Implicit and explicit learning of socio-emotional information in a dynamic interaction with a virtual avatar

The vast amount of evidence for the existence of IL has been obtained using non-social stimuli and tasks, such as the Artificial Grammar Learning task (AGL; Reber, 1967) or the Serial Reaction Time task (SRTT; Cleeremans & McClelland, 1991; Nissen & Bullemer, 1987). In these paradigms, participants learn regularities that structure nonsocial stimuli such as letter strings (in AGL) or spatial locations (in the SRTT) and are generally considered to provide robust evidence for IL effects (e.g., Dienes & Seth, 2018; but see Shanks, 2004, 2010). In contrast, there is very limited direct evidence that IL can operate upon socially relevant stimuli, the evidence being restricted to the aforementioned studies (e.g., Norman & Price, 2012) and few others (but see Hendrickx et al., 1997). Hence, the claims that IL could be involved in social functioning does not primarily rest on studies using social stimuli, but rather on the assumption that IL is a general process which unfolds in the same way, regardless of the specific stimuli on which it operates (e.g., Pothos et al., 2006)—e.g., if IL operates on letter strings, it might operate roughly similarly with social stimuli. Consequently, for instance, the studies that attempt to test whether social functioning deficits in some disorders could be caused by deficits in implicit learning use such nonsocial tasks (AGL, SRTT) and assume to test a “general” implicit learning capacity (e.g., Exner et al., 2009; Naismith et al., 2006, 2010; Nemeth et al., 2010; Rathus et al., 1994; Zwart et al., 2018).

The assumption that IL is a general capacity is, however, at odds with a consistent corpus of empirical findings: the extant literature on the boundary conditions for IL clearly shows that we cannot implicitly extract regularities from all categories of stimuli to the same extent. For instance, participants automatically learn a task-irrelevant artificial grammar when it is instantiated by evolutionary relevant stimuli (human faces), but not when it is instantiated by evolutionary irrelevant stimuli (buildings) (Eitam et al., 2014); see also Dienes and Altmann (1997), Jimenez et al. (2020), or Scott and Dienes (2010), for further evidence for the dependency of IL on the characteristics of the surface stimuli. Closer to the social domain, Ziori and Dienes (2015) found that an artificial grammar was learned less when it structured sequences of faces (56% classification accuracy), compared to the same grammar that structured letter strings (64% accuracy; Dienes & Scott, 2005). In a direct comparison, Norman and Price (2012) found less learning of an artificial grammar that structured sequences of body postures compared to the same grammar when it structured letter sequences (53% vs 58% classification accuracy). Further evidence comes the related field of statistical language learning, which is often assumed to occur, partially, implicitly: Li et al. (2022) have found that young adults with a high level of autistic traits are able to extract statistical regularities from non-social auditory input (pure tones), but not from socially relevant auditory input (Chinese disyllables), bringing further support for the stimulus-dependent operation of implicit/statistical learning. In sum, there is robust evidence that the surface stimuli influence the amount of learning in typical implicit learning tasks. There is also emerging evidence that, at least in the AGL task, learning is lower with more complex, social stimuli, compared to the typical version that uses letters. Accordingly, given that IL functions differently with different types of stimuli, it may be unwarranted to make inferences regarding the role played by IL in the social domain based on studies that use non-social surface stimuli (see Norman & Price, 2012; Jimenez et al., 2020, for thorough analyses of this problem). Moreover, given the emerging evidence that complex social stimuli are learned less well than simpler nonsocial stimuli (e.g., Ziori & Dienes, 2015), a legitimate concern is whether there is any learning for even more ecologically relevant social stimuli than those already used, such as dynamic facial emotional expression.

A different limitation is that the few existing studies that have employed more socially relevant stimuli have used experimental paradigms in which participants are exposed to social contingencies in a relatively passive manner. For instance, in the hidden covariation detection task, participants are exposed to avatars that instantiate contingencies between physical and psychological features, but there is no change in the participants’ behavior as a result of the avatar’s behavior or vice versa. In contrast, as we will clarify in the subsequent section, the central feature of our social life is its deeply interactive, dynamical nature.

