Speaking and listening to inter-brain relationships

Studies of inter-brain relationships thrive, and yet many reservations regarding their scope and interpretation of these phenomena have been raised by the scientiﬁc community. It is thus essential to establish common ground on methodological and conceptual deﬁnitions related to this topic and to open debate about any remaining points of uncertainty. We here offer insights to improve the conceptual clarity and empirical standards offered by social neuroscience studies of inter-personal interaction using hyperscanning with a particular focus on verbal communication. © 2022 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).


Introduction
Studies investigating the physiological basis of social interaction increasingly propose a multi-brain approach to data collection. They extend model-based neuroimaging to the twoperson situation in order to provide a mechanistic account of social interaction (e.g., Bilek et al., 2022;Bolis & Schilbach, 2017;Henco & Schilbach, 2021). This approach has been promoted by conceptual frameworks emphasising the importance of interactive activities for the emergence and implementation of social cognition (Di Paolo & De Jaegher, 2012;Hari & Kujala, 2009;Hasson, Ghazanfar, Galantucci, Garrod, & Keysers, 2012;Schilbach et al., 2013;Stolk et al., 2014). Two-person neuroscience (2PN) (Hari, Henriksson, Malinen, & Parkkonen, 2015) and second-person neuroscience (Redcay & Schilbach, 2019) are examples of conceptual and methodological frameworks that prioritise simultaneous neural recordings from two or more interacting individuals (i.e., hyperscanning, Fig. 1A). The underlying idea for implementing hyperscanning setups is to be able to evoke behaviours and mental states that are contingent on the partner's behaviour. That creates the conditions for studying how mutually contingent behaviours are generated and controlled. Which is done by analysing the relationship between neural signals recorded simultaneously from multiple individuals (Montague et al., 2002). A recurrent finding from hyperscanning studies is enhanced between-brain covariance during coordinated activities (for review, Czeszumski et al., 2020) as compared to dissimilar control conditions. This enhancement is typically referred as 'inter-brain synchronisation'. In this context, inter-brain synchronisation is claimed as a phenomenon uniquely associated with social interactions that can only be observed during hyperscanning.
However, there are timely concerns about interpreting research investigating inter-brain synchronisation in social neuroscience (Astolfi et al., 2020;Hamilton, 2021;Hari, Himberg, Nummenmaa, Hamalainen, & Parkkonen, 2013). A recent critique suggests that work in the field has departed from well-understood paradigms and rigorous data interpretation (Holroyd, 2022). The main issues raised are interwoven: a shallow theoretical corpus (e.g., definitions, theories), accompanied by incomplete experimental designs (e.g., tasks, control conditions, statistics), resulting in misleading interpretations (e.g., bold or unsubstantiated claims). With 'inter-brain synchrony' being the driving force for many research programs, it is worth paying attention to these concerns. The current paper represents a proactive effort to improve the conceptual clarity and empirical standards needed to advance in the field. We here offer some insights about terminology, definitions and present ideas that could be used in further research testing mechanistic accounts of increased inter-brain similarity. The elaborations here focus on the temporal aspect of inter-brain dynamics. We emphasise verbal interactions as a unique opportunity for new methodological developments and testing current theories about inter-brain relationships.

2.
On the term inter-brain synchronisation The term 'inter-brain synchronisation' can evoke the intuitive idea of individuals rhythmically sampling the environment at precisely the same time points and the emergence (or reinforcement) of a link between individuals during action. These intuitions are correct to some extent since the Oxford English Dictionary defines synchronise as "to occur at the same time", and synchronisation is a common indication of a link between biological systems (Strogatz, 2003). However, depending on the author, the term synchronisation can have different meanings and interpretations (Ravignani, 2017). Furthermore, the same holds for the related terms of coupling (Dumas & Fairhurst, 2021) and entrainment (Bittman, 2020). We believe that agreement on one definition of synchronisation will avoid confusion and misinterpretation. A scientific definition of synchronisation phenomenon consists in the adjustment of rhythms from independent oscillators due to an interaction between these oscillators (Pikovsky, Rosenblum, & Kurths, 2001). A consequence of adopting this physical/mathematical conceptualisation is that using the term inter-brain synchronisation should be supported by an indication that self-sustained neural oscillations are observed in both of the interacting individuals (e.g., verbal interactions should be supported by separate, endogenous neural oscillations in both the speaker and the listener that achieve synchronisation during conversation). Although endogenous rhythms could be a key underlying principle for speech perception (van Bree, Sohoglu, Davis, & Zoefel, 2021), empirical criteria for demonstrating endogenous oscillations are often unsatisfied in practice (Obleser & Kayser, 2019) and, to our knowledge, have not thus far been shown for conversation.
