2. A primer on probability spaces

The project of moral uncertaintism only gets off the ground if individual beliefs can be represented by a probability distribution of some kind. So what is a probability distribution? Simply put, it’s a numerical representation of uncertainty about what will happen. Let $ \Omega $ be some arbitrary nonempty set of elementary outcomes; in a simple model of a dice toss, the set of outcomes might be 1,2,3,4,5, and 6. Next, define a family $ \mathrm{\mathcal{F}} $ of subsets of $ \Omega $ such that $ \mathrm{\mathcal{F}} $ is closed under complement, countable union, and countable intersection: if $ A,\hskip1.5pt B $ are both elements of $ \mathrm{\mathcal{F}} $ , then $ A\hskip1.5pt \cup \hskip1.5pt B $ , $ A\hskip1.5pt \cap \hskip1.5pt B $ , and $ B\backslash A $ are as well. $ \mathrm{\mathcal{F}} $ is what is called a $ \sigma $ -algebra over $ \Omega . $ $ \mathrm{\mathcal{F}} $ is the set of possible events, i.e., propositions that might be true or false and that you could bet on; in our dice-toss model, the events will include “1,” “2,” etc., as well as “even,”, “odd,” “less than six,” etc. Finally define a function $ p:\mathrm{\mathcal{F}}\to \left[0,1\right]. $ $ p $ is a probability distribution if and only if it satisfies three conditions:

1. 1. (Nonnegativity) $ p(E)\ge 0\hskip1.5pt \forall \hskip1.5pt E\in \mathrm{\mathcal{F}} $ .

2. 2. (Unitarity) $ p\left(\Omega \right)=1. $

3. 3. (Countable additivity) for any countable sequence $ {\left\{{A}_n\right\}}_{n=1}^{\infty } $ of pairwise disjoint sets in $ \mathrm{\mathcal{F}} $ , $ p\left({\cup}_{n=1}^{\infty }{A}_n\right)={\sum}_{n=1}^{\infty }p\left({A}_n\right). $

The triple $ \left\langle \Omega, \mathrm{\mathcal{F}},p\right\rangle $ is called a probability space; $ \Omega $ and $ \mathrm{\mathcal{F}} $ are called, respectively, the sample space and the event space. I shall interpret probabilities subjectively, as representing degrees of belief. Thus, the elements of $ \mathrm{\mathcal{F}} $ are our objects of belief. In our dice-toss model, if you think the dice is fair, then $ p\left(``\mathrm{even}"\right)=p\left(``\mathrm{odd}"\right). $

If we want to use this sort of framework to represent individual beliefs about moral theories, we have no choice: either moral theories will have to be represented as events (i.e., elements of $ \mathrm{\mathcal{F}} $ ), or as sampling outcomes (i.e., elements of $ \Omega $ ). This is not an indifferent modelling choice: each entails a very different way of thinking about how moral theories are related to each other, and how our beliefs about them are structured. Under the first modelling choice, the objects of our beliefs are moral theories themselves and, formally, they are subsets of some common underlying sample space. Much more importantly, it means that the set of moral theories is a $ \sigma $ -algebra, and it therefore comes with nonnegotiable structure and content. Moral theories must bear set-theoretic relations to each other (e.g., some moral theories will be subsets of others), and all sets which can be obtained by countable intersection or union or complement from other moral theories themselves feature among the set of moral theories—and we must have beliefs about the moral theories thus obtained. Notice that under this first modelling choice multiple moral theories may be true: if $ A $ is a subset of $ B $ then $ p(A)\le p(B) $ , and thus the truth of any moral theory implies the truth of any other theory of which it is a subset. Furthermore, if moral theories are events, then a number of principles of rationality can kick in which regulate how we must distribute our beliefs among them. For example, so-called regularity principles will require us to assign a nonzero credence to every nonempty event—i.e., every moral theory that isn’t the empty set.

