Introduction
Suppose that I am a member of the jury in a criminal trial of a defendant who is prosecuted for the offence of making off without payment.Footnote 1 The defendant is accused of leaving a petrol station without paying for 45 litres of petrol. I hear a first-hand witness testimony confirming that the defendant committed the offence at issue, and this testimony appears very reliable. I am explicitly or implicitly conscious of the fact that around 95 per cent of the defendants against whom such a reliable first-hand testimony is available are guilty. I will most probably convict the defendant based on this testimony, and hardly anyone would argue that I should have done otherwise. Considering the unproblematic nature of the generalisation about the guilt of defendants against whom a strong first-hand testimony is available, we can call it ‘an uncontroversial generalisation’.
Now suppose that I am a member of the jury in another criminal trial for making off without payment, and this time I am presented with statistics: out of the 100 customers who visited a kiosk before entering the cinema and left with a bottle of champagne, 95 failed to pay for the bottle, but there is no knowing who; and the defendant has been randomly selected out of the 100. Thus, there is 95% likelihood that the defendant is guilty.Footnote 2 As it raises the likelihood of the defendant's guilt compared with the likelihood of guilt of a random member of the population, the evidence is relevant and admissible. And as it indicates a very high likelihood of guilt, it is hard to see why the evidence should not satisfy the criminal standard of proof. Yet here, it is reasonable to predict that I would acquit the defendant.Footnote 3
There is hardly any specific authority in English law that confirms the status of such statistical generalisations about the guilt of defendants most of whose group-members participated in the crime at issue, and so the question remains to be explicitly decided. Legal experience gained in other jurisdictions confirms that such generalisations might be admissible, but they cannot provide an appropriate basis for conviction. In the recent Israeli Supreme Court decision in the matter of Betzalel,Footnote 4 the appellant had been convicted by the district court of assault occasioning grievous bodily harm together with other members of his organisation (‘La Familia’). After considering the evidence that had been available against the appellant, Justice Stein concluded that most of it was in fact unreliable; hence, the appellant's conviction hung on the fact that almost all the members of his organisation who were present at the scene of the crime participated in the assault. Deciding to acquit the appellant, Justice Stein explained:
The State has proved that there was high probability that the appellant participated in the assault… however, the proof is a vague statistical proof that does not provide sufficient basis for conviction. Since not every La-Familia member who was present in the scene of crime participated in the assault, we ought to allow the benefit of doubt to every member who was not individually identified as one of the participants. The appellant is one such member. The probability of his innocence is low, perhaps very low. But even if we assume that this probability is a negligent one of 0.01 (1%), and that the probability that the appellant participated in the assault thus reaches 0.99 (99%), we would still have to acquit him on the basis that there remains a reasonable doubt about his guilt.Footnote 5 (translation by the author)
Other similar generalisations tend to prove statistical propensity to commit the crime at issue. Such is the generalisation that people of the defendant's class are highly likely to commit crimes of the same description or category as the crime at issue. In this context too, there is no specific authority available about the status of such generalisations. Yet here practice is abundant, and there seem to be no attempts to introduce such generalisations in criminal trials. In addition, only very limited indication is available in the literature that such generalisations might be admissible as propensity evidence under the Criminal Justice Act 2003.Footnote 6 It can, therefore, be concluded that such generalisations are probably inadmissible in English courts.
We can thus call ‘controversial generalisations’ all those relevant generalisations that the law seems to reject either altogether as inadmissible, or merely as potentially sufficient for conviction even when indicating very high probability of guilt.
But why is this so? Why is it that a first-hand witness testimony which is 95% likely to identify a guilty perpetrator can meet the criminal standard of proof, but statistics which are similarly 95% likely to identify a guilty perpetrator cannot meet this standard? Why are we inclined to rely on generalisations about, say, the truthfulness of first-hand witness testimonies when we make legal judgements, but we are reluctant to rely on seemingly identical generalisations that are expressly statistical?Footnote 7 Questions about generalisations are foundational to the law of evidence. Almost every inference from legal evidence to a legal finding involves a generalisation.Footnote 8 One could, therefore, expect that the law of evidence would have a clear theory of generalisations, and particularly of generalisations that lead to criminal convictions. This, however, is hardly the case. The principles guiding the use of generalisations in the criminal justice process are still far from clear.
The literature has explored different possible grounds for a distinction between controversial and uncontroversial generalisations.Footnote 9 As Justice Stein has noted in Betzalel, a significant and lively portion of the literature has focused on the status of generalisations in epistemology, namely on the fact that only uncontroversial generalisations can justify belief in that which the evidence indicates and generate knowledge of it.Footnote 10 Still, the value of knowledge for the legal process has remained uncertain,Footnote 11 and while some have chosen to leave the question open,Footnote 12 others have already suggested that knowledge has no value for the law and that legal practice should be altered: controversial and uncontroversial generalisations should be treated in the same way.Footnote 13
A while ago, I proposed a possible account of the value of knowledge for legal decision-making.Footnote 14 The focus of that account was on substantive responsibility of decision-makers for the content of their decisions; it was suggested that for legal decision-makers to be substantively responsible for the content of their decisions, these decisions must assert facts based on knowledge. I would now like to dig down deeper; I would like to expose the source of any problems with responsibility to which conversational generalisations give rise, and to further uncover the full implications for the justifiability of legal decisions.