Moreover, whereas synchronisation of oscillators should be assessed by measuring frequency and phase parameters (Pikovsky et al., 2001), other measures can be used to assess temporal contingencies between signals from different brains. For example, measures from information theory like mutual information (MI) and transfer entropy do not require phase consistency between biological systems and hence expand the space of meaningful neural similarities during an interaction. Recent comparisons suggest that MI measures (e.g., Ince et al., 2017) are more flexible and sensitive when measuring the coupling of central and peripheral neural signals (Gross et al., 2021). They might similarly be better suited to assessing inter-individual similarity in brain activity in general and between speaker and listener neural responses in particular (Hasson & Frith, 2016).
Another crucial related observation is that synchrony measures are unable to encompass the whole spectrum of interactions. Pattern switching is an essential foundation for coordination in living things (Kelso, 1995). Moreover, during intention-based cooperation, the patterns of alignment go beyond mirroring (imitation) of the other's actions. Social interaction requires implementing complementary actions and achieving synergies (Hasson & Frith, 2016). This is especially relevant in the case of dialogue, where interlocutors Hyperscanning (Panel A) implies true inter-personal interaction and simultaneous brain recordings. Pseudo-hyperscanning (Panel B) is a non-interactive condition in which verbal intervention stimulation is pre-recorded and later presented to the corresponding partner. Then, the different brain signals are aligned off-line using the common physical signal. The following statistical analysis will comprise two steps, (i) comparing the hyperscanning and pseudo-hyperscanning against surrogate data to find effects beyond chance levels and (ii) directly comparing hyperscanning and pseudo-hyperscanning. This design allows a test for the exclusive emergence of inter-brain relationships associated with social interactions. The figure adapts the original version in Hari et al. (2015). c o r t e x 1 5 9 ( 2 0 2 3 ) 5 4 e6 3 continuously cooperate and adapt to one another. During conversation, these complementary patterns of behaviour are better predictors of conversational performance (Fusaroli & Tylen, 2016), as measured using recurrence quantification analyses (Marwan, Carmen Romano, Thiel, & Kurths, 2007). In other words, phase synchrony analyses are limited for capturing complementary dynamics, and a successful interaction (e.g., good coordination or effective conversation) does not equate to pure synchronisation.
Recently, Zhang, Rose, and Yartsev (2022) used animal models (bats) to propose a computational model explaining how inter-brain activity patterns and dynamic social interactions co-evolve and feedback on each other. Interestingly, their model focused on the differences between brains and how that difference relates to the shared neural activity pattern . They use the term "inter-brain relationships", which can be more widely used to describe a dynamic range of covariance patterns between neural activities.
In a nutshell, beyond the conceptual ambiguity and current over-use of the term 'inter-brain synchronisation', synchronisation is too specific a measure of the systematic relationships between activity in different interacting brains. A more inclusive (yet no less precise) term like "inter-brain relationships" is to be preferred. The inter-brain relationship can be defined as the statistically significant relationship between the neural activity of socially interacting individuals in a form that depends on the interaction.

