<i>Inquiry Under Bounds</i>

All theories of rationality face a challenge. Our intuitions about rationality make unrealistic demands of us. So how can a theory take our imperfect reasoning to be rational while at the same time preserving our unrealistic intuitions? David Thorstad has written a book that tries to square this circle. His questions are pressing and important. His grasp of the relevant literature is enviable. His ability to bring a wide range of philosophical and empirical considerations to bear in support of his view is impressive. And his final theory has many virtues. But does he manage to vindicate both our imperfect reasoning and our unrealistic intuitions? Let’s take a look.The Standard Picture reflects our unrealistic intuitions. Rationality on this picture is reserved for ideal agents unburdened by pesky human limitations. The attempt to apply Standard Picture expectations to real people has “bleak implications” for human rationality (Nisbett and Borgida 1975). Thorstad rejects these bleak implications—at least with respect to inquiry.Consider an example. This morning, I randomly chose a pair of gym socks from my sock drawer. I didn’t conduct an exhaustive survey to find the best pair of socks. Was my behavior rational? For Thorstad, it was—but not because I reasoned like an ideal agent. I didn’t. In fact, on the Standard Picture, my sock-choosing behavior was just more confirmation that, rationality wise, I’m a big disappointment. For Thorstad, however, my sock-choosing behavior was rational because it satisfied two conditions. First, it was my best option. Of all the ways I might have chosen gym socks this morning, random choice maximized expected utility, where this is calculated using a wide range of values (more on this soon) and probabilities derived from my available evidence. And, second, my behavior was prompted by considerations that made it the best option: I chose randomly because I’ve got better things to do with my morning than search for the perfect pair of gym socks. Thorstad’s Reason-Responsive Consequentialist View (or RRCV) is a theory of bounded rationality. It accounts for a reasoner’s limitations, if any, from the get-go. It’s a unified normative framework for all agents, bounded and unbounded.RRCV takes rationality to be environment-relative (or “ecological,” in the jargon of psychologists). Random choice was rational because—not to get too braggy—my sock drawer is always full of comfy gym socks. Random choice wouldn’t have been rational had the drawer been mostly full of mittens. This is a simple but important point. A large literature in psychology shows that in typical (but not all possible) environments, several of our default heuristics are surprisingly effective, despite the scorn heaped on them by the Standard Picture.RRCV delivers a vindicatory epistemology: Our inquiries are typically rational. They are rational far more often than the Standard Picture allows. This is because we work smart, not hard. We seldom use expensive ideal strategies (e.g., closely evaluating every sock in the sock drawer). We usually use cheap heuristics (e.g., random choice) that deliver good (and occasionally ideal) results. If we focus on the consequences of a single episode (e.g., choosing this pair of socks), we’ll usually get somewhat better results with ideal strategies. But the benefit of cheap heuristics is that they free us to do other valuable things with our finite resources (like drive the kids to school, go to the gym in socks that are plenty, if not perfectly, comfy, and still get to work on time). If we tried to live up to the ideals foisted on us by the Standard Picture, the juice just wouldn’t be worth the squeeze.That’s the basic outline of Thorstad’s theory. Now let’s dig into the details. His overall strategy is to defend a theory of practical rationality, apply that theory to inquiry, and establish that human inquiry is typically rational (203). The theory consists of three elements.It’s worth noting that utility-maximizing theories are prototypical Standard Picture theories. But since RRCV doesn’t require agents to calculate expected utilities, it is a theory of bounded rationality. The sock example illustrates that RRCV provides a natural way to account for how limited creatures engage in rational inquiry. It illuminates how we ought to select topics, seek evidence, sift evidence, halt inquiry, and arrive at conclusions. Curiously, however, Thorstad does not apply RRCV to the products of inquiry.Thorstad does not discard the Standard Picture. Instead, he apportions part of our cognitive lives (inquiry) to a theory of bounded rationality and another part of our cognitive lives (the attitudes that result from inquiry) to the Standard Picture. This is how Thorstad tries to square the circle. He vindicates human rationality with RRCV and our unrealistic intuitions with the Standard Picture. Thorstad’s theory takes our inquiries to be rational more often (and perhaps far more often) than the products of our inquiries (178).It’s worth noting just how much tradition Thorstad preserves. Even his theory of bounded rationality has the look and feel of a Standard Picture theory. As I see it, this philosophical conservatism leads to three problems. The first is caused by level separation. For level separation to work, there needs to be separation between what’s rational and what’s irrational. But there’s not. Suppose RRCV says that it’s rational for Robin to inquire using a heuristic that employs a fraction of their total evidence. The heuristic’s final step, the decision rule, says: Accept p. But