To pursue the task effectively, I shall, for now, limit its scope and focus mainly on criminal judgements. My assumption throughout the discussion will be that criminal convictions are justified on deontological grounds; they are not justified as instruments for achieving any goals that are external to defendants. Accordingly, criminal convictions involve moral judgements.Footnote 15 The paper will argue that criminal judgements cannot be based on controversial generalisations, since such generalisations introduce illegitimate arbitrariness into the judgement. Because of this arbitrariness, controversial generalisations cannot justify the moral judgements made in criminal trials; in addition, they undermine the responsibility of the legal decision-maker for the judgement.
To develop this argument, the paper will first consider the propositional structure of moral judgements in criminal law. Next, it will be demonstrated that controversial generalisations cannot provide reasons in defence of all the propositions of a moral judgement; some of the propositions thus remain random. The paper will continue by suggesting that to be justifiable and legitimate, moral judgement cannot incorporate random or indeed arbitrary propositions. Rather, all the propositions contained explicitly or implicitly in a moral judgement should be defensible with reference to reasons. The grounds for the requirement of full reasons will be considered with reference to the objection to moral luck and to its various ethical and institutional implications, and with further reference to requirements of responsibility and legitimacy.
Next, the relation between the proposed analysis and Jonathan Cohen's seminal account in The Probable and the Provable will be assessed.Footnote 16 It will be demonstrated that the proposed analysis captures many of Cohen's insights about controversial generalisations, and that it further provides answers to some questions that Cohen left open.
Before concluding, the paper will suggest that controversial generalisations can still justify some decisions in the criminal justice system, such as non-condemnatory moral decisions or instrumental decisions. It will further be shown that controversial generalisations could have played a more prominent role in the criminal justice system had this system been justified purely on consequentialist grounds and had not sought to morally condemn defendants.
The conclusions will suggest some implications of the proposed analysis for the law of evidence, and they will further extract from the proposed analysis a general method for thinking about the legitimacy of controversial generalisations in different contexts.
1. Moral judgement in criminal law: propositional structure
The focus of this paper is on criminal convictions, even though – as we shall see below – the proposed analysis carries over to criminal acquittals and to other legal judgements of responsibility or its absence. One working assumption is that criminal convictions include, or should include, a moral judgement; and that moral judgement is a cognitive exercise of reason.Footnote 17 This working assumption provides a platform for the current investigation, yet the method of investigation that can be extracted from the current investigation is applicable to non-moral platforms.
It is thus assumed that criminal convictions involve, or should involve, a moral judgement. This judgement is that the defendant is morally responsible and guilty (and she can thus be held legally responsible and guilty). Moral judgement in the sense used here is a subcategory of moral decisions;Footnote 18 the making of moral judgement involves a cognitive exercise of critical and condemnatory moral evaluation.Footnote 19 In analytical terms, this exercise includes: (a) the identification of critical moral norms of responsibility and guilt; and (b) the application of these norms to the facts of specific cases.
Each of the two analytical tasks involved in the making of moral judgement requires the establishment of relevant propositions; and each new proposition is established by way of drawing inferences from previously established propositions. For now, it is enough to note that according to a common and plausible understanding, some of the propositions of a moral judgement are factual and others are normative or moral. For example, we start with one factual proposition, such as: the defendant confessed to deliberately taking the victim's wallet without permission; we infer from it another factual proposition, namely that the defendant deliberately took the victim's wallet without permission; and we then infer from this last factual proposition and from additional moral positions about property, responsibility and guilt that the defendant is responsible for and guilty of committing the wrong of theft.
The proposition that a defendant (D) is morally responsible for and guilty of a specific wrong is the main proposition of the moral judgement incorporated in criminal convictions. Yet the judgement usually contains additional propositions that affirm the main proposition. These affirming propositions negate competing hypotheses: the proposition that D is responsible for and guilty of a specific wrong further contains a proposition negating, say, the hypothesis that D lacks moral agency and hence is not responsible and guilty. Significantly, the moral judgement about D's responsibility and guilt often contains a proposition negating the competing hypothesis that it is another potential defendant A who is responsible for and guilty of that wrong. We can call this proposition the ‘D rather than A’ proposition; it will play an important role the analysis that follows.
We should now direct our attention to the justificatory structure of moral judgement. Accordingly, it is important to identify the exact combination of propositions and inferences that provide reasons capable of justifying moral judgement. This is not the place for a full account of the justification of moral judgement. For current purposes, the relevant question is only this: whether a well-supported main proposition can alone provide justification for moral judgement, or whether well-supported affirming propositions are also required if moral judgement is to be justifiable. This question amounts to the following: whether only the main proposition of a moral judgement needs to be defensible with reasons for the judgement to be justifiable, or whether the affirming propositions also need to be similarly defensible.
The answer to this question will have implications for an account of generalisations. Generalisations are propositions that provide reasons in support of other propositions. For example, the generalisation that most defendants who confess of a wrong have committed that wrong provides support for the proposition that our confessing defendant committed the wrong he has confessed to. Significantly, different types of generalisations can support a different range of propositions. As we shall see further below, some generalisations are only capable of supporting the main proposition of a moral judgement (that D is responsible and guilty), while other generalisations are capable of further supporting the affirming propositions (like the proposition that D rather than A is responsible and guilty). Therefore, if we conclude that all the propositions of a moral judgement should be defensible with reasons, some generalisations will not do.
Accordingly, if a full account of generalisations is to be provided, two questions require consideration: (1) which propositions are supported by different types of generalisations? and (2) must all propositions of a moral judgement be defensible with reasons for this moral judgement to be justifiable? The rest of the paper will take these questions in turn.