Emergent property versus shared stimulation
Evidence suggests that inter-brain relationships are not exclusively driven by sensory stimuli and/or motor behaviours. There are non-trivial inter-brain relations driven by the generation of an interpretational context (Stolk et al., 2013(Stolk et al., , 2014. However, sometimes it is unclear from a study whether the emergence of a two-brain network is exclusively associated with social interaction. An inter-brain relationship can instead be a byproduct of time-locked responses to sensory events in a shared environment (Hari, Sams, & Nummenmaa, 2016). Generating a null distribution by (post-hoc) randomly pairing participants is a commonly used method for testing the hypothesis of enhanced inter-brain similarity in real interacting pairs. However, with this procedure we are eliminating the common sensory events generated or present during the interaction (e.g., vocal sounds and backchannel responses). We are generating a null distribution of interaction-related effects but also a null distribution of shared sensorial stimulation effects. In other words, inter-brain relationships effects obtained by comparing with random paired data distributions, might still be caused by shared sensory events. To rule out this possibility it is critical to compare data collected from hyperscanning recordings with, for example, simultaneous recordings of (i) noninteracting individuals without any explicit task, (ii) noninteracting individuals exposed to simultaneous identical stimulation and (iii) separated individuals interacting with a computer (or a decoy) that executes an identical sequence of actions. Another possibility is comparing hyperscanning and sequential brain imaging of two unidirectional communicating participants (Hari et al., 2015). This is a form of social communication but in a non-interactive situation. Pseudohyperscanning is one such form of sequential brain imaging in which the vocal (and manual, etc.) responses of one participant are used as stimuli for another participant (Fig. 1B). Then, neural signals during production and perception are aligned offline using the common physical signal in the two recordings as a reference. Given that the actions of the pre-recorded participant are not modified by the online actions of the subsequent participant, the use of pseudo-hyperscanning reduces or eliminates the action-perception loops that characterise social interactions. This is especially true in situations where social interaction unfolds rapidly, for example, during a conversation.
In the specific case of verbal communication, pseudohyperscanning could be performed by precise off-line alignment of the neural activity of a participant speaking and the neural signal of a different participant listening to a replay of the produced speech. Although a communicative act (i.e., meaning transmission) takes place, the listener's responses (either verbal or non-verbal) do not help shape the content of the speech. Simply put, there is no linguistic interaction, and we might thereby ask whether this manipulation is sufficient to remove inter-brain relationships. It should be noted that a situation closer to a real interaction could be achieved by inducing participants to think there is a real-time turn-taking verbal interaction (i.e., deceiving participants). To the degree that speech perception and speech production (and arguably, attention) remain intact while removing the interpersonal interaction that is only present in true hyperscanning, we can disentangle whether inter-brain relationships exclusively emerge from the interactive exchange during the conversation; this experimental comparison of hyperscanning and pseudo-hyperscanning is depicted in Fig. 1. It rules out shared environmental signals or other confounds as causing apparent inter-brain relationships.

3.1.
Auto-pseudo-hyperscanning proposal A further possibility related to pseudo-hyperscanning is aligning the neural activity of the same participant when speaking and when listening to the same shared speech signal. In other words, and unlike pseudo-hyperscanning, we can align neural signals corresponding to speech production and self-listening in a single participant. This particular type of 'inter-brain' analysis has potential shortcomings but also major advantages. We call this approach auto-pseudohyperscanning.
On the one hand, auto-pseudo-hyperscanning does not capture the (social) prediction dynamics that are still present (although to a lesser extent) in pseudo-hyperscanning. For example, starting to listen to an unfamiliar person may be associated with poor prediction. Then, throughout continuous listening, new information will accumulate, leading to the generation of a more accurate situation model (Zwaan & Radvansky, 1998) and more accurate expectations for the speaker. In contrast, during self-listening, the participants reexperience themselves. The situation model already exists, and it is not updated but retrieved from memory. The action can be described as a continuous memory recovery (access) c o r t e x 1 5 9 ( 2 0 2 3 ) 5 4 e6 3 with minimized surprise. Thus, it could be argued that our approach would only resemble periods of total convergence of beliefs within a successful linguistic exchange.