We start from the observation that in both real-life and experimental environments, information is being exchanged in loops. Specifically, in most cases of real-life interaction, an individual’s behavior determines or encourages a response from the social environment; further, the individual typically reacts again to the social environment’s response—thus perpetuating a loop of informational exchange. Similarly, in experimental contexts, a participant’s behavior determines a response from the research paradigm (e.g., advancing to the next stimulus) to which the participant typically reacts again, thereby perpetuating a loop of informational exchange.

However, the loops through which information is exchanged in most IL tasks are fundamentally different from those in which information is exchanged in social environments. Specifically, in the real-world environment, if an individual behaves in different manners, s/he should expect different responses—i.e., information is being exchanged via feedback-driven interactive loops (Becchio et al., 2010). By contrast, in most IL paradigms, participants respond to overly complex, predefined sequences of stimuli which, crucially, do not adapt in reaction to their responses—i.e., information is being exchanged via noninteractive loops. For instance, in the acquisition phase of the AGL participants are exposed to a predetermined list of letter strings; there is no modulation in the behavior of the task or of the stimuli as a consequence of participant’s behavior. The same principle applies to the SRTT and to the hidden covariation detection task.

Here, we emphasize that an instrument that aims to assess the role of IL in social interactions—besides using socially relevant surface stimuli—should also simulate the dynamic, feedback-driven manner in which information is being exchanged in such environments. One example of such a method is the classic Dynamic Systems Control task (DSC; Berry & Broadbent, 1984)—which will be detailed in the paragraphs below.

In the DSC, participants gradually learn to control the value of a dynamic system’s output variable by manipulating the value of an input variable. For instance, in the Sugar Production condition of their seminal paper, Berry and Broadbent (1984) asked participants to reach and maintain a certain output of a small sugar factory. Participants could achieve their task simply by manipulating the number of employed workers in any given trial. Unbeknownst to the participants, however, the relationship between their inputs and the amount of sugar production was mediated by a counterintuitive equation such that the same input from the participant determined different outputs depending on the previous state of the system. In other words, the system interacted with the participants, by adapting its responses according to participants’ successive inputs, thereby implementing an interactive loop of informational exchange. Berry and Broadbent (1984) found that participants improved their performance maintaining the desired output as the task progressed, suggesting that learning occurred. Furthermore, as assessed by a written post experimental questionnaire, the authors found that the performance improvement was not associated with reportable knowledge of the underlying rule, suggesting that learning had been implicit. Congruent results were also obtained by several other investigations (e.g., Berry & Broadbent, 1995; Dienes & Fahey, 1995; Fahey & Dienes, 1998).

Berry and Broadbent (1984) also attempted to adapt their Sugar Production task to assess the role of IL in social interactions. For this purpose, in the Personal Interaction condition of their experiment, the authors told the participants that they would be interacting with a virtual person named Clegg. Participants were further informed that their task was to keep Clegg’s mood at a prespecified level. Participants could enter their inputs via a keyboard and Clegg’s responses were presented on a display as linguistic labels (i.e., text vignettes). The following adjectives were used both as Clegg’s possible moods and as participants’ response options: “Very Rude, Rude, Very Cool, Cool, Indifferent, Polite, Very Polite, Friendly, Very Friendly, Affectionate, Very Affectionate or, Loving”. Each of the adjectives was assigned a numerical value from 1 to 12. Unbeknownst to the participants, the relationship between their inputs and Clegg’s mood was mediated by the same equation as that used in the Sugar Production condition. Berry and Broadbent (1984) found that participants improved their performance to keep Clegg’s mood on the target state as the task progressed, suggesting that learning occurred in the task. As in the Sugar Production condition, learning was also found to be mostly implicit.

Importantly, we suggest that while the Personal Interaction task implements interactive loops of informational exchange between the participants and the task, its use of linguistic labels as surface stimuli keeps it abstract in a way that severely limits its relevance to the social domain.

While we consider that the existing studies on the involvement of IL in our social functioning provide encouraging starting results, we also think that their relevance for the social world is restricted by two main factors: (a) the artificial nature of the surface stimuli (especially in the case of AGL and SRTT) and (b) the static nature of the paradigms themselves, which seldom, if ever, afford the kinds of interactive contexts that nevertheless constitute the core feature of social life.