2. Generalisations as reasons: full and partial
This section analyses generalisations as reasons that support the propositions of moral judgement. It demonstrates that different generalisations support different ranges of propositions: some generalisations can provide reasons in support of all the relevant propositions of a moral judgement, while other generalisations cannot produce a moral judgement that is fully defensible with reference to reasons. When it comes to the latter, while the generalisation provides reasons to establish that there is an overall high likelihood that the main proposition of the judgement is true, some of the affirming propositions cannot be defended with reference to reasons that demonstrate they are likely to be true.
Let us begin by focusing attention on the place of generalisations in inferences from one fact to another in specific cases. Such inferences start with a contingent fact of the specific case (the tomato at issue is green). This is followed by identification of a generalisation that applies to similar contingent facts (green tomatoes are unripe). An inference of reason is then drawn from the original fact to a new fact based on the generalisation (the tomato at issue is unripe).
The question whether the last inference of reason can be fully supported with reference to the generalisation, or whether it can be only partly supported, depends on the nature of the generalisation. More specifically, it depends on the instances that the generalisation captures. The question is always whether the generalisation essentially captures some instances for which it is not valid; or whether the generalisation does not essentially capture such instances. Where a generalisation essentially captures some instances for which it is not valid, it entails some errors as a matter of logical certainty, and it then cannot fully support inferences between facts. But where a generalisation does not essentially capture instances for which it is not valid, and hence it does not entail errors as a matter of logical certainty, it can fully support inferences from one fact to another.
To see this, we can start with generalisations that do not necessarily capture instances for which they are not valid, and hence they do not entail errors as a matter of logical certainty. Such generalisations would logically allow getting all the relevant cases right, and in some cases they would even logically entail getting all the relevant cases right.
For purposes of convenience we can use here as an example a generalisation that entails getting all cases right (a generalisation that only allows getting all cases right will be discussed in the next section). Assume that we are observing a batch of tomatoes, some of which green and some not-green (or red). Assume further that the fact that a tomato is green means that it is unripe, so that the generalisation ‘green tomatoes are unripe’ is a valid generalisation. Here, the generalisation does not capture any instances where it might not be valid: by definition, there are no green tomatoes that are ripe. Accordingly, we can draw inferences about the unripeness of any green tomato without error (in this example, the inferences are deductive;Footnote 20 but as we shall see further below, this is of limited significance for current purposes).
Where a generalisation does not entail errors as logical certainty, inferences based on it would be fully defensible with reference to reasons: first, the generalisation provides a reason in support of the main proposition (that the green tomato at issue is unripe). And second, there seems to be no relevant competing ‘A rather than D’ hypothesis that is implicitly being negated here, and hence no relevant affirming ‘D rather than A’ proposition that is part of the judgement and might require support. If at all, an affirming proposition would be something like ‘this green tomato, rather than another red tomato, is unripe’. Such an affirming proposition can be supported with reasons: the generalisation provides reasons in support of the proposition that the green tomato is unripe; and at the same time it provides no reason to think an alternative red tomato might be unripe. Hence, to the extent that a relevant affirming ‘D rather than A’ proposition is at all available, the generalisation provides reasons in support of it too.Footnote 21
Clearly enough, a person who draws inferences about the unripeness of a specific green tomato might still be wrong. But even when wrong, this person's inference would not be random or whimsical. The error would be in the application of full and appropriate reasoning to the facts of the specific case: maybe the person is colour-blind, and maybe his impression about correlation between colour and ripeness is misguided. Thus, full (even if contingently erroneous) reasons would be available for that person when trying to defend the proposition that the tomato at issue is unripe.
We can now consider generalisations that essentially capture instances for which they are not valid, and hence they entail errors as a matter of logical certainty. Such generalisations do not logically allow getting all the relevant cases right. An example would be the generalisation ‘green tomatoes are unripe’ where based on evidence that 95 per cent of green tomatoes are unripe. If only 95 per cent of green tomatoes are unripe, a person who observes a sufficiently large batch of tomatoes and tries to determine if specific green tomatoes are unripe will at some point necessarily encounter a green tomato that is ripe, and in this case the generalisation would lead her to an error.Footnote 22
Where the generalisation entails errors as a matter of logical certainty, an inference based on it would not be fully defensible with reference to reasons. In such conditions the main proposition (that the tomato at issue is unripe) can indeed be defended with reference to the generalisation about the unripeness of green tomatoes, which is based on evidence that 95 per cent of the green tomatoes are unripe; but the relevant affirming proposition (here: that another green tomato in the batch rather than the tomato at issue belongs with the 5 per cent ripe tomatoes) is indefensible. In conditions where 5 per cent of the green tomatoes in the batch are ripe, there is no reason to think that another green tomato rather than the green tomato at issue belongs with them. The question which green tomato belongs with the 5 per cent ripe ones is beyond reason; it is merely randomly determined. Accordingly, the inference that a specific green tomato is unripe is not fully defensible with reference to reasons; it contains a random ‘D rather than A’ proposition.Footnote 23
Thus, as far as the question of full defensibility of inferences with reasons is concerned, the relevant distinction is between evidential generalisations that do, and evidential generalisations that do not, entail errors as a matter of logical certainty.
3. Generalisations: controversial and uncontroversial
The distinction between generalisations that do not entail errors as logical certainty and hence can fully support inferences from one fact to another, and generalisations that do entail errors as logical certainty and hence support such inferences only in part, is echoed in the distinction in legal practice between uncontroversial and controversial generalisations.