On the other hand, auto-pseudo-hyperscanning removes inter-subject variability inherent in pairing different participants and allows for further control of confounding factors known to influence inter-brain patterns in conversational dyads, like gender distribution (Baker et al., 2016) or emotional factors (e.g., empathy) (Smirnov et al., 2019). Thus, 'interbrain' measures from auto-pseudo-hyperscanning might provide an idealised inter-brain relationship. That is, in the absence of inter-subject variability and with 'perfect' linguistic alignment. Measurements derived from auto-pseudohyperscanning may provide a useful reference for comparing other, more 'real' or 'ecological' inter-brain relationships. Moreover, single subject level 'inter-brain' results can help to define which individual within a pair or group is driving an atypical inter-brain pattern. Perhaps most importantly, we can compare speaking and listening without stimulus differences and set up an 'inter-brain' situation in which accurate predictions for upcoming speech signals are possible and complete comprehension is expected (P erez et al., 2022). This 'perfect prediction' of upcoming speech constitutes an interesting feature for experiments testing the mutual prediction theory.

The mutual prediction theory
Inter-brain relationships have been typically framed within theories of joint action, which emphasise prediction as the critical element facilitating coordination (Sebanz, Bekkering, & Knoblich, 2006). Building on prediction, Frith (2015a, 2015b) conceptualise communication as involving two dynamic systems coupled via sensory information and operating according to principles of active inference by minimising prediction errors. In fact, prediction error computations play a critical role in perceiving social signals such as speech (Sohoglu & Davis, 2020) and faces (Apps & Tsakiris, 2013). The generalised synchrony model proposes an active inference framework for communication, suggesting that when an observer is modelling the behaviour of another person who is modelling the observer, a state of generalised synchrony can emerge (Friston & Frith, 2015a;2015b). Kingsbury et al. (2019) recently provided single neuron evidence that inter-brain relationships can result from mutual inference processes during social interactions in mice. They further elaborated these theories and formalised an explanation of the resulting inter-brain relationship. According to the mutual prediction theory, if individuals use the same neural mechanisms to predict others' actions and enact their own actions, social interaction or other situations involving joint prediction will inevitably lead to dynamic neural similarity within interacting pairs (Kingsbury et al., 2019).
Actually, in the case of verbal communication, the linguistic operations needed to understand and produce an utterance rely to a great extent on similar networks of the brain (Menenti, Gierhan, Segaert, & Hagoort, 2011). Yet, differences between the neural representation of self and others during interaction limit the degree to which exact inter-individual similarity in neural activity is expected. Moreover, explaining inter-brain relationships through mutual internal prediction states could be challenging in practice (also see: Rabagliati & Bemis, 2013). For example, in some types of social interactions like mothereinfant interactions, accurate prediction is more likely to be observed in the adult since the infant is learning to predict (Wass, Perapoch Amad o, & Ives, 2022). In general, reliable access to internal prediction processes is challenging. We argue that conversation denotes a unique opportunity.

4.1.
Conversation as a tool to test and refine mutual prediction theories Conversation is a type of joint action in which language plays the predominant role (Clark, 1996). Insofar as it is a form of social interaction whereby two (or more) interlocutors coordinate language in time (e.g., turn-taking) and space (e.g., head and body position so as to be audible) to create a change in the environment and in others (e.g., the mind of the interlocutor is altered). A distinctive feature of conversation is that it needs to accommodate long-range dependencies between participants' signals (e.g., shared conceptualisations). This feature is not usually considered in joint action settings, which are focused on near-instantaneous perceptioneaction relations. Thus, in the context of studying human inter-personal relations, it is worth emphasising that language is only one among a series of tools that humans use to communicate referentially (Levinson, 2022). Conversation (where interlocutors are committed to communicative success) constitutes a joint action that contributes to interpersonal coordination by using the communicative capacity of language.
Pickering and Garrod (2021) emphasise the predictive nature of interactive language use. They have recently integrated the theory of prediction and joint action with their interactive linguistic alignment (ILA) theory (see: Pickering & Garrod, 2021, for the definition of alignment). According to ILA theory, during verbal communication, production and comprehension processes become aligned at different levels: phonetic, phonological, lexical, syntactic and semantic (Pickering & Garrod, 2004). We hypothesise that mutual alignment and predictability across the many levels of the linguistic hierarchy can be associated with inter-brain relationships during linguistic interactions.
