2. Transformation as awareness growth

Paul picks out phenomenological facts—what it feels like to have certain kinds of experiences—as being beyond one’s epistemic access in some unusual sense, prior to actually having that kind of experience. Paul starts from the idea that all the knowledge that one can amass about third-person matters does not yield knowledge of some first-person matters—what it feels like to occupy certain positions in the world. She extends the lesson of Jackson’s (Reference Jackson1986) Mary case to other sorts of evidence for first-person facts. According to Paul (e.g., Reference Paul2015, 154), it is hard to think of any good source of evidence on which Mary might draw in her black-and-white room—including, say, others’ testimony about what it is like to see color—that would allow her to fully appreciate this entirely new kind of experience.

Even if Paul underestimates testimony as a source of evidence for phenomenological experience, we think that her examples are in fact paradigms of some unusual kind of epistemic limitation that can be relevant for decision-making. The way to make sense of this, we contend, is to see the epistemic transformations as episodes of awareness growth. We defined this earlier as an augmentation in the agent’s possibility space. The agent comes to see that there are practically important ways that the world might be that she hadn’t previously considered. Such augmentations are the result of coming to better appreciate a property of the world or random variable: it may be that one recognizes a further value that some property/variable might take (what is called awareness growth by expansion), or it may be that one recognizes an entirely new property/variable (what is called awareness growth by refinement). In the case of parenthood, you might experience awareness growth by expansion if you come to realize that there is a further state of fatigue that you had not recognized—a kind of tiredness that nonetheless allows one to function reasonably well—that matters in accounting for the costs of lack of sleep. On the other hand, you might experience awareness growth by refinement if you come to realize that your child’s level of sociability matters to how much you enjoy, as a parent, spending time in the company of others.

Why do we claim that Paul’s epistemic transformations are best understood as awareness growth? The best argument for this claim, we suggest, is a negative one: it is hard to see that there are any better alternatives for understanding epistemic transformation and the preceding epistemic limitations. Paul speaks of a lack of knowledge (e.g., Reference Paul2015, 155–58). Is it simply that the agent initially lacks knowledge of which possible phenomenological experience will occur? That kind of epistemic shift is surely not sufficiently remarkable. An initial lack of knowledge about some decision-relevant matter, say, due to not believing or having low confidence in the truth of the matter, is an entirely ordinary predicament and the paradigm case for standard decision theory. One could turn to more radical, nonstandard kinds of uncertainty, such as a temporary inability to specify precise probabilities for the relevant possibilities.Footnote ¹ But even this kind of epistemic predicament does not seem sufficiently extraordinary to capture the cases at hand. An inability even to specify or grasp the relevant possibilities is a much better candidate for spelling out this unusual lack of epistemic access. One might put it this way: it is not that the agent lacks knowledge of which of some possible phenomenological experiences she will experience upon taking some course of action; rather, she lacks knowledge of what are the possible experiences associated with the course of action.Footnote ²

Admittedly, limited awareness and subsequent awareness growth are not limited to personal phenomenological experience. Any kind of novel course of action, whether the novelty be due to the phenomenology of the possible outcomes or other “third-person” properties of the outcomes, can be characterized by limited awareness. But we do not see why this lack of exclusivity should be a problem. We thus proceed to model the decision problems of interest as ones in which the agent recognizes herself to have limited awareness, partly in view of the fact that she anticipates that some of her options would lead to awareness growth. Hence we will refer to the decision maker’s predicament as one of anticipated awareness growth.

Let us continue with the example of the agent deciding whether to become a parent. It is useful first to consider a decision problem for which there is no anticipated awareness growth, as per Table 1. The agent assumes there are n states of the world (from $s_1$ to $s_n$ ) that determine the outcome of the available options, which we assume here to be simply “become parent” (p) and “remain childless” (c). We will assume that, in addition to her uncertainty about other relevant matters, the states capture the agent’s uncertainty about the phenomenological experience of having a child and of not having a child. Finally, $p(s_i)$ is the outcome that obtains when p is chosen and $s_i$ is the true state of the world.

