Epistemic sensitivity and evidence

ABSTRACT

In this paper, we put forth an analysis of sensitivity which aims to discern individual from merely statistical evidence. We argue that sensitivity is not to be understood as a factive concept, but as a purely epistemic one. Our resulting analysis of epistemic sensitivity gives rise to an account of legal proof on which a defendant is only found liable based on epistemically sensitive evidence.

KEYWORDS:

• Legal epistemology
• sensitivity
• statistical vs. individual evidence

Acknowledgments

This paper would not exist were it not for Katie Steele who introduced me to its topic and provided insightful and extensive comments. I am grateful for detailed comments by Atoosa Kasirzadeh, Alan Hájek, Pamela Robinson, and the team of editors and referees of Inquiry. Furthermore, I would like to thank the audience at a Humanising Machine Intelligence seminar for their valuable feedback.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 This fictional example and the one to come are based on the examples due to Blome-Tillmann (2015).

2 These authors include Enoch, Spectre, and Fisher (2012, 198), Blome-Tillmann (2015, 103), Gardiner (2018, 181) and Di Bello (2019, 1045–1046, 1080). Ross (2021), by contrast, argues that there are some cases, where fact finders should find liable based on statistical evidence.

3 For example, Cohen (1977), Enoch, Spectre, and Fisher (2012), Enoch and Fisher (2015), Haack (2014), Smith (2010, 2018, 2020), Pritchard (2015, 2018) and Moss (2021).

4 For details, see Stalnaker (2003).

5 We would like to thank an anonymous referee for bringing this possibility to our attention.

6 See Enoch, Spectre, and Fisher (2012, 204) and Enoch and Fisher (2015, 574).

7 See, for example, Pardo (2018, fn. 10).

8 For details on the notion of normic support, normalcy, and explanation, see Smith (2010, 13–19).

9 For a criticism of normic support along similar lines, see Steele (ms).

10 Belief in the first conditional p > e can be seen as an epistemic version of a safety condition: the agent believes that in all ‘nearby’ worlds, where p is true, the evidence e obtains. Our analysis can thus be conceived of as combining a safety and a sensitivity condition.

11 Formally, $\mathrm{\$}$ is a set that contains a smallest element S and finitely many supersets $S^{\prime }$ of S. For details, see Grove (1988, 159).

12 On our model, an agent neither believes p nor $\mathrm{\neg p}$ iff p is true at some of the worlds in S and $\mathrm{\neg p}$ is true at some other worlds in S. As a consequence, our agents believe all logical truths and perhaps all necessary truths. This is, of course, an idealisation. But we think this idealisation is not overly harmful. Nobody will be convicted because a fact finder believes a tautology.

13 Technically speaking, $P^p$ is the general image of P on p. For details, see Lewis (1976, 308–312) and Gärdenfors (1982, 751).

14 The decision-theoretic argument fixes the numerical threshold s such that convicting has higher expected value than acquitting. For details, see for example, Cheng (2013, 1259–1261&1275–1278), Steele (ms, 3–4) and Di Bello (2019, 1054–1055).

15 $\mathrm{\$}=\{S,S^{\prime }\}$.

16 Initially, $w_1$ is assigned a higher probability value than $w_2$. That is, $P(p\mid e)=P\left(w_1\right)=0.7$ and so, by the probability calculus, $P(\mathrm{\neg p}\mid e)=P\left(w_2\right)=0.3$. By supposing p, we consider the general image on p. Since the eye-witness is more reliable than unreliable, $P^p\left(w_1\right)>P^p\left(w_3\right)$. Hence, the unique most likely p-world in $S^p$ according to $P^p$ is $w_1$.

17 The binomial coefficient $(\begin{array}{c}1000\\ 700\end{array})$ equals approximately $5.29\cdot {10}^{263}$.

18 $\mathrm{\$}=\left\{S\right\}$.

19 See Blome-Tillmann (2015, 105) and Di Bello (2019, 1050).

20 It follows that eye-witness testimony need not be individual according to our analysis of epistemic sensitivity. Suppose a witness is a perfect randomiser: she reports p in any hypothetical scenario, where p is the case, in exactly half of the cases, and she reports $\mathrm{\neg p}$ in any hypothetical scenario, where $\mathrm{\neg p}$ is the case, in exactly half of the cases. Then her testimony is epistemically insensitive on our analysis. And indeed, her testimony is random and cannot be used – not even for a ‘stop-and-frisk’ search.

21 Here, we go over concerns of forward- and backward-looking individual evidence. Moreover, I presume that Thomson's account is purely epistemic. The textual evidence is not definite. If her account is factive, wrongful convictions are conceptually impossible, just like for Enoch et al.'s sensitivity account.