Schillbach et al. make a strong case that studying human social behaviour requires more than our traditional “isolation paradigms.” They detail a number of methods which will allow the neuroscientists of the future to study social interaction in a more meaningful way. They also suggest that new theoretical approaches are needed, going beyond traditional cognitive theories. However, they dedicate much less space to specifying what these new theories should be. In this commentary, I would like to reiterate the need for strong theories to drive forward the field of second-person neuroscience. Such theories must be clearly specified so that other researchers can test them, and must be falsifiable because theories that accept all data do not make strong predictions.

Here I consider three possible categories of theories drawn from different research traditions: computational models, dynamical systems, and cognitive models. I suggest that cognitive/information-processing models remain the most promising, but that all three approaches may be able to converge.

First, advances in the non-social domains of reinforcement learning and motor control have led to computational models of social interaction. Such models often use tasks or games derived from game-theory in which people interact within a set of rules. Impressive localisation of specific computational parameters in the brain has been possible with these models (Behrens et al. Reference Behrens, Hunt and Rushworth2009; Hampton et al. Reference Hampton, Bossaerts and O'Doherty2008; Tomlin et al. Reference Tomlin, Kayali, King-Casas, Anen, Camerer, Quartz and Montague2006). Other related approaches include adapting motor control models to control not just physical objects (e.g., a tennis racket), but also social objects (e.g., another person) (Wolpert et al. Reference Wolpert, Doya and Kawato2003). These computational approaches are very powerful in the cases where the model can be specified. However, a current limitation is that these models are only applied to abstract, rule-bound contexts (e.g., iterated trust games). Such tasks are somewhat artificial and lack many of the behavioural cues (eye gaze, emotion, etc.) of real social interactions.

Second, research derived from ecological psychology and dynamical systems has led to an enactive approach to social neuroscience (Thompson Reference Thompson2007). This approach rejects traditional cognitive models, together with ideas of symbolic information processing and representation. Instead, it relies on dynamical systems (Port & van Gelder Reference Port and van Gelder1995; Thelen & Smith Reference Thelen and Smith1996). A key idea is that social cognition exists in the interactions between agents rather than in the information processing within the head of a single agent (De Jaegher et al. Reference De Jaegher, Di Paolo and Gallagher2010). The target article endorses these approaches, and they seem particularly useful in thinking about infant development. However, there seems to be some tension between the claims of the strong enactivist models, and the neuroimaging method, which remains routed in studying activation within one brain at a time. This is clear in Figure 1 of the target article, which assumes localised information processing systems within each brain as the originators of the dynamic social interaction.

These diagrams of brain systems seem more compatible with a third approach, that of embodied information processing models. This type of model emphasises the overlap of motor, proprioceptive, linguistic and affective information processing streams (Prinz Reference Prinz and Hurley2005). It thus rejects the strict modularisation of Fodor but retains the cognitive idea that the brain is an information processing device and that we can localise specific types of processing to specific areas of cortex. Such information processing models have in the past been used to describe performance on “spectatorial” tasks and in contexts without dynamics. However, this does not mean that they should always be used in this way.

An example of a more socially engaged information processing model is the STORM (social top-down response modulation) model (Wang & Hamilton Reference Wang and Hamilton2012). This is based on the idea the human brain contains a visuomotor processing stream (Cisek & Kalaska Reference Cisek and Kalaska2010) embedded within parietal and premotor cortex. This visuomotor stream is called the perception–behaviour expressway in social psychology (Dijksterhuis & Bargh Reference Dijksterhuis and Bargh2001) and is also present in ideomotor theories of action (Prinz Reference Prinz and Hurley2005). It is likely to incorporate the “human mirror neuron system” and its behaviour is almost certainly determined by associative learning over the lifetime (Heyes Reference Heyes2011). The central claim of the STORM theory is that information processing within the visuomotor stream is subtly and dynamically modulated by other social brain systems. For example, the tendency to spontaneously mimic other people is rapidly modulated by eye contact signals (Wang et al. Reference Wang, Newport and Hamilton2010), and this modulation is implemented by processing in medial prefrontal cortex and enhanced connections from medial prefrontal cortex to superior temporal sulcus (STS) (Wang et al. Reference Wang, Ramsey and Hamilton2011). Thus, key regions of the social brain (medial prefrontal cortex, mPFC) regulate information processing in the visuomotor stream (superior temporal sulcus, STS). In other contexts, other social brain systems such as those linked to reward and motivational processing are also likely to modulate the basic visuomotor stream. The key predictions of this model are that during dynamic interactions top-down influences on the visuomotor stream should be clearly visible. However, upward information flow from motor systems to mentalising systems (as suggested by simulation models; Gallese & Sinigaglia Reference Gallese and Sinigaglia2011) should be less prominent. Such predictions can be tested behaviourally or by means of neuroimaging combined with dynamic causal modelling (Friston et al. Reference Friston, Harrison and Penny2003).

It is not yet clear which of these three quite different categories of theory will provide the clearest answers to core questions about human social interaction and the functioning of the social brain. Here, I make the case for using information processing models in second-person neuroscience. Just because information processing approaches have sometimes been “spectatorial” or have lacked dynamics does not mean that they should always be this way. These models can be closely linked to brain activation, can be applied across a large number of contexts including psychiatric disorders, and are testable using methods like dynamic causal modelling.

Whichever type of theory ultimately prevails, having published theories will help to drive the field forward. In the new realm of interaction neuroscience, strong theories and clear models will help us choose which of the many possible experiments are worth pursuing, and will make our work more than just butterfly collecting.