Under the second modelling choice, the objects of our beliefs are not moral theories themselves, but rather sets of moral theories, interpreted as events in which one of the moral theories in the set is true. So for example, the event $ \left\{{T}_1,{T}_2\right\} $ is the event in which either theory 1 or theory 2 is true, and the “atomic” event $ \left\{{T}_1\right\} $ is the event in which $ {T}_1 $ is true. On this modelling approach, moral theories need bear no set-theoretic relations to one another, but they are all mutually exclusive. By definition, $ p(A)=1 $ implies $ p\left(\Omega \backslash A\right)=0 $ . Thus, if moral theories are the elementary outcomes of our sample space, then no two theories can both be true, as all atomic events are pairwise disjoint. But most significantly, because the sample space is not a $ \sigma $ -algebra, the space of moral theories does not come with nonnegotiable content and structure: we can include what we like in our sample space, and neither probability theory nor EU theory can force us to include any particular theory in the sample space. By analogy, in our dice-toss example, the only outcomes I included were 1,2,3,4,5, and 6—I didn’t also include the possible outcome “the dice vanishes in a puff of smoke”; I certainly could have included this possible outcome, but the laws of probability did not require it. This may turn out to be a blessing and a curse, as we will see.

The way in which the moral uncertainty framework is usually deployed suggests that moral theories are being thought of as events. The table in the introduction, which is typical of the literature, suggests that theories are supposed to be our objects of belief, and therefore our events. And it is widely assumed that theories can be fine or coarse grained arbitrarily: MacAskill and Ord (Reference MacAskill and Ord2018), for example, claim that we can always equivalently represent your credences between theories as being split 0.4-0.6 between deontology and utilitarianism, or instead as being split 0.4-0.3-0.3 between deontology and two slightly different, more specific versions of utilitarianism that disagree with one another but which both imply the more general version. This assumption provides the basis for an important argument against the ‘My Favourite Theory’ view (MFT), which holds that what you ought to do is just what “the” theory you find most plausible: no such thing, they claim, because the theory that gets the highest credence isn’t the same depending on your choice of partition. Whether or not this argument is persuasive, it presupposes a model of theories as events, since some theories can imply others.

This immediately raises the question: What kinds of events?

3. Theories as events

I wish to consider three different ways of modelling theories as events: as sets of sentences, as preorders, and as choice functions. All three fail for a general reason, which is that on this approach, moral theories are “opinionated events,” and the set-theoretic relations of different events force pathological combinations of opinions. Accordingly, the reader is invited to read just one of the following three subsections, then skip to section 4. For although the impossibility results shown throughout these subsections vary in detail, their inner logic is much the same.

3.c Theories = sets of choice functions

So our second attempt was also a flop, but let’s not give up yet. One last try. Suppose we try to define moral theories as choice functions instead of as preorders. If $ A $ is our set of possible acts, a choice function is a function which selects from every nonempty subset $ B $ of $ A $ a nonempty subset of $ B. $ Intuitively, a choice function selects from a menu of alternatives a submenu of alternatives, each of which is permissibly chosen. Formally, a choice function $ c $ is a subset of the Cartesian product $ {\mathcal{P}}_0(A)\times {\mathcal{P}}_0(A), $ where $ {\mathcal{P}}_0(A)=\mathcal{P}(A)\backslash \varnothing, $ with the property that for any two nonempty $ B,C\in A,\left(B,C\right)\in c\hskip0.35em \Longrightarrow C\hskip0.35em \subseteq \hskip1.5pt B $ .

Let us now define the probability space $ \left({\mathcal{P}}_0(A)\times {\mathcal{P}}_0(A),\varepsilon, p\right) $ , where Ɛ is any $ \sigma $ -algebra over $ {\mathcal{P}}_0(A)\times {\mathcal{P}}_0(A) $ that includes all possible choice functions over $ A $ . Unfortunately, since Ɛ is a $ \sigma $ -algebra, it cannot include only choice functions, it must also include as elements the complements of choice functions, and all countable unions and countable intersections of choice functions and their complements. And these other elements are not themselves choice functions. We need some way of associating choice-worthiness judgments with every element in Ɛ. One way of doing this is to associate with every element in Ɛ a “pseudo” choice function that selects from every set of alternatives a subset of that set whose elements are permissibly chosen—but unlike with ordinary choice functions, this subset may be empty. If the pseudo choice function maps a set of alternatives to the empty set this means that nothing is permissibly chosen, as might be the case if we face a true moral dilemma. So for every $ R\in \varepsilon, $ define the pseudo choice function $ {C}_R $ as:

$$ {C}_R=\left\{\left(B,C\right)\in \mathcal{P}(A)\times \mathcal{P}(A)|C\hskip2pt \subseteq \hskip2pt B\;\hskip2pt \mathrm{and}\hskip2pt \forall x\in C,x\in B\hskip2pt \iff \hskip2pt \exists D\in \mathcal{P}(A):\left(C,D\right)\in R\hskip2pt \mathrm{and}\hskip2pt x\in D\right\}. $$