suppose that the belief that ∼p is rational given our preferred Standard Picture theory (e.g., it’s best supported by Robin’s total evidence). There’s a tension: It’s rational for Robin to accept p (RRCV) but believe ∼p (Standard Picture). This problem is exacerbated by the fact that Thorstad takes rationality to be deontic: One is rationally required to X just in case one ought to X (32). Every time inquiry is rational but the result of that inquiry is irrational (or vice versa), Thorstad’s theory yields a pair of ‘ought’ claims that cannot both be satisfied. This is especially problematic given that Thorstad accepts the ‘ought’ implies ‘can’ principle (128). (Level separation causes analogous problems for practical inquiry. And insofar as practical attitudes, such as preferences and intentions, turn out to be irrational, it’s hard to see how this irrationality doesn’t bleed into a lot of theoretical inquiry.)The second problem for RRCV is that it seems to be self-defeating. By its own lights, it’s too unwieldy to inform rational inquiry into rational inquiry. RRCV evaluates each choice-point in an inquiry by comparing its expected utility to the expected utilities of other possible choices. And all those expected utilities are a function of the inquirer’s total evidence and every value in Thorstad’s rich axiology. This is not a theory humans can use. Compared to RRCV, the models of rational inquiry psychologists build are absurdly simple (162). Thorstad takes these models to be simplified applications of RRCV. And that’s fair, but it doesn’t really address the objection. To see why, consider that Thorstad is a philosophical lumper. Lumpers prefer fewer large categories, whereas splitters prefer a larger number of small categories. A philosophical splitter will propose different, but related, normative frameworks for empirical inquiry, prudential inquiry, aesthetic inquiry, moral inquiry, and so on. Thorstad lumps all these disparate kinds of inquiry into a single framework. Splitters seem to be in a much better position to develop frameworks for specific types of inquiry (empirical, prudential, moral, aesthetic) that we can use to understand and predict the behaviors of real rational agents. What’s more, it’s possible that a splitter’s frameworks could be applied to improve how people reason about (say) causation, sunk costs, moral obligations, or film. By taking us so far away from the action, by its own lights, RRCV risks being far less rational than more specific and useful splitter theories.The third problem for RRCV concerns normative dilemmas. Here’s a classic dilemma: Epistemic considerations pull Alex to believe p while moral considerations pull Alex to withhold judgment and seek more evidence. Splitters can explain dilemmas (“Moral values versus epistemic values!”) but have trouble resolving them. Lumpers resolve dilemmas so efficiently that they have trouble explaining them (“Just believe that p, what’s the dilemma?”). But this isn’t my worry. Thorstad’s framework has resources to explain normative dilemmas. My worry has to do with the way RRCV resolves normative dilemmas. It lumps values together and then swamps some of them in its expected utility calculations. For example, Robin’s use of a cheap heuristic might not maximize expected epistemic utility, but it’s rational because it maximizes overall expected utility. This explanation is a signature success of RRCV. But there’s a dark side to values getting swamped in the consequentialist maw.Consider Draper, a sexist living in a sexist culture. Some of Draper’s heuristics reflect his sexist attitudes. For example, he automatically infers that unfamiliar women at work are secretaries, and in his environment, this heuristic is highly accurate. Thorstad regards “it as a non-negotiable datum” that this sort of stereotype-based inference is rationally impermissible, despite its “resemblance to structurally similar cases of rational heuristic inference” (99). A classic objection to consequentialism is that it delivers fewer non-negotiables than we’re comfortable with, and that’s the problem here. The consequentialist machinery that swamps epistemic values (to vindicate Robin’s epistemically imperfect inquiries) will just as efficiently swamp moral values (to vindicate Draper’s sexist inquiries). In fact, this swamping is a practical certainty in Draper’s sexist environment. A lumping-swamping theory of rationality is not going to make Thorstad’s “non-negotiable datum” stick. Splitters, on the other hand, are better positioned to prevent moral values from being swamped in the evaluation of Draper’s inquiries.Let me pause here to emphasize that I don’t take these objections to be disqualifying. Remember, I think that every theory of rationality carries significant costs. Every theory has to pick its poison. So when I push Thorstad to be a splitter or to abandon the Standard Picture, he can (with justification) push back by noting that this means flouting some very robust intuitions. And, in fact, that’s exactly what he does (58–60, 178–86). Ultimately, the issue for Thorstad’s theory—and, really, for any theory—is whether the juice is worth the squeeze.Philosophers interested in rationality have a choice. We can have a theory of bounded rationality that explains how we can be both rational and imperfect. Or we can have a Standard Picture theory that respects our unrealistic intuitions. Thorstad tries to do both. The result is a book that’s admirable, impressive, and in some ways successful. But in philosophy, as in life, we tend to get ourselves into trouble when we try to have it all.