Because language is not a succession of disconnected words, natural language processing models (NLP) using machine learning methods can generate predictions about upcoming words. Recent work has shown how probabilistic predictions for upcoming words (and speech sounds) can be used to predict neural activity during speech comprehension (e.g., Armeni, Willems, van den Bosch, & Schoffelen, 2019; Donhauser & Baillet, 2020; Koskinen, Kurimo, Gross, Hyv€ arinen, & Hari, 2020;Schrimpf et al., 2021). Furthermore, since a conversation is not a succession of disconnected remarks, NLP models can generate the same predictions while also considering the interlocutor's identity and mental state. NLP models that include the content of previous speaker interventions to generate predictions are increasingly being applied (See, Roller, Kiela, & Weston, 2019). Within a few c o r t e x 1 5 9 ( 2 0 2 3 ) 5 4 e6 3 words of the start of a conversation, it is possible to obtain dynamic prediction measures of verbal interaction. This opportunity for prediction quantification as a hallmark of a neural process is unique and makes it possible to test whether inter-brain relationships are associated with successful predictions. See Fig. 2 for a schematic representation of an experimental test of this idea.
Interestingly, the neural configuration and timing of each specific linguistic feature may depend on the language system or code used. Therefore, different linguistic interaction types can yield different patterns of inter-brain relationships. In line with this, we have shown that verbal communications in native and foreign languages are accompanied by distinct topological/topographical patterns of enhanced brain similarities (P erez, Dumas, Karadag, and Duñabeitia, 2019).
In a nutshell, we propose to explore the mutual prediction theory using linguistic interactions. The rationale is as follows: since there is a statistical regularity in speech, we can discretise its predictability using information theoretic measures (see, Gwilliams & Davis, 2022). This speech predictability structure can be estimated at the level of phonemes, morphemes, words, syntactic structures, and semantic content. Moreover, the neural responses to (continuous) speech can be modelled as a function of these measures (e.g., phoneme surprisal, word embeddings). Thus, we can explore whether there is a relationship between speech predictability structure and the inter-brain relationship across different linguistic levels (Nastase, Gazzola, Hasson, & Keysers, 2019). This association will add to the model of inter-brain relationships as reflecting predictive processing while describing the contribution of different linguistic features in the prediction.
How the signals from different brains are aligned to show reliable inter-brain relationships varies across studies. Sometimes enhanced inter-brain similarity is shown using lagged delays between the two different brain signals. In other studies, the effect is instantaneous (i.e., with zero-lagged timing). Nonetheless, the timing of alignment that shows enhanced inter-brain similarity has conceptual implications. For example, showing an effect in which one brain activity consistently precedes the other, may be meaningful in interpreting leader-follower dynamics during joint actions. However, lagged delays can also be interpreted as a result of the neural alignment between action implementation/execution in one individual and action perception in the other, which could still be present in non-interactive situations. In other words, for interpreting in the context of the mutual prediction theory, inter-brain relationships should not be reducible to Fig. 2 e Schematic of an experimental design testing mutual prediction theory with a conversational setting. Top: Audio signal containing natural speech and text transcription. Bottom: EEG time-series from two participants (only one channel for simplicity) from which inter-brain relationship can be obtained. Middle: Output from the Natural Language Processing model (e.g., quantification of word prediction or meaning similarity) is compared with the time series of inter-brain relationship. The figure adapts the original version in Koskinen et al. (2020). Example dialog taken from Weizenbaum (1966).
perceptual information about the other or motor contingencies as expressed in the lagged misalignments.