We can start with generalisations that are uncontroverisally applied in legal practice, such as the ones made about the guilt of defendants against whom a seemingly reliable first-hand witness testimony is available. We shall see below that such a generalisation does not necessarily capture instances for which it is not valid, and hence it does not entail errors as a matter of logical certainty. It is true that, unlike in the equivalent example where green tomatoes are defined as unripe, here the generalisation may well capture instances for which it is not valid. But the important point for current purposes is only this: that the logical structure of the generalisation does not dictate that it must capture such instances; and accordingly, the generalisation does not logically entail errors. Let us begin by clarifying why this is the case.
In the first-hand witness testimony example, the relevant generalisation is ‘defendants against whom first-hand witness testimonies with XXX marks of truthfulness and hardly any marks of untruthfulness are available are guilty’.Footnote 24 In fact, there are two relevant generalisations here: (1) ‘first-hand witness testimonies with XXX marks of truthfulness and hardly any marks of untruthfulness are reliable’; and (2) ‘defendants against whom reliable first-hand witness testimonies are available are guilty’.
The second generalisation is a relatively easy case. This generalisation provides reasons in defence of the main proposition ‘D is guilty’; and it further provides reasons in defence of an affirming proposition ‘D rather than A is guilty’, where A is another potential defendant against whom no reliable testimony is available: since no reliable testimony is available against A, A's case is not captured by the generalisation, and hence there are no reasons available in support of a conclusion about A's guilt, while at the same time there are reasons available in support of a conclusion about D's guilt. Once again, this does not mean that errors are impossible,Footnote 25 but it does mean that reasons are available in support of all the proposition of the judgement.
But what about the first generalisation that establishes the reliability of the testimony? Does it not rely on the assumption that most (rather than all) first-hand witness testimonies with XXX marks of truthfulness are reliable, thus capturing also instances of similar yet unreliable testimonies (in which case there would be no reason for assuming that the testimony at issue, rather than any other similar testimony, is reliable)?
In a generalisation like the one about the reliability of a testimony with XXX marks of truthfulness, ‘XXX’ is an open concept receptive of a virtually infinite range of values (there can be ever more marks of truthfulness to our first-hand witness testimony that would distinguish it from other first-hand witness testimonies, even if we cannot immediately put our finger on them). Accordingly, the instances that the generalisation captures are open to further and further individuation. This is one kind of a generalisation that does not necessarily capture instances for which it is not valid;Footnote 26 logically, the instances that the generalisation captures might include no unreliable testimonies at all. When relying on such a generalisation, a hypothetical sufficiently skilled decision-maker who is capable of fully unlocking the potential of testimonies by identifying all their marks of truthfulness would never get it wrong. That such a decision-maker is merely hypothetical does not matter for current purposes; it is the logics of the generalisation (and the subsequent defensibility of the inference with reasons) that matters. Accordingly, such a generalisation does not entail errors as a matter of logical certainty.
The same analysis would apply to other uncontroversial generalisations, such as the generalisation that motives with characteristics XXX have the power to move agents to act. If we are to abstract, then the same analysis would apply to any generalisation tracing a connection between the defendant and a specific crime that may well be stronger than the connections between any other potential defendant and this crime. Wherever the generalisation traces such a connection, it does not necessarily capture instances for which it is not valid.
As we have already seen, where a generalisation does not necessarily capture instances for which it is not valid, and hence it does not entail errors as a matter of logical certainty, the decision-maker's inference would be fully defensible with reference to reasons. We can return to the case of first-hand witness testimonies to demonstrate this. In such a case, the generalisation supports the main proposition (that this testimony is reliable); and there seems to be no relevant ‘D rather than A’ affirming proposition that requires support. If at all, an affirming proposition would be ‘this testimony, rather than another testimony that lacks similar marks of truthfulness, belongs with the reliable testimonies’. Such an affirming proposition is defensible with reasons: the generalisation provides reasons to think that this testimony is reliable; and at the same time, it provides no reasons to think that the other testimony that that does not have XXX signs of truthfulness is reliable. Here too, the decision-maker might still be wrong; but her inference would not be random or whimsical. Rather, the error would then be merely in the application of full reasons to the facts of the specific case (for example, there may have been only XX marks of truthfulness in the testimony). Thus, while not error-proof, judgement based on such uncontroversial generalisations is fully defensible with reference to reasons.
We can now move on to controversial generalisations, such as most statistical generalisations. The same analysis does not apply to controversial generalisations. Here, the generalisation essentially captures some instances for which it is not valid, and accordingly it logically entails errors and logically forbids getting all the relevant cases right. In such conditions, the main proposition is defensible with reasons, but the affirming proposition is not. Where the generalisation ‘tall customers made off without payment’ is based on evidence that 95 out of 100 tall customers made off without payment, the affirming proposition that it was D rather than another random tall customer A who made off without payment is not defensible with reference to reasons. As the generalisation captures guilty and innocent customers, it applies with equal force to D and to all other tall customers including A. Thus, none of them is more likely than the others to be guilty. The answer to the question who out of all potential defendants (all of whom are captured by the generalisation) is guilty is thus randomly determined; it is beyond human reason. Accordingly, judgement based on such controversial generalisations cannot be fully defended with reference to reasons.