Under this definition, an event $ R\in \varepsilon $ regards some act $ x $ as permissibly chosen out of a set of alternatives $ B $ if and only if $ B $ is $ R $ -related to another set $ D $ which contains $ x $ . This definition has the attractive property that if $ R $ is a choice function, then $ {C}_R=R $ . The intuition here is that, ordinarily, if some set $ A $ is $ R $ -related to a set $ B $ by a choice function $ R $ , we take this to mean that the elements of $ B $ are those acts which are permissibly chosen given that $ A $ is the set of available alternatives. Thus, if we want to interpret the elements of Ɛ as telling us something about what can be chosen, we should try to preserve the idea that if some set $ A $ is $ R $ -related to another set $ B $ by some relation $ R\in \varepsilon $ , then the elements $ B $ are acts which are permissibly chosen given that $ A $ is the set of alternatives. The trouble is that if $ \varepsilon $ is not a choice function, it may $ R $ -relate $ A $ to several different sets, including sets which don’t contain any elements from $ A $ ! A conservative proposal for how to interpret the choice instructions $ R $ is giving us is then that if $ A $ is $ R $ -related to $ B $ , then if $ B $ contains elements from $ A $ , then those elements must be permissibly chosen from $ A $ . With this definition in hand, let’s consider some reasonable conditions on a theory of decision-making under moral uncertainty.

V3—Valuation 3: $ {f}_R(x)\ge {f}_R(y)\hskip2pt \iff \hskip2pt \forall B\hskip2pt \subseteq \hskip2pt A:x,y\in B,y\in R(B)x\in R(B). $

N2—Nontriviality 2: it is possible for two acts $ x,y $ to be unequally choice-worthy, i.e., it is not the case that with probability 1 $ x $ and $ y $ are equally choice-worthy.

Theorem 3.6. V3 and N2 are inconsistent.

Proof. Trivial. By unitarity, $ p\left({\mathcal{P}}_0(A)\times {\mathcal{P}}_0(A)\right)=1, $ and by definition of the pseudo choice function $ {C}_{{\mathcal{P}}_0(A)\times {\mathcal{P}}_0(A)}(B)=B $ for any $ B\in {\mathcal{P}}_0(A) $ , because all pairs of the form $ \left(B,C\right) $ are elements of $ {\mathcal{P}}_0(A)\times {\mathcal{P}}_0(A). $ By V3 it then follows that $ {f}_{{\mathcal{P}}_0(A)\times {\mathcal{P}}_0(A)}(x)={f}_{{\mathcal{P}}_0(A)\times {\mathcal{P}}_0(A)}(y) $ , which violates N2. □

Thus, all three attempts so far at modelling theories as events have foundered. The impossibility results we’ve gone through appear to be robust, which suggests that the difficulties with modelling theories as events is not sensitively tied to just what kind of event we choose to model them by. Rather, the problem seems to be with trying to be with trying to represent theories as events in the first place—specifically, with trying to represent them as “opinionated” events. And since the opinionation of events is a consequence of the fundamental commitment of moral uncertaintism (here expressed by V1–V3) to the claim that we ought to desire an act to the extent that a theory says the act is desirable, the problem cannot be avoided. In sum, it is doubtful that the project of moral uncertainty can succeed by representing theories as opinionated events.

4. Theories as possible outcomes

But if all our problems stem at least partly from the fact that we wanted to treat moral theories as events, the alternative is to treat them as outcomes. Our sample space is therefore some set $ \Omega $ of well-formed theories, or choice functions, or preorders. Without loss of generality, let’s say that $ \Omega $ is a set of preorders, and let’s take the full powerset $ \mathcal{P}\left(\Omega \right) $ as our event space. Intuitively, we can interpret the event $ \left\{R\right\} $ as the event in which the moral theory defined by $ R $ is true, and the event $ \left\{{R}_1,{R}_2,\dots, {R}_n\right\} $ as the event in which one of the theories defined by $ {R}_1,\dots, {R}_n $ is true. We can then define a probability distribution $ {p}_i $ over $ \mathcal{P}\left(\Omega \right) $ representing individual $ i $ ’s beliefs about which moral theory is true.