In the particular case of verbal communication, we have supported the idea of instantaneous inter-brain relationships between the speaker and the listener. First, we provided evidence of zero-lagged enhanced inter-brain similarities during a turn-taking conversation (P erez, Carreiras, & Duñabeitia, 2017). Second, we provided evidence of differential timing for the engagement to the speech audio (namely, speech tracking) depending on the conversational role (P erez et al., 2022). Specifically, we described a timeline for speech tracking in speakers and listeners with maximal speech tracking after the auditory presentation during perception (approximately 110 ms) and before vocalisation during speech production (approximately 25 ms). Based on these delays, significant inter-brain similarities obtained by shifting the listener's brain time series to lag behind that of the speaker by approximately 135 ms can be mediated by joint processing of the auditory envelope in speech perception and production. Following this rationale, we predict that the alignment between the speaker's articulatory system and the listener's auditory system when the speaker's brain activity precedes the listener (Liu et al., 2020) can be still present in a noninteractive situation (e.g., pseudo-hyperscanning). Critically, the instantaneous inter-brain relationship in verbal communications does not seem to be caused only by the shared physical signal of the speech.
According to a non-stationary conception of social interaction, we should expect changes in the direction of information flow at different time points during a conversation. Depending on the context, mutual dependency between brains could be dynamically lagged across the interaction. Therefore, we anticipate that different types of inter-brain relationships may be observed when we attempt to teach, refute, surprise or cajole our conversational partners. We also speculate that the articulatory-auditory (i.e., lagged) inter-brain relationship would be the primary neural signature of a verbal exchange that is trivial or irrelevant in meaning (e.g., small talk or phatic expressions such as saying "hello").
A further consideration for the timing of the inter-brain relationships is the distinction between how the signals from different brains are aligned and when the effect occurs, taking a given event as reference (e.g., an overt behaviour). The later case regards the role of inter-brain relationships for specific social cognition processes. For example, an inter-brain relationship preceding the initiation of social interaction may suggest a causal role in social decision-making (Omer, Zilkha, & Kimchi, 2019).

6.
Inter-brain relationship as a putative mechanism The allure of proposing inter-brain relationships as a causal mechanism for social interaction seems straightforward: if shown, this would imply the ground-breaking possibility of externally perturbing brain activity to modify (disrupt or enhance) inter-personal (or inter-animal) behaviours. However, these observations of inter-brain relationships in neural activity are best described as showing neural consequences of computations during social interaction (Dikker et al., 2017). Causal evidence is needed to show that inter-brain relationships make a mechanistic contribution to social interaction (Novembre & Iannetti, 2021).
A widely accepted definition of a biological mechanism requires four basic features: a phenomenon, parts, causings, and organisation (Craver & Tabery, 2017). In the present manuscript, the phenomenon would be the inter-brain relationship. The parts could be the dedicated neural substrates for social interaction processing in the (primate) brain (B aez-Mendoza, Mastrobattista, Wang, & Williams, 2021;Sliwa & Freiwald, 2017) or the specific networks supporting speech planning during interactive language use (Castellucci, Kovach, Howard, Greenlee, & Long, 2022). Moreover, the organisation would include the two brain patterns' brain localisation and temporal dynamics. Finally, the causing would be given by a causal role of the inter-brain relationships over the emergence and maintenance of the social interaction. The specific internal organisation of inter-brain relationships and their causal role on behavioural outcomes (e.g. communication or decision-making) will, however, remain unidentified until the behavioural impact of perturbing inter-brain relationships can be shown. This requires experiments in which joint brain stimulation is applied during social interaction (Novembre & Iannetti, 2021). Achieving this goal requires new methods for coordinated brain stimulation in multiple individuals (e.g., hyper-tACS or hyper-TMS). These methods permit inter-brain relationships between neural activity to be enhanced (by coherent or related forms of stimulation in two or more individuals), or suppressed (by incoherent or unrelated stimulation).
'Closing the loop' between multi-brain stimulation and interpersonal behaviour also requires an objective assessment of the success or failure of information transmission during different inter-brain situations. In the case of language communication, this goal can be achieved by using conversational analysis techniques (Sidnell & Stivers, 2013). For example, we can measure the number of 'recruitments' during the interaction, a concept developed to encompass the linguistic and embodied ways assistance may be sought (i.e., an index of cooperation) (Kendrick & Drew, 2016). Quantitative measurements of the quality of information transmission can be obtained from models developed within the NLP and Computational Linguistics. For example, 'coherence of speech' is a measure capturing digression from a topic that is obtained from latent semantic analysis (LSA) (Hoffman, Loginova, & Russell, 2018). New forms of quantification of social behaviour such as measures of inner group relationships (Dikker et al., 2017;Nozawa, Sasaki, Sakaki, Yokoyama, & Kawashima, 2016;Toppi et al., 2016) are of particular importance. The (non-verbal) motion relationship between interacting participants tracked from inertial measurement unit (IMU) wearables (Lahnakoski, Forbes, McCall, & Schilbach, 2020; or video recordings (Sabharwal et al., 2022) is another interesting possibility for quantifying interpersonal coordination.