4. Moral judgement: justifiability, responsibility, legitimacy
For a moral judgement of responsibility and guilt to be justifiable and legitimate, all the propositions contained in it should be defensible with reference to reasons. These consist of the main proposition and its affirming propositions, including the ‘D rather than A’ proposition (the proposition that it is D rather than A who is responsible and guilty). That D is most probably morally responsible and guilty cannot yet justify judging him morally responsible and guilty;Footnote 27 further, it cannot provide basis for a judging agent's responsibility for such a judgement, and it cannot legitimise such a judgement. Justifiability, responsibility and legitimacy require the availability of further reasons that indicate that D is more likely to be responsible and guilty than any other potential actor A. This is because to be justifiable and legitimate, a moral judgement of responsibility and guilt should be free from random or arbitrary propositions; and only where such further reasons are available is the moral judgement against D free from random or arbitrary propositions. Accordingly, controversial generalisations that support only the main proposition of a moral judgement of responsibility and guilt but not its affirming propositions cannot provide basis for such a judgement.
We can start by considering the justifiability of moral judgement. The objection to randomness or arbitrariness in the reasoning underlying moral judgement seems to have intuitive power. It is easy to accept that when I judge that D is morally guilty and hence implicitly that A is morally innocent, I should be able to provide reasons in defence of my propositions about both D and A. Saying ‘well, it just so randomly happened that I had to make a judgement about D and not about A, and this is what led to my judgement against D and implicitly in favour of A’ seems straightforwardly objectionable.
But beyond its intuitive power, the objection to randomness in moral judgement is also founded on well-established principles. To start unfolding these, let us observe the content and nature of moral judgement. Moral judgement is the exercise of criticism over D's engagement with moral reasons, and accordingly it is justified on the basis that D could have avoided this judgement through proper engagement with reasons. Such engagement with reasons would have allowed D to do the following: (1) to refrain from or stop pursuing the action or attitude that provokes criticism; or (2) to act and then refute the criticism: D could engage with the reasons underlying the criticism and demonstrate that they are not valid. Where moral judgement involves random propositions, D's ability to avoid moral criticism in either of these ways is impaired, and accordingly the judgement is unjustifiable.
We can begin with D's ability to refrain from or stop the action or attitude that provokes criticism (as in (1)). Where moral judgement includes random propositions, this ability is impaired in a way that, first, gives rise to an objection from moral luck.Footnote 28 Consider the case where the ‘D rather than A’ proposition is random, namely the available evidence suggests that the likelihood that D is guilty is equal to the likelihood that A is guilty. In the absence of reasons supporting the proposition that D rather than A is guilty, the moral criticism directed at D hangs only on her bad luck. The difference between D and A is purely a matter of chance and accordingly beyond their control: D is merely situated so that her actions or attitudes happen to be within the reach of the judging agent's assessment, and A is situated so that his actions or attitudes happen to be outside the reach of the judging agent's assessment. In such conditions, D cannot escape moral criticism by refraining from the action that calls for criticism; she would be similarly situated within the reach of the judging agent's assessment whether she acts (and hence belongs with, say, the 95 per cent who acted) or not (and hence belongs with the 5 per cent who did not). And the objection to moral luck tells us that the chance event that situated D and A differently lacks moral significance. If the difference between A and D is due to merely lucky factors beyond their control, neither of the two is more deserving of moral condemnation or praise than the other.Footnote 29 Accordingly, allowing D's bad luck to impact moral judgement (so that D rather than A is judged morally responsible and guilty) is objectionable.Footnote 30
Arguments about moral luck usually go to the consequences, causes or circumstances of the assessed agent's actions rather than to the circumstances of their moral assessment.Footnote 31 Arguably, these are the only appropriate spheres for the objection to moral luck: luck in the circumstances of the assessment is not similarly objectionable. First, moral assessment is always conducted in imperfect conditions, and hence luck in its application is practically irreducible. For example, the existence of a first-hand witness who is available to testify is itself a matter of luck. Objecting to luck in the circumstances of the assessment would thus prohibit moral judgement altogether. Secondly, it is arguable that a moral assessment is merely that (an assessment); but responsibility and guilt exist independently of their assessment, and only responsibility and guilt resist luck convincingly.
We can start with the second objection; its resolution will also resolve the first. It is hardly the case that moral responsibility and guilt exist independently of moral assessment or judgement. Rather, moral judgement creates responsibility and guilt. When we speak of a person being responsible and guilty, we speak of nothing but the availability of reasons whose accumulation establishes, as a matter of reason, a justified judgement of responsibility and guilt.Footnote 32 Any concrete judgement by a specific judging agent can of course be right or wrong, depending on the validity of its underlying reasons; but this does not change the fact that responsibility and guilt are created by the reasoning of moral judgement.
If this is accepted, then arguments about moral luck are in fact arguments about valid reasoning in moral judgement. The objection to moral luck is then an objection to the incorporation of luck as a valid reason in moral judgement. Thus, a distinction should be drawn between a situation where the judging agent happens to be confronted with D only against whom evidence happens to be available, and a situation where the judging agent happens to be confronted with D and implicitly also with A, against both of whom the same evidence happens to be available. In the first situation, the judging agent need not refer to luck as a ‘reason’ for the judgement: it is enough to refer to the available evidence and to the reasons it provides for judging only D guilty. But in the second situation, the judging agent has no choice but to refer to luck as one of the ‘reasons’ for the judgement. Since the same evidence is available both against D and against A, the ‘reason’ for attributing guilt to D and innocence to A is in these circumstances nothing other than D's bad luck. Thus, where understood as an objection to the incorporation of luck as a reason in moral judgement, the objection to moral luck seems to carry over only to the second situation. And this situation, for which the objection to moral lack is applicable, is the situation that controversial generalisations confront us with.