This way of doing things involves a significant departure from the way moral uncertaintists want us to think about moral uncertainty. And in fact, the motivation for moral uncertaintism is sharply undermined if we have to represent theories as sampling outcomes. It is central to the motivation behind this entire literature that moral theories can be thought of as events: one of the key reasons defenders of MEC are attracted to it is that it is supposed to show that even anti-utilitarians who have no confidence in utilitarianism must nonetheless, as a requirement of rationality, substantially follow its directives. Here’s how the argument is supposed to work. We should maximize expected choice-worthiness, according to MacAskill, Bykvist, and Ord (Reference MacAskill, Bykvist and Ord2020), because what you should prefer to do ought to be sensitive to what those theories you have at least some confidence in say you ought to do. Given that it is irrational—according to the so-called ‘regularity’ principle—to assign zero credence to nonempty events, you ought to have some confidence in every logically possible theory, including utilitarianism. And because utilitarianism has very strong opinions about what is right and wrong (the choice-worthiness values it assigns exhibit higher variance than those assigned by other theories), this entails that even if you don’t have much credence in utilitarianism, you should heed its demands to a good degree. The badness (in utilitarianism’s opinion) of not following utilitarianism’s demands swamps its improbability.

I have my doubts that regularity really is rationally mandated, but let’s grant that it is. In that case, the argument works, but if and only if moral theories are events! Since any subset of our sample space defines an event, and therefore a theory that you have no choice but to consider, the regularity principle forces you to assign this theory a nonzero probability. This argument appears frighteningly strong; in fact: even the moral views expressed by Nazism will occupy some subset of our sample space, and Nazism is also very opinionated, and therefore regularity will require our own attitudes to be strongly positively responsive to the views of the Nazis.

But if moral theories are modelled as outcomes, the argument falls apart. Regularity is a requirement on the distribution of real numbers to events, not a requirement on our choice of sample space. Remember, the laws of probability don’t require me to include the outcome “the dice vanishes in a puff of smoke” in my model of a random dice roll. And indeed, if you go to a casino, you will find that the odds they give you in Craps ignore the possibility of such outcomes. Therefore, if theories are outcomes, then rationality cannot force you to consider theories you would prefer to ignore. If I am attracted to moral uncertaintism as a mode of deliberation, but the only moral theories I want to consider are deontology and virtue ethics, rationality can’t stop me. I can easily construct a well-behaved probability space with only two outcomes, whose only atomic events are therefore $ \left\{{R}_{\mathrm{deontology}}\right\} $ and $ \left\{{R}_{\mathrm{virtue}\ \mathrm{ethics}}\right\} $ , and these two events will jointly exhaust the space of possibilities in my model. Every nonempty event in my event space will receive a nonzero probability, satisfying the regularity principle, and yet utilitarian considerations do not feature in my deliberations. You can try to persuade me that it’s somehow unwise to exclude a theory you like from my sample space (I’m skeptical: if a Nazi made that argument to me I wouldn’t be persuaded to include Nazism in my sample space), but you can’t show it’s irrational. In sum, if you are attracted to moral uncertaintism because you find the regularity argument persuasive, then you can’t allow moral theories to be represented as outcomes. Conversely, the fact that we have to represent theories as outcomes should disenchant you with moral uncertaintism. If moral theories are sampling outcomes, then the hope of persuading nonutilitarians to be utilitarians despite themselves fails.