To overcome these limitations, we developed a novel task consisting of a learning phase and of an awareness test phase. In the learning phase, based on the classic DSC task, participants responded to the cinematic facial expressions of a realistic avatar, with the aim of bringing its facial expression in a predefined state. Unbeknownst to the participants, their responses were related to the avatar’s expressions via a complex equation modeled after Berry and Broadbent (1984): the equation likewise took into account both the avatar’s previous expression and the participants’ response. Accordingly, the avatar’s expression was influenced by the participants’ response, while the participant had to adapt his/her response to the avatar’s previous expression; in other words, the avatar and the participant were in a continuous interactive loop throughout the learning phase. A significant increase in the number of trials in which the participants managed to bring the avatar in the target state would indicate that they have acquired knowledge from the task. The primary features of the learning phase were that participants were in a continuous interaction with the avatar and the fact that the avatar’s emotional facial expression changed in a cinematic, realistic manner—aspects that give our task an unprecedented level of external validity for studying implicit social learning.

The awareness test phase measured to what extent participants were aware of the information required to regulate the avatar’s state. As the issue of measuring awareness remains highly controversial, we discuss our choices in the following subsection.

The main objective of this study was to determine whether IL can be involved in the acquisition of the complex regularities present in a situation involving dynamic interaction with a life-like virtual agent. Based on the previous DSC studies (e.g., Dienes & Fahey, 1998), our hypotheses were (H1) that participants will acquire the regularity, (H2) that they will possess accurate judgement knowledge, (H3) that their accurate judgement knowledge will be based both on unconscious structural knowledge and on conscious structural knowledge. The detection of accurate judgement knowledge based on unconscious structural knowledge would enable us to conclude that IL occurred in the task. However, compared to the existing research, our task features substantially different stimuli (dynamic facial expressions) and modes of interaction with the task (responding with a facial expression). As discussed in the previous sections, the nature of the surface stimuli and the mode of interaction with those stimuli has been shown to influence the nature and the extent of learning (e.g., Jiminez et al., 2020). Accordingly, it is entirely possible that our results may diverge from those we hypothesize and that have been found in tasks that have used non-social or less complex stimuli.

We determined our sample size considering the statistical power needed to test H3, because it has a smaller expected effect size than that of H1 and H2. We expected a small to medium effect size for the unconscious learning effect stipulated by H3, based on previous studies that used similar methods to measure conscious and unconscious knowledge (e.g., Fu et al., 2012, 2018). Our power analysis indicated that a one-tailed test can detect a potential difference between two paired means (i.e., within-subjects design) that has an effect size of Cohen’s d_z = 0.3 with a statistical power of 1 − β = 0.9 in a sample of 97 participants. Note that for the other hypotheses, we expected large or medium to large effect sizes (for H1, a d_z = 0.798 based on Dienes & Fahey, 11998 for H2, a η²_p = 0.329, based on Fu et al., 2012; for the conscious learning effect stipulated by H3, a d_z = 0.67, based on Fu et al., 2012), and 97 participants provided a statistical power > 99% for all these effects. Therefore, we aimed for a sample size of at least 97, but, as participants were rewarded with partial course credit, a higher number of persons enrolled. A total of 115 first-year undergraduate students in psychology from the Babeș-Bolyai University, (99 female, M_age = 19.74, SD = 1.27) participated in this research. All participants had normal or corrected-to-normal vision, enrolled voluntarily, gave written informed consent, and were told that they could withdraw from the experiment at any time, without any negative consequences. Participants consented to have their anonymized data being made publicly available. This study respected the regulations of the Babeș-Bolyai University’s Research committee and have therefore been performed in accordance with the ethical standards laid down in the 1964 Declaration of Helsinki and its later amendments.

The stimuli were designed with iClone (Version 7.2; Reallusion; 2017); the JavaScript experiment was coded in PsychoPy/PsychoJs (Peirce et al., 2019) and ran on the Pavlovia.org servers. The experiment, including all raw stimuli, is accessible for download at osf.io/q9bac. For additional information on how to access the resource package, see the Supplementary material A.

In a within-group design, we used a two-step task with a learning phase and an awareness test phase. In the learning phase, participants were presented with a socially relevant environment that made it possible to quantify the on-line acquisition of knowledge. In the awareness test phase, we assessed the implicit/explicit status of the acquired knowledge.

