Cushman argues that rationalization is rational because it lets us construct useful fictions, which on average will yield “true beliefs and adaptive desires” (sect. 3.5, para. 5). He presupposes that the benefit of rationalized beliefs will offset any such costs incurred when false beliefs are produced. But how often may we expect the benefit-to-cost ratio to go the right way? It seems to us to be an empirical question that Cushman perhaps does not adequately address. (And note that the deck starts stacked against him, because rationalization will always involve at least some false beliefs about the recent past states of one's own mind.) We will stipulate that it is prima facie reasonable to assert that cognitive processes are adaptive within their evolutionarily relevant domains. Cushman extends this domain beyond the Paleolithic era by bringing in the idea of cultural evolution (which puts us on a time scale of decades), but in an age of Google, algorithmic stock market trading, and six-tier pension plans, it is unreasonable to expect culture to adapt in time to technological changes.
That leaves the process of reinforcement learning as an adaptive mechanism that might calibrate my rationalization process into something useful. For high-frequency, feedback-rich tasks like driving to work, I might be expected to recalibrate, eventually changing my intuitions if I move from Boston to Brisbane. For complicated, low-repetition, poor-feedback tasks like saving for retirement, I can at best assert a form of bounded rationalization, subject to all the same cognitive challenges as any other information-processing task (for a recent treatment, see Benjamin Reference Benjamin2018).
For example, myside biases might manifest as overconfidence in my ability to pick stocks – and to retain that overconfidence after years of failing to beat the market (see Daniel & Hirshleifer Reference Daniel and Hirshleifer2015, or see our father's investment returns for the last 40 years). In general, in the presence of confirmatory biases, even large amounts of feedback may well fail to drive people to correct beliefs (see, e.g., Rabin & Schrag Reference Rabin and Schrag1999).
It gets worse. As Cushman himself notes, our theory of mind capacities are “biased to perceive all behavior as rational, even though much behavior is not” (sect. 3.5, para. 6). But he has unfortunately failed to follow up on a consequence of this psychological fact: This bias will also afflict the process of rationalization. Rationalization will thus only entertain a radically restricted subset of possible hypotheses, and accordingly, whenever the relevant truths about the environment fall outside that subset, rationalization will not be beneficial.
Cushman's diagram in Fig. 2b presents three unconscious systems: instincts, norms, and habits. He argues that these systems join with reasoning to influence action, and those actions then provide signals to the underlying systems. However, this list of unconscious processes leaves out such possible intuitively irrational influences as myside biases, present-biased discounting, primacy and recency effects, and so forth. Because such biases will not be perceived as rational, rationalization will both fail to extract any information value from them where they have such value and make it harder for agents to correct for them when they lack such value.
Consider the example of saving for retirement. If I observe that older workers are relying heavily on their defined benefit pensions, and that fact unconsciously becomes the basis for rationalizing skimping on my own retirement contributions, then I will have made a very costly error. Even if I have a defined benefit pension plan, it is very likely to be stingier than that of the older workers whose choices have informed my rationalization. If I knew that my rationalization was really based on the older workers, it would be easier to notice that they have a different plan than I do.
Cushman argues that rationalization will “occasionally be maladaptive” (sect. 1.1, para. 10). We believe this statement is highly optimistic. We believe that modern society offers a large number of domains that are far removed from the conditions in which our biology and culture evolved, and which are too dynamic or feedback-poor to have calibrated reinforcement learning (see Shanteau Reference Shanteau1992). These domains include high-cost, low-frequency decisions like what education to get, how to save for retirement, and whether/whom to marry – none of them choices that provide for much learning, but which are extremely consequential. How often will people be calibrated enough for rationalization to be reasonably adaptive? What are the relative costs of adaptive versus non-adaptive rationalizations? Without positive evidence to resolve such questions, rationalization cannot yet be viewed as rational.
However, we would conclude by emphasizing that we view this as an open empirical question. One possibility Cushman may wish to explore, in trying to arrive at a positive verdict, is that some of these biases will not merely corrupt rationalization, but also be compensated by it. Consider our example of being an overconfident investor. If I was otherwise perfectly adapted to the modern stock market, then a misbelief could immediately push me into bad choices. However, behavioral scientists have documented entire hosts of “non-rational influences.” For example, Benartzi and Thaler (Reference Benartzi and Thaler1995) demonstrate that a different influence, myopic loss aversion, can paralyze investors. These countervailing biases, one promoting investment and one suppressing it, could lead someone to better investment decisions than they would make if subject to only one of them. See Bahaddin et al. (Reference Bahaddin, Weinberg, Luna-Reyes and Andersen2019) for further discussion of countervailing biases and Akerlof and Yellen (Reference Akerlof and Yellen1985) for the implications of starting far away from the optimum. The presence of multiple biases supports Cushman's claim that even inaccurate rationalizations could potentially be adaptive.
We find Cushman's descriptive claim that rationalization can be viewed as a form of representational exchange to be a fruitful one, and we applaud his opening up the question of whether rationalization has perhaps gotten an unfairly bad rap. We contend that his further normative claim on behalf of rationalization unfortunately falls short. We do not, at this time, have reason to think that rationalization is rational, at least, not based on Cushman's arguments.