Granted, some moral uncertaintists might reply that their efforts to convince nonutilitarians to be utilitarians despite themselves don’t really depend on regularity. Rather, they might just assert that, as a matter of empirical fact, most people do find utilitarianism at least somewhat plausible and so do consider its claims in their deliberations. If that’s the argument, then it’s no great loss to represent theories as outcomes rather than events; but then again, if that’s the argument, then the argument loses its force. If your efforts to persuade people to be utilitarians despite themselves depends on their already being disposed to act according to utilitarian recommendations, persuasion is either unnecessary (because they’re already disposed to do what you’re nagging them to do) or futile (because your persuasion strategy appeals to dispositions they lack). What was supposed to be impressive about these “hedging” arguments was their potential to force resolute anti-utilitarians to take utilitarianism more seriously despite being generally unwilling to consider its claims in their deliberations, simply by appealing to weak norms of rationality that stand independently of utilitarianism.

Likewise, the central argument, which we discussed earlier, by MacAskill and Ord (Reference MacAskill and Ord2018) against the My Favourite Theory view is in trouble. Recall that the thrust of the objection is that it doesn’t make sense to talk of “the” theory you find most credible because we can always fine grain or coarse grain the level at which we describe theories, and so the theory which gets the highest credence depends on an arbitrary choice of partition of the state space. This objection, though, assumes that theories are modelled as events. If instead we model theories as possible outcomes, the objection has no force, because it suffers from a catastrophic presupposition failure: to say that $ R $ is your favourite theory is to say that the atomic event you give highest credence to is $ \left\{R\right\} $ . But remember that $ \left\{R\right\} $ is not a moral theory—it is the atomic event in which some particular moral theory is true. Since it is an atom, it cannot be fine grained. You can of course coarse grain an atomic partition of the state space, but the events you end up with are not increasingly general moral theories, just increasingly disjunctive events in which one of many moral theories is true: the event $ \left\{{R}_1,{R}_2\right\} $ is the event “either $ {R}_1 $ is true or $ {R}_2 $ is true,” not some highly general moral theory which is implied by both of $ {R}_1 $ and $ {R}_2 $ . Moral theories under this definition cannot be fine grained or coarse grained.

Note that the point I am making here is quite different to the point Gustafsson and Torpman (Reference Gustafsson and Torpman2014) make in response to MacAskill’s argument: their response is to simply assume that there is a most natural partition—not necessarily the atomic partition—in terms of which the claims of MFT should be couched. The natural partition is something like the level of description of moral theories that you prefer, the one you find most salient. Your favourite theory is therefore the theory you prefer at the level of grain you find most salient. But note that this way of thinking about things necessarily models theories as events in an event space which can in principle be fine or coarse grained. If theories are sampling outcomes though, this whole response also suffers from presupposition failure.

As compensation for this departure, though, we avoid some of the absurdities of our earlier approach. It’s no longer the case that the individual must assign probability 1 to it being true that every act is just as choice-worthy as any other. Instead, the individual must only believe with probability 1 that some theory (possibly including error theory) is true—a much more plausible conclusion. This is because the necessary event $ \left\{{R}_1,{R}_2,\dots \right\} $ is not itself a moral theory. In particular it is not the union of all preorders representing the moral theories under consideration. It is the set of all such preorders, an event which we interpret as the event in which one of the theories under consideration is true.

Further, the process of partitioning the sample space works just as you would hope: if $ \Omega $ contains the ten moral theories you find yourself hesitating between as elements, then you can partition $ \Omega $ into ten subsets, each containing a single element corresponding to one of those ten theories you’re uncertain about. We now even have a way, it seems, of defining expectations of moral choice-worthiness. Suppose, for simplicity, that $ \Omega $ is finite, and that every element of $ \Omega $ is a complete preorder.Footnote ⁶ Then each element of $ \Omega $ can be represented by a real-valued function $ {f}_R $ . From here, we define the expected value of an act $ x $ as $ E(x)={\sum}_{R\in \Omega}{p}_i\left(\left\{R\right\}\right){f}_R(x) $ . Thus, the choice-worthiness of an act is the probability-weighted sum of its choice-worthiness under each moral theory. Provided that every moral theory included as a possible outcome in the sample space is representable by a real-valued function, $ E(x) $ is well-defined, and you can meaningfully instruct individuals to maximize expected choice-worthiness. Finally, the process of moral learning works roughly as you would hope: if you read the works of Peter Singer and become convinced that eating meat is wrong, then you’ve effectively “observed” the event in which at least one theory which implies that eating meat is wrong is true—thus by updating by conditionalization you come to assign credence 0 to the atomic events in which theories which fail to imply this are true, and you raise your credence in each atomic event in which one of the remaining theories is true.