Table 1. Prospective parenthood without anticipated awareness growth

	$s_1$	$\ldots$	$s_n$
Become parent $(p)$	$p(s_1)$	$\ldots$	$p(s_n)$
Continue childless $(c)$	$c(s_1)$	$\ldots$	$c(s_n)$

What does it mean to say that Table 1 does not make room for awareness growth? It means that the states $s_1$ to $s_n$ are thought, by the agent,Footnote ³ to exhaust the set of possible contingencies. Moreover, the outcomes that these contingencies yield under each option, $p(s_i)$ and $c(s_i)$ , are thought to be maximally specific in all ways that might be relevant to the decision.

Consider now the matrix in Table 2. We focus initially on the anticipation of awareness growth by refinement that is represented here. Suppose that $E_1$ to $E_n$ is the agent’s finest and most exhaustive “partition” of the space of contingencies. (We see later that it is not really a partition, according to the agent.) Each $E_i$ is taken to be, not a (fully specified) state of the world, but rather an event that amounts to a set of states. It follows that the outcomes— $p(E_i)$ , $c(E_i)$ —are not thought to be fully specified in all ways that matter to the decision. Perhaps the agent recognizes that there is some aspect of the experience of having a child that she thinks matters to her evaluation, but she cannot really specify what this aspect of experience is, letalone the different ways it may be realized. The agent recognizes her limited awareness, let us assume, but does not know how to further partition the $E_i$ into the states that fully determine the outcomes of the options in all relevant detail.

Table 2. Prospective parenthood with anticipated awareness growth

	$E_1$	$\ldots$	$E_n$	??
Become parent $(p)$	$p(E_1)$	$\ldots$	$p(E_n)$	$p(??)$
Continue childless $(c)$	$c(E_1)$	$\ldots$	$c(E_n)$	$c(??)$

Now we turn to the anticipation of awareness growth by expansion that is represented in Table 2. The “??” column in the table represents that there is some further event that the agent cannot specify and yet considers part of the event space, such that the other events, $E_1$ to $E_n$ , are not an exhaustive set of mutually exclusive events. We call the “??” event a “subjective catchall” (Steele and Stefánsson Reference Steele and Orri Stefánsson2021b, sec. 6.3). In other words, the agent is aware of the fact that there might be contingencies that are inconsistent with all of $E_1$ to $E_n$ and which would, if they materialize, determine the outcome of the two available options.

Note that, formally speaking, both types of anticipated awareness growth are cases of refinement. Although we call the latter case “expansion,” it is modeled as if the agent anticipates refining her catchall. Indeed, if an agent anticipates awareness growth of any kind, then she must already account for it in some way, be that via crudely described outcomes, the most extreme being a subjective catchall. This is further reason why our account lines up well with Paul’s characterization of decision problems that have the potential to bring about transformative experience. For Paul, the crucial limitation is not that some possible outcomes are not accounted for at all but rather that their values are unknown. Our model is not inconsistent with this basic idea. But it will be shown in the next section that the details of anticipated awareness growth matter; although the agent does not know for sure the value of some outcomes, she may nonetheless assign them an expected value.

3. Managing awareness growth

The question now is whether anticipated awareness growth presents a special obstacle to decision-making. Does the shift from a decision problem like that in Table 1 to one like that in Table 2 require a radical shift in the notion of rational choice? We will argue that it does not, and in this sense, our account of transformative decisions diverges from that of Paul. Paul’s position is that one cannot choose rationally insituations where the options may result in a transformative experience. Here we argue that this is too pessimistic—agents can reason about the value and probability of catchall and other not-fully-specified outcomes in a relatively ordinary way. In the next section, we respond to objections and compare our view with a rival approach.