You have
Access
Cushman argues that rationalization is rational because it lets us construct useful fictions, which on average will yield “true beliefs and adaptive desires” (sect. 3.5, para. 5). He presupposes that the benefit of rationalized beliefs will offset any such costs incurred when false beliefs are produced. But how often may we expect the benefit-to-cost ratio to go the right way? It seems to us to be an empirical question that Cushman perhaps does not adequately address. (And note that the deck starts stacked against him, because rationalization will always involve at least some false beliefs about the recent past states of one's own mind.) We will stipulate that it is prima facie reasonable to assert that cognitive processes are adaptive within their evolutionarily relevant domains. Cushman extends this domain beyond the Paleolithic era by bringing in the idea of cultural evolution (which puts us on a time scale of decades), but in an age of Google, algorithmic stock market trading, and six-tier pension plans, it is unreasonable to expect culture to adapt in time to technological changes.
That leaves the process of reinforcement learning as an adaptive mechanism that might calibrate my rationalization process into something useful. For high-frequency, feedback-rich tasks like driving to work, I might be expected to recalibrate, eventually changing my intuitions if I move from Boston to Brisbane. For complicated, low-repetition, poor-feedback tasks like saving for retirement, I can at best assert a form of bounded rationalization, subject to all the same cognitive challenges as any other information-processing task (for a recent treatment, see Benjamin Reference Benjamin2018).
For example, myside biases might manifest as overconfidence in my ability to pick stocks – and to retain that overconfidence after years of failing to beat the market (see Daniel & Hirshleifer Reference Daniel and Hirshleifer2015, or see our father's investment returns for the last 40 years). In general, in the presence of confirmatory biases, even large amounts of feedback may well fail to drive people to correct beliefs (see, e.g., Rabin & Schrag Reference Rabin and Schrag1999).
It gets worse. As Cushman himself notes, our theory of mind capacities are “biased to perceive all behavior as rational, even though much behavior is not” (sect. 3.5, para. 6). But he has unfortunately failed to follow up on a consequence of this psychological fact: This bias will also afflict the process of rationalization. Rationalization will thus only entertain a radically restricted subset of possible hypotheses, and accordingly, whenever the relevant truths about the environment fall outside that subset, rationalization will not be beneficial.
Cushman's diagram in Fig. 2b presents three unconscious systems: instincts, norms, and habits. He argues that these systems join with reasoning to influence action, and those actions then provide signals to the underlying systems. However, this list of unconscious processes leaves out such possible intuitively irrational influences as myside biases, present-biased discounting, primacy and recency effects, and so forth. Because such biases will not be perceived as rational, rationalization will both fail to extract any information value from them where they have such value and make it harder for agents to correct for them when they lack such value.
Consider the example of saving for retirement. If I observe that older workers are relying heavily on their defined benefit pensions, and that fact unconsciously becomes the basis for rationalizing skimping on my own retirement contributions, then I will have made a very costly error. Even if I have a defined benefit pension plan, it is very likely to be stingier than that of the older workers whose choices have informed my rationalization. If I knew that my rationalization was really based on the older workers, it would be easier to notice that they have a different plan than I do.
Cushman argues that rationalization will “occasionally be maladaptive” (sect. 1.1, para. 10). We believe this statement is highly optimistic. We believe that modern society offers a large number of domains that are far removed from the conditions in which our biology and culture evolved, and which are too dynamic or feedback-poor to have calibrated reinforcement learning (see Shanteau Reference Shanteau1992). These domains include high-cost, low-frequency decisions like what education to get, how to save for retirement, and whether/whom to marry – none of them choices that provide for much learning, but which are extremely consequential. How often will people be calibrated enough for rationalization to be reasonably adaptive? What are the relative costs of adaptive versus non-adaptive rationalizations? Without positive evidence to resolve such questions, rationalization cannot yet be viewed as rational.
However, we would conclude by emphasizing that we view this as an open empirical question. One possibility Cushman may wish to explore, in trying to arrive at a positive verdict, is that some of these biases will not merely corrupt rationalization, but also be compensated by it. Consider our example of being an overconfident investor. If I was otherwise perfectly adapted to the modern stock market, then a misbelief could immediately push me into bad choices. However, behavioral scientists have documented entire hosts of “non-rational influences.” For example, Benartzi and Thaler (Reference Benartzi and Thaler1995) demonstrate that a different influence, myopic loss aversion, can paralyze investors. These countervailing biases, one promoting investment and one suppressing it, could lead someone to better investment decisions than they would make if subject to only one of them. See Bahaddin et al. (Reference Bahaddin, Weinberg, Luna-Reyes and Andersen2019) for further discussion of countervailing biases and Akerlof and Yellen (Reference Akerlof and Yellen1985) for the implications of starting far away from the optimum. The presence of multiple biases supports Cushman's claim that even inaccurate rationalizations could potentially be adaptive.
We find Cushman's descriptive claim that rationalization can be viewed as a form of representational exchange to be a fruitful one, and we applaud his opening up the question of whether rationalization has perhaps gotten an unfairly bad rap. We contend that his further normative claim on behalf of rationalization unfortunately falls short. We do not, at this time, have reason to think that rationalization is rational, at least, not based on Cushman's arguments.