Our learning phase is inspired by the classic DSC task (Berry & Broadbent, 1984), but brings important modifications. Amongst which the most important is the different types of surface stimuli. Specifically, on the one hand, Berry and Broadbent, as well as the bulk of the research implementing this task, used surface stimuli under the form of written words with an affective valence (e.g., “extreme anger”). On the other hand, we used dynamic facial expressions of a cinematic virtual male avatar (i.e., the video of a virtual human depicting—for instance—an actual, standardized, facial expression of extreme anger). The reason for which we decided not to design a female avatar is informed by the AGL study of Ziori and Dienes (2015), who found that the perceived attractiveness of stimuli depicting female faces affects learning performance of both male and female participants. Given that this effect did not appear for stimuli depicting male faces neither for male nor for female participants, we prevented the potential stimulus attractiveness effects on IL by designing socio-emotionally relevant surface stimuli that were expressed by a male avatar. In the next subsection, we detail the characteristics of these emotional facial expressions.

Next, we will present an overview of our procedure—for a more detailed account, consult the Supplementary material D. Participants were first asked to give written informed consent prior to completing the experimental activities. Second, they were asked for their demographic information. Third, they completed the learning phase and, fourth, they completed the awareness test phase. The entire experiment lasted around 25 min.

In the learning phase, before initiating the experiment, participants were given the following written instruction: “You will interact with John, a fictional character from an unknown culture; he is able to display only a limited number of facial expressions and is unable to regulate his facial expressions in the way that we do. John attends an important task and in order to be successful, he must not express intense facial expressions—neither positive nor negative. Your task is to assist him in regulating his facial expressions, aiming to bring him into a neutral state as many times as possible. Finally, John may respond to your inputs in ways that might seem atypical”. After participants read the instructions, we presented the feedback messages that could appear in this phase of the experiment (i.e., the positive feedback, the feedback for repeated responses and the feedback for late responses). Before initiating the experimental trials of the learning phase, we verified that participants knew how to interact with the avatar by asking them to perform 10 practice trials. Here, the avatar reacted randomly to participants’ inputs and these trials were dropped from the analyses.

After they completed the learning phase, participants were given the written instructions for the awareness test phase. Participants were informed that they would be presented with a facial expression of the avatar and they would have to choose the response they thought would bring the avatar in the target state (for the inclusion task) or in any other state but the target (for the exclusion task). They were further instructed to indicate the subjective basis of their response by choosing one of the four possible response options (i.e., Guess, Intuition, Rules or Memory). After participants indicated that they understood their task, we presented them with one practice trial (from the inclusion or the exclusion phase, depending on the condition). The practice trial was also discarded from the analyses. After completing the practice trial, participants started the awareness test phase. The definition of the response options appeared on the screen after each trial of the PDP.

After the awareness test phase was completed, participants were thanked for their involvement in this research and were given the contact information of the principal investigator in order to address their potential questions.

The raw dataset generated for this study is available on the Center for Open Science repository (osf.io/q9bac). If participants acquired knowledge from the task, we would expect an increase of the number of On-target trials as the task progressed. A one-way repeated measures ANOVA revealed a significant effect of Block (1–10, within-subjects) on the number of On-target trials, F(9, 114) = 38.33, p < 0.001, η²_p = 0.252. A follow up repeated measures t test indicated that participants generated significantly more On-target trials in the 10th acquisition block (M_proportion = 0.309, SD = 0.198) than in the 1st acquisition block (M_prop. = 0.135, SD = 0.106), t(114) = − 9.05, p < 0.001, Cohen’s d = 0.84. Altogether, these results clearly show that learning had occurred during the task (see Fig. 4).

Participants’ ability to flexibly conform to the opposing requirements of the inclusion and of the exclusion tasks would be enabled by their possession of accurate judgement knowledge. To determine whether they possessed this accurate judgement knowledge, we compared the number of On-target trials from the inclusion task with those from the exclusion task.