In addition to the objection from moral luck, D's inability to avoid criticism by refraining from action undermines the justifiability of the judgement in another way: it robs D of the ability to escape a fate of condemnation in a way that is profoundly tragic and (relatedly) also wrongful.Footnote 33 Randomness like the one of the affirming ‘D rather than A’ proposition is unique: it usually manifests the existence of contingent forces that push both D and A towards condemnable action (for example, because both D and A are situated so that they have similar reasons to perform condemnable action). On the background of these forces, both D and A seem fated to condemnation. Errors are instances where D has heroically manged to overcome these forces and avoid condemnable action; yet almost astonishingly, she still ends up being condemned because of her bad luck that distinguishes her from A. The incorporation of luck in the judgement thus means that D could not escape her bad fate despite every heroic effort. This outcome is profoundly tragic in an Aristotelian sense,Footnote 34 and allowing it to unfold is arguably wrongful: where possible, we should allow people the freedom to escape any ‘fate’ assigned to them by their contingent circumstances. These tragic and wrongful implications for D further undermine the justifiability of the judgement.
Notably, in the legal context, where moral judgement is public and institutionalised, the tragic and wrongful nature of unlucky errors is most acute. In such a public and institutionalised context, D has no real chance of escaping the social outreach of the judgement or indeed her fate; she cannot hope to find a social sphere where condemnation will not hang over her head. Furthermore, to the extent that the legal system has a communicative justification, this wide social outreach of the condemnation is not just incidental to the system; rather, it is its most essential characteristic. The tragic nature of error is then embedded in the justification of the judgement as public. The case against randomness in moral judgement is thus strongest here.
A third way in which arbitrary affirming propositions undermine the justifiability of moral judgement concerns D's ability to act and then refute the criticism through engagement with reasons. By allowing criticism that is not fully based on reasons but rather incorporates some arbitrary propositions, we deny or at least diminish D's ability to avoid the criticism through engagement with reasons. Thus, we pull the carpet from underneath the criticism: it is hard to criticise the defendant for not engaging well with reasons, where some relevant reasons supporting this criticism are unavailable. Furthermore, a critical moral judgement that contains random propositions and hence cannot be fully defended with reference to reasons is plainly unfair to the person whose action is being judged. It denies him the opportunity to defend himself and refute the criticism with reference to countering reasons.Footnote 35 In addition, such a judgement fails to respect him as a moral agent who can engage with it and respond to it rather than merely passively accepting it. For these reasons too, arbitrariness undermines the justifiability of moral judgement.
This last foundation for a requirement of non-arbitrariness also gains its full force where the moral judgement is public and institutionalised. In such contexts, D's robust ability to refute the reasons underlying the judgement further ensures the ability to effectively scrutinise the judgement. The ability to scrutinise the judgement contributes, in turn, to its justifiability as a public judgement, and it also works to legitimise the judgement (as discussed further below). This point goes to the legal meaning and significance of non-arbitrariness.Footnote 36 Furthermore, D's robust ability to refute the reasons underlying the judgement contributes to its effective communication to D and to the public. Thus, to the extent that a legal system is based on communicative assumptions, non-arbitrariness ensures justifiability of the judgement as a public judgement in this way too, and arbitrariness thus undermines it.
We can now move from justifiability to responsibility and legitimacy. Random or arbitrary affirming propositions undermine the responsibility of the judging agent for the judgement. Like moral action, moral judgement presumes responsibility. This is not the place for an exhaustive exploration of responsibility;Footnote 37 instead, the following can be briefly noted. Responsibility captures a link between an agent and an object, and it traces legitimate expectations from the agent that concern her reasoning about the object.Footnote 38 Responsibility thus creates a route for dealing with eventual errors in judgement about the object with reference to reasons. In providing this route, responsibility has independent moral role and possibly also independent value: it allows the judging agent and the relevant public to contain errors as active subjects who can learn and improve; it thus prevents helplessness in the face of errors and subsequent indifference to them. These independent role and value of responsibility further feed into two normative foundations of judgement: first, they contribute to the justifiability of the judgement; and it is arguable that they are preconditions for a judgement's justifiability.Footnote 39 Secondly, responsibility legitimises the judgement.Footnote 40 This is the case both if ‘legitimacy’ receives an empirical meaning referring to approval by the relevant population, and if it receives a normative meaning referring to various standards that should be met if a judgement or an institution is to be legitimate.Footnote 41
Yet where a judging agent incorporates random or arbitrary affirming propositions in her judgement, her responsibility for the judgement is undermined for at least three related reasons. First, where an agent cannot identify reasons underlying a ‘D rather than A’ proposition, she cannot constructively engage with an eventual ‘A rather than D’ error with reference to such reasons and hence as an active subject who is be able to learn from error and to improve. Secondly, we saw that the incorporation of arbitrary affirming propositions makes the judgement unjustifiable for various other reasons. Where moral judgement is already unjustifiable, a judging agent who is required to pursue it despite its unjustifiability cannot be held responsible for errors, as the necessary link between the judging agent and the error is then missing. Thirdly, the incorporation of arbitrary affirming propositions also undermines the ability of the agent to trust her findings;Footnote 42 and I have suggested elsewhere that this prevents the judging agent from asserting the facts that form the basis of her judgement and, relatedly, it further undermines her substantive responsibility for the judgement.Footnote 43 For all these reasons, responsibility is undermined, and hence also the justifiability of the moral judgement and its legitimacy.