We offer a way to reason about the value of the catchall. (We focus on anticipated “expansion,” to simplify our analysis, but all of what we say applies to anticipated “refinement” too.) The first thing to note is that there is general agreement among scholars working on (anticipated) awareness growth that it is possible to infer from an agent’s preferences over options what is the value the agent assigns to a catchall. We can examine an agent’s conditional preferences. Suppose that one prefers becoming a parent to remaining childless, conditional on events $E_1$ to $E_n$ . Still, overall, one prefers to remain childless. Is this inconsistent? No, it simply reveals that one prefers remaining childless to becoming a parent conditional on the catchall, and sufficiently so that this overturns the preference for becoming a parent conditional on the other events (in light of the probabilities for these events).Footnote ⁴

But Paul is unlikely to be impressed by this appeal to a “preference-first” (Pettigrew Reference Pettigrew2015) conception of decision theory and catchall evaluation. Paul is interested here in a “deliberative” conception of decision theory (to continue with Pettigrew’s terminology), whereby the agent appeals to decision theory to arrive at well-reasoned preferences. We want to know not simply that the agent assigned some value to outcomes under a catchall event that yielded consistent overall preferences but why, or on what basis, she assigned those values and thus had those preferences.

What we propose is a coherence principle that indirectly bears on deliberation in that it suggests a procedure for arriving at the values of outcomes associated with a catchall. We say more on the deliberative procedure later. First, let us state the coherence principle, which we dub Awareness Reflection (outcome values). It concerns how the agent’s current evaluation of an outcome relates to her predicted future evaluations of that outcome (insofar as she makes such a prediction and considers these future attitudes to be more enlightened) when she anticipates awareness growth. As the name suggests, the principle is inspired by a similar principle known simply as Reflection and defended by van Fraassen (Reference van Fraassen1984), concerning probabilistic belief and applicable in ordinary situations that do not involve anticipated awareness growth.Footnote ⁵ In short, the reflection principle that we will explore states that an agent’s evaluation of an outcome, conditional on having some evaluation given awareness growth (and no other personal change), should just be that more aware evaluation.Footnote ⁶

We now state the core principle more formally. Let v and $v+$ represent cardinal value functions that are comparable in units and levels. Assume that v represents the agent’s current value function. Let ${\bf v}^+$ stand for the proposition that the agent will have value function $v^+$ at time t, having experienced awareness growth and no other personal change that would mean she was less enlightened. Then Awareness Reflection (outcome values) can be stated as follows:

Awareness Reflection (outcome values). For all options f and all events $E_i$ (including the catchall), and for any v and $v^+$ , and for all ${\bf v}^+$ of which the agent is aware,

$$v(f(E_i)\mid {\bf v}^+) = v^+(f(E_i)).$$

The agent’s current value for any catchall outcome is thus constrained by her predicted future values for this outcome.

In special circumstances, the agent’s current value for any given outcome is the expectation of her future values for this outcome. The special circumstances are those in which the following conditions are met. First, the agent is sure she will experience awareness growth and no other personal change. Moreover, she predicts that her probability distribution over the event space, $E_1, \ldots, E_n,$ plus the catchall, ??, will be constant upon awareness growth (more on this assumption later). So she predicts that only her values will change upon awareness growth. The various candidates for her future value function form a partition of her possibility space, represented by $\{{\bf v}_1^+, \ldots, {\bf v}_z^+\}$ .Footnote ⁷ Furthermore, the agent is indifferent to (i.e., neither desires nor detests) whatever values she will have once her awareness grows (since she assumes these will be the result of only having undergone awareness growth). The latter means that we can assume that $v(f(E_i)\mid {\bf v}^+)=v(f(E_i)\& {\bf v}^+)$ .Footnote ⁸ Now, because $\{{\bf v}_1^+, \ldots, {\bf v}_z^+\}$ is a partition of the agent’s possibility set, Jeffrey’s (Reference Jeffrey1965) conditional expected value formula implies that (where P is the probability function)

$$v(f(E_i))= \sum _{i=1}^z P({\bf v}_i^+\mid f(E_i))\cdot v(f(E_i)\& {\bf v}_i^+).$$

But then Awareness Reflection (outcome values) and the preceding conditions together imply that

(1) $$v(f(E_i))= \sum _{i=1}^z P({\bf v}_i^+)\cdot v_i^+(f(E_i)).$$

One of the foregoing assumptions is that the agent predicts her probabilities over the event space to remain constant upon awareness growth. Otherwise, the current value of outcomes must be sensitive to how future values and future probabilities are predicted to covary. This more complicated case can be captured by a more general reflection principle concerning how current all-things-considered preferences should relate to predicted more aware all-things-considered preferences. See Steele and Stefánsson (Reference Steele and Orri Stefánsson2021b, sec. 7) for the formulation of such a principle, dubbed Preference Awareness Reflection. Its defense turns on the importance of avoiding sure loss in a sequential-choice setting (Steele and Stefánsson Reference Steele and Orri Stefánsson2021b). The specific reflection principle outlined here—Awareness Reflection (outcome values)—can be derived from the more general Preference Awareness Reflection. See the appendix for further details.