In conclusion, characterising inter-brain relationships as a causal mechanism for social interaction remains premature until stimulation and behavioural methods can be combined.

Naturalistic interactions
Inter-brain relationships are most commonly assessed in a face-to-face context to emulate the multisensory nature and dynamically changing demands of social interaction. However, mimicking the multiple attributes in real-world environments could be challenging if we want to isolate the contribution of the voice, the face or other movements for communication. Although naturalistic approaches offer many advantages over controlled stimuli (e.g., in language, Hamilton & Huth, 2020), we can build on the classical unidirectional stimulus-response approaches to determine the neural coding of social signals. In this sense, the field can learn from more traditional paradigms about isolating an aspect of interest. Thus, we believe that a paradigm could be both, efficient and ecologically valid for the study of interactions while isolating one of the channels of the interaction. Nonetheless, we stress that more available channels (e.g., face-toface as compared with a telephone conversation) will improve the quality of communication, and will yield a more robust inter-brain relationship pattern.
Hyperscanning and the inter-brain relationship analyses can be substituted by clever multi-brain designs that do not directly compares between individual brain activities. For example, we can compare within participants between socially relevant conditions or to determine whether individual neural activity can be explained by models including other's behaviours (e.g., strategies) (e.g., Ong, Madlon-Kay, & Platt, 2021). Another alternative is using controllable interactive tasks in the scanner like those developed by Schilbach and colleagues (e.g., gaze-contingent paradigms) (Brandi et al., 2021;Pfeiffer et al., 2014;Schilbach et al., 2010). While these tasks only focus on one brain, they still provide the opportunity to engage in a reciprocal exchange.
Finally, social aspects like social goals and status are relevant. Future studies should attempt to disentangle features of shared linguistic information vs social factors in interactive alignment and communicative success (Garrod, Tosi, & Pickering, 2018). The further naturalistic setting will move beyond interactions as dyadic and consider dynamic multiperson links (Kingsbury & Hong, 2020). Social groups are a central aspect of human social interactions. It is highly desirable to bring central social network mechanisms and conceptsdsuch as homophily, transitivity, and popularitydto the context of interactions (Hoffman, Block, Elmer, & Stadtfeld, 2020). Yet, dyadic inter-brain relationships remain the building block of these larger groups and should be central.

Final remarks
Multi-brain imaging approaches have opened the door to the non-invasive exploration of the social brain. In this context, exploring inter-brain relationships is a valid methodological approach to investigating the biological nature of social interactions by comparing different brain signals. The validity and usefulness of inter-brain relationships are, thus far at least, independent of whether they are a mechanism for social cognition. However, if clear associations between the interbrain relationship and specific processes of social cognition are meaningful they will have numerous potential practical applications. For example, insofar as effective psychotherapy successfully modifies behaviour and effective teaching allows learning, it may be accompanied by changes in the patterns of inter-brain relationships (e.g., psychotherapy, Ellingsen et al., 2020). Thus, we face the intriguing possibility that as hyperscanning techniques improve, they will be helpful for situations that go beyond basic research and towards monitoring the status and enhancing the progress of natural social interactions. The study of inter-brain relationships already constitutes a new opportunity to advance Cognitive Neuroscience (Hasson & Nusbaum, 2019) and a foundational base in the emergence of the Neuroscience of Social Interactions (Hari et al., 2016). Let us researchers studying how mutual understanding takes place, continue speaking and listening to each other.

Declaration of competing interest
None.

Acknowledgement
The scope and content of this manuscript was enriched by the comments of three anonymous reviewers.