The legal context exacerbates the difficulty here too. In the legal context, problems with the judging agent's responsibility call into the picture institutional considerations; specifically, they might undermine the legitimacy not only of specific judgements but also of the legal institutions that make these judgements.Footnote 44 For this reason too, the case against randomness or arbitrariness in moral judgement is strongest in institutional contexts like the legal one. Moreover, undermined legitimacy essentially translates to a communicative failure. To the extent that legal judgements are justified on communicative grounds, undermined legitimacy thus further undermines the justifiability of the judgement as a public judgement.
It can thus be concluded that arbitrary affirming propositions undermine the justifiability of moral judgements of responsibility and guilt; the responsibility of the judging agent for such judgements; and the legitimacy of the judgements and possibly also of the institutions that make them.
Before moving on, the scope of the argument about non-arbitrariness can be briefly and cautiously considered. Until now, the argument has focused on condemnatory moral judgements of responsibility and guilt. However, other moral judgements have a similar structure. These include judgements of innocence and judgements about mere moral responsibility or its absence (where moral responsibility is a component of legal liability). Rather than a full-blown condemnatory attitude, the critical moral stance might then take the form of moral praise, or it might take more subtle forms of critique (where the evaluation is negative) or of approval (where it is positive).
Much of the proposed analysis carries over more broadly to all these moral judgements that feature in the legal system. Moral luck is similarly objectionable in moral judgements of innocence and, some have suggested, in moral judgements regarding responsibility or its absence.Footnote 45 Furthermore, the judging agent's responsibility is similarly required and similarly undermined when incorporating arbitrariness in judgements about innocence or about responsibility. Hence, it can be cautiously suggested that the proposed analysis applies with similar force to the use of controversial generalisations in raising reasonable doubt about prosecution evidence indicating guilt in criminal trials,Footnote 46 and to the use of controversial generalisations in establishing moral responsibility or its absence in other legal contexts that incorporate moral judgement.
5. Moral judgement and forms of inference
We can now consider Cohen's analysis of inferential processes in The Probable and the Provable Footnote 47 and observe how it stands in relation to the proposed analysis. Cohen has put the spotlight on the logical form of the inference (that can include mainly deduction or one of several forms of inductionFootnote 48). He focused on Bacon's ground-breaking account of eliminative induction that is commonly used in the experimental sciences. In eliminative induction we seek to establish generalised hypotheses through systematic observation of specific instances whereby these hypotheses might be negated. This process is in fact one of negating competing hypotheses. We infer that bees discriminate between different colours by negating the possibility that it is odour or another factor or the combination thereof that gives them clues about the location of their feeding-place.Footnote 49 In Cohen's words, in eliminative induction we work to ‘eliminate alternative, antecedently not unlikely, explanations of the phenomenon’.Footnote 50 Thus, in fact, in eliminative induction we infer the main proposition through validation of its affirming propositions. Cohen then relied on this method of inference (eliminative induction) to develop a method of quantification that he termed ‘inductive probability’.Footnote 51
Part of Cohen's argument was that eliminative induction and inductive probability worked best to explain legal practice in a range of areas, including the legal aversion from controversial generalisations. He suggested that legal propositions of responsibility and guilt are often inferred in the same way as the general hypotheses of the experimental sciences, namely through eliminative induction (negation of competing hypotheses). Accordingly, relevant probability estimations would be the ones of inductive probability. And in inductive probability, evidence that establishes a controversial generalisation lacks any weight, as it does not serve to negate a competing hypothesis that can explain the phenomenon at issue. In other words, controversial generalisations are not probative because they do not support affirming propositions.
Cohen did not explain why the use of eliminative induction and inductive probability might be the most appropriate in some contexts but not in others. However, by now we can understand why this might be the case. Because in eliminative induction propositions are established based on negation of competing hypotheses (or based on proof of their affirming propositions), it is a method of inference where affirming propositions do not remain indefensible. What previous sections have suggested is that this is the crucial point for the justifiability and legitimacy of moral judgement: because of the nature of moral judgement, inferences that leave affirming propositions indefensible cannot justify or legitimise it.
Clearly, this leaves room for the use of different forms of inference when making moral judgement in law, as long as these forms of inference are not applied in a way that leaves the affirming propositions indefensible. Deductive reasoning and eliminative induction would always be appropriate; but other forms of enumerative and non-enumerative reasoning might also be appropriate, with the constraint that they should not rely on controversial generalisations where they seek to provide basis for moral judgement of responsibility and guilt in a specific case. This conclusion fits the flexible way in which forms of inference are used in legal discourse.
6. Beyond judgements of responsibility and guilt
Controversial generalisations thus cannot provide a foundation for moral judgements of responsibility and guilt in specific cases. This, however, does not imply that such generalisations cannot provide a foundation for other decisions that sometimes feature in the criminal justice system, such as decisions about the rationality, rightfulness or instrumentality of courses of action.
We can start with decisions about the rationality and rightfulness of courses of action.Footnote 52 Here, controversial generalisations often provide appropriate justifying reasons. In such decisions, the generalisations usually support propositions about risk expectancy; and once the risk expectancy is determined, the right course of action is identified. Decisions about risk expectancy and the right course of action do not include judgement about any specific potential bearer of risk D (even though they might affect D). Thus, it is not necessary that a ‘D rather than A’ proposition be defensible with reference to reasons. Rather, an overall high likelihood of truth of the main proposition about the course of action would provide enough grounds for ensuring non-arbitrariness.