This approach is a slighly more general version of that proposed by Riedener (Reference Riedener2020) and Dietrich and Jabarian (Reference Dietrich and Jabarian2021). They model moral theories as vNM utility functions, then postulate a relation on the set of probability distributions over these functions (i.e., implicitly, the vNM functions are treated as sampling outcomes), which they assume also to satisfy the vNM axioms. As Dietrich and Jabarian helpfully explain, this model effectively imagines that individuals are behind a veil of ignorance, and don’t know what their values will be. They then exploit results like Harsanyi’s theorem to argue that the maximization of expected value is the only rule for aggregating the different vNM functions which satisfies the ex ante Pareto axiom (i.e., if every theory expects A to be better than B, then you should prefer A to B). The only difference between these approaches and the one I’ve just sketched is that I don’t assume that moral theories must satisfy the vNM axioms, nor that any acceptable decision rule under uncertainty must also satisfy these axioms. It’s worth noting that Riedene, Dietrich, and Jabarian are, to my knowledge, the only authors to provide an explicit probability model of moral uncertainty, so it may speak in favour of representing theories as sampling outcomes that this is how they themselves do it.

Maybe all of this is compensation enough for heterodoxy. Unfortunately, while this setup may avoid the blatant absurdities of the previous approach, its modelling capacities are so limited that it fails as a general model of agents experiencing moral doubt, and it cannot successfully be used as a tool in deliberation by such agents either. This can be shown by another impossibility result. Define a choice rule for decision-making under moral uncertainty (or “choice rule,” for short) as a map from a set of possible probability spaces $ \left(\Omega, \mathrm{\mathcal{F}},q\right) $ to a set of preorders over $ A $ , such that $ \Omega $ is a set of binary relations on $ A $ . Intuitively, given that you are uncertain between a given set of moral theories and your uncertainty over these theories is representable by a probability distribution, the choice rule tells you what to do. Now consider the following conditions.

This is another way of saying that individuals are allowed to form doxastic attitudes to any moral theory. Individuals may not be required to consider any particular theory in their deliberations, as argued earlier, but surely they must be permitted to consider any moral theory they wish. Our theory of decision-making under uncertainty must be able to guide and advise any agent who is uncertain of which moral theories are true. If there is some moral theory in which they in fact express some confidence our choice rule must be able to guide them. A choice rule which only works when restricted to a narrow range of possible moral theories fails as a general solution to the problems of moral doubt. A slightly less permissive, but still quite permissive alternative to U is the following.

By strengthening U to Q, we prohibit the individual from taking doxastic attitudes to the empty set, but also to pathological moral theories that do claim that some acts are more morally choice-worthy than others while also denying that any act is at least as choice-worthy as itself. After all, maybe it’s too much to ask a choice rule to provide sensible guidance when the individual has beliefs in moral theories that are simply unintelligible. But at the very least, a theory of decision-making under uncertainty should work for any agent who expresses doxastic attitudes to moral theories that are actively debated by moral philosophers. No such theories (other than error theory) are excluded by Q. Next we define MEC and MFT in this new modelling framework.

This condition can be extended to lexical order in case there are ties.

Note that theorems 4.2 and 4.3 hold if we replace U with any domain condition that requires the choice rule to yield a preorder for probability spaces for which the error theory is a possible sampling outcome. U is much stronger than we need.

How damaging are these results? One can certainly escape theorem 4.1 by imposing a more restrictive domain condition than Q. This is what Riedener (Reference Riedener2020) and Dietrich and Jabarian (Reference Dietrich and Jabarian2021) do by assuming all moral theories satisfy the vNM axioms (including transitivity and continuity), and their approach delivers well defined expectations of moral choice-worthiness.Footnote ⁷ But, for my money, their domain assumption is so restrictive that it simply defeats the project of moral uncertaintism. They assume that the only moral theories you care to consider in your deliberations are ones that are already in the business of maximizing expected value. Therefore, if in the course of your moral deliberations, you wish to entertain and engage with anti-aggregative moral theories like contractualism, or ‘limited aggregation’ theories (Voorhoeve Reference Voorhoeve2014; Steuwer Reference Steuwer2021), or value theories like Larry Temkin’s, which recognises value cycles, then your state of mind is not one which can be modelled on their approach. This is a problem.

