Tradeoffs between accuracy and parsimony are inherent in most scientific endeavors. Murayama and Jach (M&J) argue that contemporary motivation theories, which operate nearly exclusively at the construct level, are making the wrong tradeoffs between accuracy and parsimony. They challenge the idea that motivation constructs directly cause behavior. Instead, they argue (1) constructs are essentially epiphenomenal byproducts, and (2) process-oriented computational models are necessary to unpack mechanisms of motivated behavior and advance the field of motivation science.

We agree with the first premise. Motivation constructs are not literal causes of behavior; researchers will never find a need-for-competence dial in the brain (at least, not in the way that latent variable models assume). Rather, the need-for-competence and other motivational constructs act as labels that summarize patterns emerging from yet to be defined processes. M&J argue that the best path forward is to begin attempting to “unpack the black box” and define these processes. Their proposed solution of computational modeling, however, shares many of the same faults as the construct-level approach they critique; they simply trade one black box for another. Where construct-focused approaches assume a need-for-competence drives behavior, their exemplar knowledge acquisition model assumes an intrinsic-reward-for-knowledge drives behavior. Importantly, we could ask the same question of the reward-for-knowledge as M&J ask of the need-for-competence: What process creates this drive? As acknowledged by the authors themselves, their suggestion to move down a level of analysis does not solve the black-box problem. It merely changes the black boxes we use.

Because neither construct nor computational models address the black-box problem, we need an alternative way to evaluate between them. One might even argue that the black-box language obfuscates the real point: That the best measure of a model is its accuracy and that process-oriented models provide a way toward greater accuracy. We disagree with the former notion. The question of which level of mental process is best to model is not a question of “Which level is most accurate?” Of course, the answer to that will always be the next level down.

Rather, we believe it better to ask “Which level is most useful?” Given our background in applied psychology, and specifically our experiences training pre-service teachers, we know that one of the ultimate goals of motivation theory is to generate insights with practical importance for teachers, students, bosses, workers, and so on. These experiences have led us to adopt a more pragmatist philosophy of science, wherein a key feature of any worthwhile theory is that it can be used to make an impact (Elder-Vass, Reference Elder-Vass2022; James, Reference James and Delbanco1907/2001). History tells us this is where computational models struggle.

For example, educational researchers including Carroll (Reference Carroll1963) and Bloom (Reference Bloom1976) made a strong push for computational and process-oriented “models of school learning” in the mid-twentieth century (see also Bjork, Reference Bjork1973). However, those models were difficult to understand, even for researchers. Consequently, these theories received minimal adoption and faded in importance (Harnischfeger & Wiley, Reference Harnischfeger and Wiley1978). More recently, researchers in self-regulated learning – ourselves included (Schuetze, Reference Schuetze2024) – have put forth a number of non-construct-focused models based on discrepancy reduction (e.g., Ackerman, Reference Ackerman2014; Carver & Scheier, Reference Carver and Scheier1990; Thiede & Dunlosky, Reference Thiede and Dunlosky1999). Many researchers have testified to the benefits of building these sorts of models for the purpose of theory development (e.g., Aubé, Reference Aubé1997; Guest & Martin, Reference Guest and Martin2021; van Rooij & Blokpoel, Reference van Rooij and Blokpoel2020). However, due to their complexity and relatively narrow areas of focus, these process-oriented models have struggled to make the same impact on school systems and business leaders as construct-focused understandings of human behavior, such as self-efficacy and growth mindset.

Our contention here is that even if we create highly accurate theories of motivated behavior, if they are not usable or interpretable by those who are in positions to apply them, more needs to be done. Applied to M&J's proposal, we believe that computational models of motivation can be useful to the world at large – but this will require additional work. Part of this work may mean finding creative places to implement motivational theories, such as in intelligent tutors, where theoretical complexity is managed by a technical system, as opposed to by teachers or managers (Yan, Sana, & Carvalho, Reference Yan, Sana and Carvalho2024). Other parts of this work may mean creating a hierarchy of mutually compatible theories operating at different levels of analysis. Insights derived from lower levels can be distilled and moved up to higher (perhaps construct) levels that require less time and expertise to put into practice (Anderson, Reference Anderson2002; Donoghue & Horvath, Reference Donoghue and Horvath2016). In essence, the researcher's theory of motivation doesn't necessarily need to be the same as the practitioner's. Different groups can understand the same phenomenon in distinct levels of detail. Indeed, divisions of this sort help molecular biologists, pharmacists, and medical doctors create and administer medical treatments that work despite having very different levels of focus. Similarly, the average driver does not need to understand how their car's engine functions. The mechanic, however, must. Again, theories at different levels should aim not for the utmost accuracy, but for the utmost utility to those who are using it.

With this in mind, we find ourselves agreeing with M&J's second premise as well: That computational and process-oriented models can help us better understand motivation. Not because they help us solve the black-box problem, but because they show promise in helping us make more useful and applicable theories of motivated behavior. However, pitfalls accompany this promise, and they must be kept in mind. Given the history of computational modeling and the associated increases in time, effort, and expertise required to use these models, we believe that great care will be needed to translate them into practically applicable formats. Until we achieve broadly usable computational models, we cannot fault motivation researchers or practitioners for sticking to tried-and-true construct-oriented models. Pending that development, the construct level of analysis will continue to be the primary lever of change outside the lab.