Veissière et al. convincingly argue that we collectively build niches and cultural practices, which guide our attention towards salient information, facilitating cultural learning, and the acquisition of shared expectations about norms and conventions. This allows us to “acquire culture by being immersed in specific, culturally patterned practices that modulate salience.”

Here, we consider that not only we guide each other's attention, prediction, and learning processes towards salient information at a long time-scale (e.g., niche construction); but also at a faster time-scale (e.g., during teaching and joint action).

In active inference, the salience of stimuli depends on both their quality and on the agent's belief about the world (Parr & Friston Reference Parr and Friston2017a). Salient stimuli are those that are expected to change the agent's belief, such as those about which the agent is uncertain, but (once gathered) would clearly disambiguate the agent's alternative hypotheses. Conversely, stimuli that were predicted, are of poor quality or ambiguous (and if gathered, would not disambiguate the agent's hypotheses) have little salience, as they would not change the agent's belief significantly.

By directing our information-gathering actions (e.g., saccades) to high-salience locations or stimuli, we gain the most (in information terms) from our engagement with the environment. Information-gathering actions are sometimes called epistemic actions, as they aim at changing one's contextual beliefs (e.g., when exiting from an unknown underground parking, looking around to resolve uncertainty about one's current location); and are distinguished from pragmatic actions, which aim to achieve goals (e.g., drive to an intended destination, after having resolved the above contextual uncertainty) (Friston et al. Reference Friston, Rigoli, Ognibene, Mathys, Fitzgerald and Pezzulo2015; Reference Friston, Redish and Gordon2017; Pezzulo et al. Reference Pezzulo, Rigoli and Friston2015; Reference Pezzulo, Donnarumma, Dindo, D'Ausilio, Konvalinka and Castelfranchi2018).

Crucially, during social interactions, we can perform epistemic actions for the sake of others – or social epistemic actions. These actions aim at gathering salient (belief-changing) information for others and hence are quintessentially communicative. However, they don't need to be verbal, but can exploit sensorimotor channels, that is, sensorimotor communication.

An extensive body of research shows that during social interactions, we modify our behavior in communicative ways, to render it more “legible,” easier to understand and to predict by others (Pezzulo et al. Reference Pezzulo, Donnarumma, Dindo, D'Ausilio, Konvalinka and Castelfranchi2018). A well-known example is that of a mother who amplifies her speech (e.g., the vowels) and exaggerates her bodily movements, when interacting with her child, that is, motherese and motionese. Such sensorimotor communication may simultaneously help capture the child's attention and simplify her learning task (e.g., by stressing what is salient).

Sensorimotor communication is ubiquitous during social interaction. It does not necessarily require “specialized” behaviors, such as pointing or gazing at some object. Rather, virtually any action can be used (or modified) to convey sophisticated communicative messages. Even simple actions, such as passing a glass to somebody can express (intentionally or unintentionally) love, hate, or deference. These (“hidden”) emotional states can be conveyed by – and inferred from – subtle kinematic cues, for example, slightly faster or slower arm movements (Becchio et al. Reference Becchio, Manera, Sartori, Cavallo and Castiello2012; Pezzulo et al. Reference Pezzulo, Donnarumma and Dindo2013).

Sensorimotor communication is especially effective during joint actions. For example, during a joint action as simple as moving a table together, we can push the table to signal in which direction we want to go or where we want to place the table. During more complex social interactions, such as during a soccer match, we can exaggerate our movements to help our teammates inferring our intention (e.g., where we want to pass the ball), or hide it from opponents (by feinting). Sensorimotor communication goes also beyond the body. For example, when driving a car, we can decelerate or move to the side of a road, to signal that we want to leave room to other drivers (Chater et al. Reference Chater, Misyak, Watson, Griffiths and Mouzakitis2018; McEllin et al. Reference McEllin, Sebanz and Knoblich2018; Pezzulo & Dindo Reference Pezzulo and Dindo2011; Vesper et al. Reference Vesper, van der Wel, Knoblich and Sebanz2011).

These examples can be conceptualized within active inference, as social epistemic actions that unveil salient information, for the sake of somebody else. Consider the case of a person, who is helping his roommate to move a table, but does not know whether she wants to place it to the left or the right of a chair. He can engage in mindreading, to infer the roommate's intention (“hidden state”) based on her movements (“observables”). Simultaneously, the roommate can select a plan that is maximally informative about her intentions (e.g., push to the left earlier and harder) – an example of social epistemic action, to optimize interactive success. Interestingly, the roommate helps her coactor inferring her intention, by “surprising” him, because her “exaggerated” force deviates from the most likely (predicted) action plan. It is this unpredicted (and informative) deviation that has high salience and signals the teammate's communicative intention to place the table to the left (Pezzulo Reference Pezzulo2011; Pezzulo et al. Reference Pezzulo, Donnarumma and Dindo2013). Importantly, not all unpredictable behaviors are salient, but (like standard epistemic actions) only those that resolve the coactor's uncertainty. Hence, selection of effective social epistemic actions requires tailoring them to the current interactive situation. Recent research showed that coactors consider elements such as others’ uncertainty, and what they can or cannot see, to modulate their social epistemic actions. Furthermore, social epistemic actions can be bidirectional, with coactors continuously helping each other inferring their intentions and predicting their actions (Leibfried et al. Reference Leibfried, Grau-Moya and Braun2015; Pezzulo & Dindo Reference Pezzulo and Dindo2011; Pezzulo et al. Reference Pezzulo, Iodice, Donnarumma, Dindo and Knoblich2017; Vesper & Richardson Reference Vesper and Richardson2014).

In sum, we flexibly use social epistemic actions to drive others’ attention on patterns and regularities that we want them to infer (or learn). Social epistemic actions may thus work synergistically with “culturally-patterned practices” to afford social and cultural learning. Of note, although social epistemic actions can be realized verbally, here we focused on sensorimotor realizations, which (in certain conditions) may signal more directly what and where salient information is. Furthermore, humans may be exquisitely sensible to recognizing under which conditions certain actions or demonstrations are executed for pedagogical purposes, and hence learn faster and more efficiently under these conditions (Csibra & Gergely Reference Csibra and Gergely2011).