For a start, it means that the Riedener-Dietrich-Jabarian (RDJ) approach offers no general answer to the question which moral uncertaintists were out to answer, namely: what should I do when I’m unsure what I should do? Moreoever, it consigns the RDJ approach to total irrelevance in the practical deliberations of anyone who wishes to engage with nonmaximizing moral theories in their moral deliberations. This will include many Bayesians and consequentialists who are otherwise sympathetic to value maximization. I, for example, am not a anti-aggregative deontologist, but I don’t think Scanlonian contractualism is insane: I feel the pull of the TV transmitter room case, and I think that responsible moral inquiry must respond to contractualist arguments and thought experiments. The RDJ approach therefore cannot feature prominently in responsible moral inquiry, since it prohibits me from entertaining any attitudes to contractualism. This self-imposed irrelevance also forestalls any possibility of convincing anyone who wishes to engage reflectively with deontology to “hedge their bets” in favour of consequentialist moral theories. This is for the same reasons as above: any strategy for persuading me to be less of a deontologist that already assumes I don’t consider any form of deontology relevant to my deliberations is a failure.

We need to be at least somewhat permissive in the range of moral theories we allow individuals to express attitudes to, else the project of moral uncertaintism fails on its own terms. Q, I think, is one of the most restrictive domain condition one can impose without vitiating the aims of moral uncertaintism.Footnote ⁸ The inconsistency of Q with MEC is therefore, I submit, a challenge to MEC’s plausibility.

That’s it for theorem 4.1. What about theorems 4.2 and 4.3? Is it such a problem that MFT bottoms out of advice when individuals are unsure of whether error theory is true? Q and LMFT are consistent (thus, by extension, Q and MFT), so LMFT will provide sensible guidance for anyone who does believe there is a right and wrong, but is merely uncertain of what makes something right or wrong. This, I concede, is already pretty good. MFT is less severely impugned by the above results than MEC, but nonetheless I claim that MFT’s inability to guide agents who are in doubt as to whether anything is right or wrong is a serious failing. Not only does MFT fail to provide guidance to all morally uncertain agents, and thus fails as a general theory of decision-making under moral uncertainty, this failure looks like a serious fault in the approach because, curiously, I don’t find it particularly difficult to think of reasonable advice for such an agent. If they find error theory such a threatening possibility, they could be invited to test their reasons for finding it so plausible by reading a healthy dose of Putnam (Reference Putnam2002, Reference Putnam2004), and Williamson (Reference Williamson, Bloomfield and CoppForthcoming).

This last remark might seem a bit flippant, but it points to what I believe is a general and underappreciated flaw with the whole approach proposed in this literature. The rules that have been proposed for decision-making under moral uncertainty are, by design, completely insensitive to the preferences and nonmoral beliefs of individuals. This is the root of many problems. For a start, it leaves these theories unable to rationalize a range of ordinary morally conscientious behaviour. For example, a morally conscientious agent who is unsure what to do might find it quite desirable to spend time reading about and trying to settle metaethical debates, or engage in moral disputation to try and settle what is right and wrong, or seek advice about what to do and how to live from their moral gurus or heroes; these strategies stand a chance of resolving their underlying uncertainty (see e.g., Weatherson Reference Weatherson2014; Sepielli Reference Sepielli2016). Just as we expect a good theory of decision-making under ordinary descriptive uncertainty to have rationalize ordinary evidence-seeking behaviour, and rationalize our willingness to pay for information, we should expect a good theory of decision-making under moral uncertainty to rationalize the evidence-seeking behaviour I’ve described. But, in fact, it’s very hard to see how the approach sketched out in this section could ever recommend the activity of moral inquiry, because no first-order ethical theory will ever advise such behaviour: from each theory’s own point of view, all moral issues are settled, and you simply ought to do what the theory says and, consequently, MFT and MEC will also advise against such moral evidence-seeking.

