Ainslie's review of willpower, and his delineation of resolve, suppression, and habit complements and extends a lifetime of work (Ainslie, Reference Ainslie1974, Reference Ainslie1975). His synthesis of disparate literatures is illuminating, but his neglect of risk preferences and subjective beliefs necessitates a reliance on hyperbolic discounting as the primary explanation for impulses and the attendant exercise of willpower. To provide a complete account of the motivational bases of choice, Ainslie should extend his framework to incorporate risk attitudes and subjective beliefs.

Risk preferences are primitives in economic theory because they define how agents respond to risk and uncertainty. Similarly, subjective beliefs about uncertain events, whether updated according to Bayes' rule or not, interact with risk preferences (and time preferences) in driving choice. In fact, subjective beliefs are the risk perceptions that define the risky choice objects one then applies risk preferences to evaluate. Atemporal risk aversion refers to aversion to variability of outcomes at a point in time, whereas intertemporal risk aversion refers to aversion to variability of outcomes over time. Ainslie's review is replete with references to “risk,” “expectations,” “contingencies,” and “prospects.” This raises the question: Why does risk not formally feature in his theoretical framework?

This neglect of risk is surprising given the economic literature that Ainslie cites. For example, Kőszegi and Rabin (Reference Kőszegi and Rabin2009) is explicitly a theory of reference-dependent consumption plans that incorporates risk attitudes, time preferences, and subjective beliefs. Furthermore, Gul and Pesendorfer (Reference Gul and Pesendorfer2001), Benhabib and Bisin (Reference Benhabib and Bisin2004), Fudenberg and Levine (Reference Fudenberg and Levine2006), and Loewenstein, O'Donoghue, and Bhatia (Reference Loewenstein, O'Donoghue and Bhatia2015) all incorporate risk. Finally, Ainslie lauds Bénabou and Tirole (Reference Bénabou and Tirole2004) for, “the most complete expression of the recursive self-prediction model in the terms of economics” but, again, they incorporate risk attitudes and subjective beliefs, Ainslie does not. Thankfully, Ainslie does not follow Rachlin, Logue, Gibbon, and Frankel (Reference Rachlin, Logue, Gibbon and Frankel1986), Rachlin, Castrogiovanni, and Cross (Reference Rachlin, Castrogiovanni and Cross1987), and Rachlin, Raineri, and Cross (Reference Rachlin, Raineri and Cross1991) who argue that choice under risk can be tied to a temporal framework by interpreting the probability of a reward as the delay to, or rate of reinforcement of, reward; this approach is refuted by Hofmeyr (Reference Hofmeyr2020). The complete omission of a stochastic component leaves Ainslie's theory incomplete.

Consider the discussion of hyperbolic discounting and preference reversals. The shape of a hyperbolic discounting curve can incline an agent who prefers a larger, later (LL) reward over a smaller, sooner (SS) reward, when both are sufficiently delayed, to switch preference to the SS reward when its receipt is imminent. This is indeed a potential implication of hyperbolic discounting, but is there a simpler, primitive explanation for preference reversal?

SS and LL rewards differ in terms of their risk of receipt, with the LL reward presumably more risky than the SS reward, in the sense that it may not actually materialise or the agent might be dead before receipt. The continuous temporal risk of death from predators or natural causes, for example, starvation, is a constant theme in modelling risky foraging behaviour of species. When both rewards are sufficiently delayed, any difference in perceived riskiness is arguably small. Just prior to receipt of a SS reward though, any risks associated with delivery of that reward presumably decrease significantly, whereas the risks inherent in waiting for the LL reward likely change little. Thus, risk aversion alone could account for the choice of a smaller, imminently-available, essentially “riskless” reward over a larger, delayed, and, hence, more risky reward. Furthermore, presumably an agent's subjective perceptions about the size of the LL reward, the likelihood that it materialises, and whether they can wait for it, also influences choice. With a tangible, riskless, and imminently-available SS reward, one can explain why this may be chosen over a delayed, risky, and inherently-uncertain LL reward without invoking hyperbolic discounting as the mechanism.

In lab experiments, risks associated with delivery of SS and LL rewards can be minimised by emphasising credibility of payment, and through (clever) experimental design, such as the use of a front-end delay to the SS reward so that the perceived risk of the SS and LL rewards is similar (Coller & Williams, Reference Coller and Williams1999). One can also ground beliefs about reward magnitudes and days of receipt by simply informing subjects of these reward attributes. Thus, in the lab one could argue that it is possible to focus purely on time preferences without worrying about risk attitudes or subjective beliefs. But even this conclusion is premature because it is crucial to incorporate the curvature of the utility function when estimating time preferences to draw valid inferences about discounting behaviour (Andersen, Harrison, Lau, & Rutström, Reference Andersen, Harrison, Lau and Rutström2008; Harrison, Lau, & Rutström, Reference Harrison, Lau and Rutström2010). Thus, even in lab settings, focussing purely on choices between SS and LL rewards is problematic.

In the world outside the lab, most choices involve risk and delayed rewards are inherently uncertain. For example, in any real exploration–exploitation trade-off, an agent must choose between harvesting an available resource with an unknown, but sampled, distribution of outcomes, and exploring, which entails deferring consumption, in the hope of finding a more profitable distribution of outcomes. Thus, risk, time, and beliefs are inextricably intertwined.

Ainslie's primary focus is not on impulses, but rather impulse control or willpower. He argues that by bundling a series of LL rewards agents can control impulses. But the implicit assumption underlying the summation of hyperbolically discounted rewards is that the intertemporal utility function is additively separable. As Richard (Reference Richard1975) shows, when additive-separability does not hold decision makers are either intertemporally risk seeking or intertemporally risk averse, which affects consumption decisions through time. Indeed, the vast majority of economic models of habit formation assume intertemporal risk-seeking behaviour. This generates intertemporal complementarities in consumption that can explain choices yielding immediate benefits but entailing long-term negative consequences. It is worthwhile investigating, therefore, how different intertemporal utility function assumptions affect Ainslie's interpretation of willpower.

In sum, Ainslie provides a masterful review of different but overlapping literatures, and a reconciliation of the misunderstood and often conflated notions of suppression, resolve, and habit. But one must insist that the theoretical framework be completed by incorporating the risk and uncertainty that are intrinsic to Ainslie's concept of willpower.