (“[I]f in a particular case all the alternatives are ruled out, we can be confident that the case presents one of those instances in which [a] rare event did occur.”).

120. One would need to know something about other evidence and details about the particular case. It would also be impossible even if we knew, in the abstract, the base rates for true and false confessions. If, for example, we had data indicating that 95 percent are true and 5 percent are false, this would not establish whether this defendant is guilty beyond a reasonable doubt. Moreover, even more extreme numbers in either direction would not change the need to assess the details and other evidence. This is not to suggest that default rules for when evidence is insufficient cannot be established. The common law corpus delecti rule, for example, required additional evidence of a crime before a confession could be sufficient to support a conviction. See Whitman, James Q., Presumption of Innocence or Presumption of Mercy?: Weighing Two Western Modes of Justice, 94 Tex. L. Rev. 933, 951 n.78 (2016)Google Scholar (discussing corroboration requirements for confession evidence). Cf. Roth, Andrea, Safety in Numbers? Deciding When DNA Alone Is Enough to Convict, 85 N.Y.U. L. Rev. 1130, 1184 (2010)Google Scholar (recommending a minimal threshold for the sufficiency of DNA evidence, but also noting that its probative value will depend on other evidence and the competing theories advanced by the parties).

121. Importantly, both safety and probative value are matters of degree. This similarity indicates another important difference between the legal and philosophical contexts. This feature may be a downside in epistemology, see supra notes 97–98, but there is no similar downside in explaining a legal concept that already shares this feature. Indeed, this confluence supports an explanatory fit between the two concepts.

122. See, e.g., Lempert, Richard O., Modeling Relevance, 75 Mich. L. Rev. 1021 (1977)CrossRefGoogle Scholar; Cheng, supra note 108; Pardo & Allen, supra note 3; Allen, supra note 35; Pennington, Nancy & Hastie, Reid, A Cognitive Model of Jury Decision Making: The Story Model, 13 Cardozo L. Rev. 519 (1991)Google Scholar.

123. The same is also true for many projects in philosophy. See Sinnott-Armstrong, Walter, A Contrastivist Manifesto, 22 Soc. Epistemology 257 (2008)CrossRefGoogle Scholar.

124. Evidence that does not by itself distinguish between the parties’ explanations may nevertheless be relevant. Indeed, some trial evidence is “overlapping” in the sense that it is consistent with the claims made by each side.

125. To be clear, I am not arguing that the law imposes—or should impose—a specific safety requirement for verdicts. See Pritchard, supra note 92, at 457 (suggesting the possibility of imposing a safety requirement on verdicts). Safety and probative value are matters of degree. See supra note 121. Thus, rather than trying to draw sharp lines around a safety requirement for law, we can recognize an important (epistemic) role for safety by revealing its role in the probative value of legal evidence. When evidence leads to unsafe inferences, its probative value diminishes accordingly.

126. Edward Craig, Knowledge and the State of Nature (1990), at 91. See also Knowledge Ascriptions (Jessica Brown & Mikkel Gerken eds., 2012).

127. See supra notes 23–24 and accompanying text. Importantly, verdicts will be overturned if they are not supported by sufficient evidence. See supra note 23. Another way of recognizing this similarity is to note, as Robert Brandom has argued, that when one ascribes knowledge, the agent ascribing knowledge is endorsing an inference from (1) commitment to the truth of a proposition to (2) entitlement (to that commitment). See Robert B. Brandom, Making It Explicit: Reasoning, Representing, and Discursive Commitment (1994), at 217. The law does something similar with verdicts. In declaring evidence sufficient to support a verdict, the law is endorsing an inference from a jury's commitment (the verdict) to entitlement (the legal judgment).

128. Consider, for example, witness testimony. The decision to rely on this evidence (inside or outside the law) should depend on how easily the witness could be mistaken or lying (safety).

129. See supra notes 28–29 and accompanying text.