Unless the moral theory finds intrinsic value in moral deliberation as such (this would be a strange theory indeed), or unless moral deliberation happens to have instrumentally valuable causal effects (e.g., it turns out that, for some arbitrary reason, an act of moral reflection will maximize welfare), any moral deliberation must necessarily compete for your time and energy with other courses of action that actively follow the theory’s guidance, and for that reason moral deliberation will be disfavoured in comparison to those alternatives. To take a stylized example, suppose the individual is uncertain between theories $ {T}_1,\dots, {T}_n $ , and has options $ {A}_1,\dots, {A}_{n+1} $ to choose from where $ {A}_i $ is the option which $ {T}_i $ regards as optimal and $ {A}_{n+1} $ is the option “take longer to reason about which moral theory is true.” From the agent’s own point of view, $ {A}_{n+1} $ might be the optimal course of action because taking the time to reason more might substantially resolve their moral uncertainty. But neither MFT nor MEC will recommend $ {A}_{n+1} $ because no particular moral theory will ever recommend $ {A}_{n+1}. $ So MFT and MEC can’t rationalize ordinary evidence-seeking behaviour.

In fact, it’s not even clear that these approaches have a complete concept of evidence. Most moral uncertainty-ists postulate agents who have no beliefs about anything other than which theory is true, and even those few who do explicitly model individual’s descriptive uncertainty alongside their empirical uncertainty (e.g., Riedener [Reference Riedener2020], Dietrich and Jabarian [Reference Dietrich and Jabarian2021], and Kernohan [Reference Kernohan2021]) do so in a way that makes it impossible to represent the dependence of moral theories on empirical beliefs.Footnote ⁹ The only way individuals can change their moral beliefs is by “observing” the truth of particular moral principles or even entire theories. But our reasons for favouring one ethical theory over another are frequently guided by beliefs about empirical matters.Footnote ¹⁰ For example, my skepticism of preference-satisfactionist utilitarianism is partly driven by my skepticism about empirical matters: I’m skeptical that preferences are cardinally measurable and interpersonally unit comparable. My skepticism could be abated if I were to learn of a new measurement procedure that allowed for truly cardinal measurements of preferences. But neither this learning event nor the change it would induce in my beliefs can be represented in a probabilistic model where the only objects of beliefs are moral theories, because neither my beliefs about measurement nor the evidence I’ve observed are about moral theories. (There’s also a separate concern about what moral uncertaintists should say about what we should do in cases where, conditional on some empirical fact, some moral theory cannot be meaningfully stated, but I put this aside.)

Likewise, the advice we give to the morally conscientious should also depend on their preferences. If it just matters to them that some moral theory—any moral theory—be true because the prospect of nihilism is just too painful to contemplate—like Ivan Karamazov, our agent needs the tears of children to matter—then it seems suitable to advise them to consult defenses of moral realism, or perhaps read good existentialist novels like The Plague by Camus. In contrast, if your moral doubt and anxiety are driven by your concern for the poor yet simultaneous discomfort with coercing people by the (implicit) threat of bayonets into giving away their assets, and what you really want is either a way of making socialism safe for libertarianism (and vice versa), or a way of overcoming your lingering attachment to one or the other, then it seems more suitable to advise you to adopt a stance of cautious liberalism while you consult the back-and-forth between different stripes of libertarian, anarchist, liberal, and socialist political philosophers. The uncertainty we care most about resolving is not the same in these two cases, and good advice to the morally conscientious should be suitably responsive to this fact. But existing models of decision-making under moral uncertainty cannot be responsive to this fact since they are by design unresponsive to people’s preferences.

This is all by way of saying that present approaches to decision-making under moral uncertainty are severely limited, and the problem we have here is fundamental to the modelling approach favoured by the moral uncertaintists. The above impossibility results show that if moral theories are to serve both as objects of belief and as constraints on the distribution of values whose expectation we are supposed to maximize, then we face a trade-off between choosing a measure space that is adequate to represent the space of possibilities which individuals are aware of/wish to consider in their deliberations, and one that does not impose constraints on the distribution of values to acts which are logically impossible to satisfy. And the brief remarks I’ve just offered suggest that they fail in other ways too, by being unresponsive to precisely those individual beliefs and concerns we should expect a good source of advice to be responsive to. Moral doubt is a poignant experience of life, but inventing a whole new decision theory featuring preferenceless agents who only have beliefs about a narrow class of objects to talk about it mires us in confusion.