Reflection principles are often thought implausible in requiring an agent to defer to her future self, even if she does not respect this future self in the appropriate sense. But note that Awareness Reflection (outcome values) is highly restrictive because it concerns only those future states of greater awareness that do not involve any other personal change that would make the agent more defective. Moreover, the agent must actually entertain these possibilities regarding her future values. So Awareness Reflection (outcome values) may rarely be applicable.

Let us finally return to how the principles described herein bear on an agent’s deliberations and suggest a template for explaining how she arrives at her current values for the catchall outcomes. We suggest that when reasoning about the value of an outcome, under the assumption that one’s evaluation of it might be affected by an anticipated (potentially transformative) growth in awareness, the best that an agent can do is to ask herself how someone who has undergone the anticipated awareness growth, but otherwise has the same preferences and beliefs, would judge the outcome in question.Footnote ⁹ In particular, she should defer to someone who is exactly like herself, except more aware. Often she will not know for sure how her more aware self would evaluate an outcome. But she may consider some possible future attitudes more and less probable, in which case, the best she can do is to try to align her current evaluation with what she currently expects her more aware evaluation will be.

4. Responding to skepticism

Our coherence principles and associated deliberative procedure leave some questions unanswered. In particular, even if one could form some opinion—in part by observing similar but more enlightened others—about one’s more-aware evaluation of a catchall, is it really plausible that one could have a probability distribution over the possible evaluations? And do the principles apply in cases of personally transformative experience, whereby there is a shift in one’s fundamental values? One might see only negative responses to these questions and therefore agree with Paul that there is no rational basis for one’s current evaluation of a catchall outcome. Perhaps an entirely different way of evaluating a catchall is called for that is not about trying to guess one’s attitude to this catchall but rather reflects one’s generic attitude about encountering the unknown. Karni and Vierø (Reference Karni and Vierø2017) make a suggestion like this, and it seems to be in line with Paul’s remarks about basing potentially transformative decisions on the “revelatory” value of discovering the content of the catchall and one’s attitudes toward it.

We think this kind of move, in terms of valuing catchall outcomes, is a far last resort. It amounts to a rather ad hoc narrowing of the content of catchall outcomes to a very inward-looking property: something like “surprise about one’s attitudes.” Moreover, there is more to say in favor of deliberating about how one’s future self will view catchall outcomes. Start with the latter of the aforementioned issues. There are two stances one may take toward a personally transformative awareness growth. The personal transformation may be predicted to yield a more enlightened self, to whom one has good reason to defer, or else the personal transformation may be predicted to yield an entirely different self, to whom one has no reason to defer. Either way, one can defer just to those future versions of oneself that are more enlightened. One’s current values should simply be one’s expectations of one’s future values conditional on an “orderly (or enlightening) awareness growth” (recall note Footnote 7). Surely one has some confidence in such an awareness growth, however little.

Moreover, we hold that there is further inductive basis for predictions about what will be one’s future values having experienced awareness growth: similar past decision problems. The inference that we have in mind, based on one’s past experience of similar circumstances, has been proposed by Grant and Quiggin (e.g., Reference Grant and Quiggin2013). Our approach might be seen as a finer-grained version of theirs. Grant and Quiggin focus on environmental intervention and appeal to something akin to “Murphy’s law”Footnote ¹⁰ in this context: “sometimes unexpected things happen, and the unexpected is always bad.” We suggest that the reference classes of similar decision problems may be narrower and the inferences more nuanced. For instance, one might infer that “in career-changing decisions, sometimes unexpected things happen, and sometimes the unexpected is moderately good and sometimes moderately bad.” More generally, it would seem reasonable to have different expectations about the (un)desirability of the unexpected, depending on what type of decision is involved.