The target article sketches the promise of combining rational principles and cognitive constraints to understand human cognition, and singles out linguistics as one domain for work along those lines. While it touches on aspects of language rooted in individual cognition like the principle of least effort (Lestrade Reference Lestrade2017; Zipf Reference Zipf1949), I want to probe the limits of the resource-rational framework by looking beyond individual minds to interactive language use, the primary ecology of human cognition (Böckler et al. Reference Böckler, Knoblich, Sebanz, Glatzeder, Goel and Müller2010; Waldron & Cegala Reference Waldron and Cegala1992). Here, under the relentless pressures of rapid-fire turn-taking (Levinson Reference Levinson2016) and always-on inferential processes (Enfield Reference Enfield2013; Goffman Reference Goffman1967), language provides a window onto how social rational agents deal with resource limitations in a noisy and uncertain environment.

Human language provides ample evidence of adaptation to capacity limits in social interaction (Roberts & Levinson Reference Roberts and Levinson2017). Articulation, relatively slow compared to processes of formulation and interpretation, forms a significant bottleneck in human communication that we can bypass thanks to pragmatic inference (Levinson Reference Levinson2000): any content that can be left to inference need not be explicitly articulated. This puts a premium on computable and efficient heuristics for formulation and interpretation (Frank & Goodman Reference Frank and Goodman2012; Van Rooij et al. Reference Van Rooij, Kwisthout, Blokpoel, Szymanik, Wareham and Toni2011). But as Lieder and Griffiths argue, people cope with computational complexity through heuristics as well as through habits. One way to think of language is as offering a culturally evolved store of habits – routinely deployable resources – that help outsource computation and streamline coordination (Clark Reference Clark, Carruthers and Boucher1998; Kempson et al. Reference Kempson, Cann, Gregoromichelaki and Chatzikyriakidis2016).

A resource-rational approach may be especially promising for understanding the ubiquity of delay markers, continuers, and repair strategies, which easily occur in up to one in five utterances (Enfield Reference Enfield2017; Fox Tree Reference Fox Tree1995). Whereas classic linguistic work has assumed such items are grammatically irrelevant (Chomsky Reference Chomsky1965) or at most symptoms of trouble (Levelt Reference Levelt1989), resource-rationality makes it possible to account for them as optimally adaptive interactional tools (Dingemanse Reference Dingemanse and Enfield2017): cognitive crutches that help optimize complex rational communication under resource limitations. For instance, delay markers like “um” help word recognition by alerting the recipient that an upcoming word might need more attention (Fox Tree Reference Fox Tree2001), and repair initiators like “huh?” or “who?” allow us to gracefully recover from impending communicative trouble, something that happens, on average, at least every 84 s in conversation (Dingemanse et al. Reference Dingemanse, Roberts, Baranova, Blythe, Drew, Floyd, Gisladottir, Kendrick, Levinson, Manrique, Rossi and Enfield2015). With interactional tools available at every turn to review, revise, and recalibrate understanding, the dynamics of human cognition in interaction diverges radically from the one-shot models assumed in many current theories.

As a consequence, interactive language use calls into question the exclusive focus of rational analysis on individual minds. Are resource-rational approaches limited to individual cognition or could they extend to socially distributed cognition? By enabling the redistribution of attentional, cognitive, and embodied resources (Clark Reference Clark2006; Hutchins Reference Hutchins1995), interactive language use alleviates individual-bound capacity limits and can optimize performance beyond the bounds of idealized one-shot communication: an interactively scaffolded form of cognitive offloading (Risko & Gilbert Reference Risko and Gilbert2016). The sheer frequency of the interactional tools mentioned above shows how much communication relies on this form of scaffolding (Fusaroli et al. Reference Fusaroli, Tylén, Garly, Steensig, Christiansen, Dingemanse, Gunzelmann, Howes, Tenbrink and Davelaar2017). This radically increases the error-tolerance and flexibility of cognition in interaction. It also creates opportunities to study the workings of resource-rationality in the relatively controlled environment of well-understood sequential patterns of interaction.

Communicating under noise and uncertainty requires constant cost-benefit analyses of formulating a response versus issuing a request for repair, factoring in the relative costs of different repair formats and their possible downstream consequences, all under severe time pressure and with limited cognitive resources. A systematic comparison of repair across languages and cultures shows that people everywhere deploy the repair system in efficient ways that minimize cost for the dyad as a social unit, rather than just for themselves as individual-based rational approaches might suggest (Dingemanse et al. Reference Dingemanse, Roberts, Baranova, Blythe, Drew, Floyd, Gisladottir, Kendrick, Levinson, Manrique, Rossi and Enfield2015): an optimal use of distributed cognitive resources. A similar interactive, distributed perspective is required to make sense of information-theoretical results about word meanings and ambiguity (Piantadosi et al. Reference Piantadosi, Tily and Gibson2012): we can cope with ambiguity in communication only to the extent that one mind picks up the slack where the other leaves off. This means that resource-rational analysis of human cognition will need to deal not just with individual minds, but with interacting minds operating in an environment of culturally evolved metacognitive resources.

Recent work in cognitive science and cultural evolution is revisiting the Vygotskyan insight that human cognition is greatly amplified by culturally evolved pieces of cognitive equipment (Bender & Beller Reference Bender and Beller2014; Clark Reference Clark2006; Heyes Reference Heyes2018). At the same time, neuroscience is increasingly concerned with understanding brain and language in the context of social interaction (Hirsch et al. Reference Hirsch, Adam Noah, Zhang, Dravida and Ono2018; Konvalinka & Roepstorff Reference Konvalinka and Roepstorff2012; Schilbach et al. Reference Schilbach, Timmermans, Reddy, Costall, Bente, Schlicht and Vogeley2013). One thing that unites these approaches is their attention to how the picture of cognitive demands and resources may change radically as a result of interactionally scaffolded, socially augmented cognition. Lieder and Griffiths do not discuss cultural evolution and social interaction as part of the environment in which heuristics and habits can be honed to become optimally adaptive, and it is unclear whether they intend resource-rational analysis to include the kinds of interactional resources discussed here: material symbols of metacognition that augment and distribute our cognitive processes. Perhaps this is the next frontier.

In sum, I applaud the call for new ways to connect psychological theory and the cognitive sciences, and would like to put forward interactive language use as a challenging yet promising domain for resource-rational approaches. As the primary ecology of human cognition, social interaction provides a rich natural laboratory for probing the leverage and limits of resource-rational analysis. Future work in this vein might focus not just on how structural aspects of language adapt to the resource limitations of individual minds, but also on how every language offers its own compendium of culturally evolved ways by which people transcend individual resource limitations and benefit from distributed cognition.