This can be clarified with reference to potentially blameworthy decisions about courses of action made by medical doctors. Consider, for example, a doctor's decision to refrain from treating a patient suffering from mild illness with the only medicine that can cure him because 95 per cent of people of this patient's race are severely allergic to the medicine. Such a decision concerns only the rationality and rightfulness of courses of action that might affect the patient; it does not involve any judgement of the patient. The proposition that this patient rather than another potential patient does not belong with the 5 per cent tolerant patients is not part of the decision, and accordingly its randomness is unproblematic. There is no room for a further challenge directed at the doctor: ‘but what made you think this patient does not belong with the 5 per cent who would tolerate the medicine?’ Thus, the decision to refrain from treatment does not include any random or arbitrary propositions; and the otherwise-controversial generalisation can justify it. It is true that in the context of a criminal trial, the next stage of evaluation might involve a moral judgement of the doctor's responsibility and guilt based on the identified right course of action. But only this last judgement, and not the decision about the right course of action, cannot be made based on controversial generalisations.
Furthermore, controversial generalisations can justify judgements about the instrumentality of courses of action. Instrumental decision-making treats specific cases as means for achieving overall desirable consequences. In the context of systems like the criminal justice system, instrumental decision-making often requires only the availability of reasons that can defend the work of a rule of decision across the entire system. And controversial generalisations are fit for this purpose, as they work to increase accuracy across the system. Consider, for example, the decision that 9-year-old children cannot be criminally liable. This decision is based on considerations of procedural efficiency across the system. Accordingly, a generalisation based on evidence that X per cent of children under a certain relevant age have limited cognitive capacities can justify such a decision.
7. Generalisations and the justification for punishment
The analysis proposed so far implies that the appropriateness of relying on controversial generalisations as bases for judgements of responsibility and guilt depends on the justification for criminal punishment in this system. Had the imposition of criminal punishment not been based on moral judgements of responsibility and guilt, controversial generalisations could have provided justification for all the judgements made across the criminal justice system.
Consider, for example, a decision to convict D based on a controversial generalisation in a criminal justice system that aims only at optimising general deterrence. In such a system, the only proposition requiring support is that convicting D would work to optimise general deterrence across the system. This proposition can be defended with reference to the fact that convicting D based on such generalisations would increase accuracy across the system, and hence it would increase overall general deterrence.Footnote 53 Controversial generalisations would thus provide reasons that are capable of justifying the decision.
However, if it is accepted that the system is justified on moral judgements of responsibility and guilt, controversial generalisations cannot do. Had they allowed moral judgement, controversial generalisations would indeed have improved the quality of judgement in each specific case (and incidentally also across the system). But since they do not allow for justifiable and legitimate moral judgement, it becomes impossible to achieve maximal accuracy and to maintain the moral nature of the judgement at the same time.
Conclusions
This paper has suggested that controversial generalisations cannot provide reasons in defence of all the propositions of a moral judgement of the type included in criminal convictions. Rather, controversial generalisations leave some of the propositions of the judgement random or arbitrary. It has been further suggested that where it contains arbitrary propositions, such a moral judgement is unjustifiable and illegitimate. Accordingly, controversial generalisations cannot provide a foundation for the moral judgement that a criminal conviction involves.
This conclusion has implications for the law of evidence. As we have seen, English law does not offer specific authority on the status of controversial generalisations. As far as propensity evidence is concerned, only minimal indication is available in the literature that the Criminal Justice Act 2003 might open the door for controversial generalisations, and thus the question might be mainly theoretical. However, controversial generalisations like the generalisation that most members of the defendant's group participated in the commission of the crime at issue might still be admitted, as happened elsewhere; and accordingly there might be a need to decide their status.
The paper has suggested that reliance on any controversial generalisations would deem legal judgement unjustifiable and illegitimate. Accordingly, the least that can be proposed is that the decision whether to convict based on such generalisations should not be left for the jury. Juries should be instructed that they cannot base their judgement on such generalisations. Such generalisations can only work to strengthen a judgement that has an independent evidential foundation and that would have been made even without the controversial generalisation. Yet this approach is not without difficulty. First, the added value of the evidence that supports a controversial generalisation would be low (as the same judgement would have been made even without the evidence). Secondly, it would be hard for the juries and for the public to identify whether the same judgement would indeed have been made even without the evidence. Thirdly and relatedly, since some evidence of a statistical nature might stimulate prejudice, the risk that juries would err in identifying whether they would have made the same judgement absent the evidence is high; such error would lead to unjustifiable and illegitimate convictions and would further reinforce prejudice. Since the added value of the evidence is low and the risk of unjustifiable and illegitimate decisions is high, a better approach would be to exclude evidence supporting controversial generalisations altogether.
But this is not where the story comes to an end. The story is not (only) about the use of controversial generalisations in moral judgements of guilt and of responsibility, as these feature in the legal system. It is a more general story about controversial generalisations and their legitimate use in a far wider range of contexts: can we make any value judgement (ethical, epistemological, aesthetic, or other) based on controversial generalisations? Can profiling be justified? Can we legitimately use statistics in distributional decisions? In developing personal attitudes towards specific persons?Footnote 54 Full examination of the use of controversial generalisations in different contexts is a pending task.
Viewed in this more general context, the contribution of this paper is in providing a method for thinking about controversial generalisations. The paper suggests that to decide whether a judgement can be based on controversial generalisations, we must identify the propositions of this judgement that should be defensible with reasons before the judgement is justifiable and legitimate.