A mixed ANOVA assessed the effects of Instruction (within-subjects: inclusion vs. exclusion) and task order (between-subjects: inclusion–exclusion vs. exclusion–inclusion) on the number of On-target trials generated in the test phase. We found a significant Instruction effect, F(1, 113) = 107.01, p < 0.001, η²_p = 0.486, indicating that participants generated significantly more On-target trials in the inclusion (M_prop. = 0.357, SD = 0.220) than the exclusion (M_prop. = 0.104, SD = 0.098) task. We failed to detect either a significant Task order effect, F(1, 113) = 0.61, p = 0.435, η²_p = 0.005, or a significant Instruction by Task order interaction effect F(1, 113) = 0.07, p = 0.794, η²_p = 0.005. These results suggest that participants acquired accurate judgement knowledge from the task. For a graphical representation and analysis of participants’ accuracy in the PDP against the chance level, consult Supplementary material E.

After having established that participants possessed accurate judgement knowledge, we now analyze to what extent it is based on unconscious and/or on conscious structural knowledge.

Following established analytical strategies in IL research (Norman & Price, 2012; Ziori & Dienes, 2015), we combined the Guess and Intuition response attributions to create Implicit attributions scores, and the Rules and Memory attributions to create the Explicit attributions scores. Roughly half of the total responses were based on Implicit attributions (in inclusion, M_prop. = 0.528, SD = 0.306 and in exclusion, M_prop. = 0.465, SD = 0.368) and the other half were based on Explicit attributions (in inclusion, M_prop. = 0.472, SD = 0.306 and in exclusion, M_prop. = 0.535, SD = 0.368). For a detailed perspective on the distribution for each of the four response bases (i.e., Guess, Intuition, Rule and Memory) across the inclusion and the exclusion tasks of the PDP, see the Table 3 below.

We then analyzed the accuracy of participants’ judgement knowledge depending on the conscious/unconscious status of their structural knowledge. First, we assessed whether participants had accurate judgement knowledge when they reported that their responses were based on explicit structural knowledge. Note that the overall analysis only included the 96 (out of 115) participants, who reported relying on an explicit attribution (i.e., Rule or Memory) on at least one trial in both the inclusion and the exclusion tasks. For responses based on Explicit attributions, a paired sample t test indicated that participants generated significantly more On-target trials in the inclusion (M_prop. = 0.558, SD = 0.347) than the exclusion task (M_prop. = 0.072, SD = 0.114), t(96) = 13.02, p < 0.001, d = 1.32. In a more granular perspective, for responses based on Rules, a paired sample t test indicated that participants generated significantly more On-target trials in the inclusion (M_prop. = 0.542, SD = 0.369) than the exclusion task (M_prop. = 0.062, SD = 0.133), t(62) = 9.38, p < 0.001, d = 1.18. The same pattern of results was observed for the responses based on Memory—participants generated significantly more On-target trials in the inclusion (M_prop. = 0.537, SD = 0.401) than in the exclusion task (M_prop. = 0.076, SD = 0.159), t(53) = 10.09, p < 0.001, d = 1.37. Collectively, the analyses above indicate that participants had accurate judgement knowledge on trials where they reported relying on conscious structural knowledge.

We then assessed whether participants had accurate judgement knowledge when they reported that their responses were based on unconscious structural knowledge. Note that the overall analysis only included the 87 (out of 115) participants who reported relying on an implicit attribution (i.e., Guess or Intuition) in at least one trial in both the inclusion task and the exclusion tasks. For responses based on Implicit attributions, a paired sample t test indicated that participants generated significantly more On-target trials in the inclusion (M_prop. = 0.200, SD = 0.209) than the exclusion task (M_prop. = 0.130, SD = 0.156), t(87) = 2.88, p = 0.006, d_z = 0.31. In a more granular perspective, for responses based on Intuition, a paired sample t test indicated that participants generated significantly more On-target trials in the inclusion (M_prop. = 0.186, SD = 0.229) than the exclusion task (M_prop. = 0.126, SD = 0.153), t(73) = 1.67, p = 0.05, d = 0.19. Furthermore, even for responses that were attributed to Guess, the inclusion (M_prop. = 0.210, SD = 0.287)—exclusion (M_prop. = 0.134, SD = 0.222) difference on the number of On-target trials that were generated by the participants remained statistically significant, t(61) = 1.76, p = 0.041, d = 0.224.

In the following, we employed an alternative method to verify if IL occurred in the task: the guessing criterion (Dienes et al., 1995)—which can be computed by comparing participants’ performance in the inclusion task against the chance level. However, to do this, it was first necessary to specify our chance level. Given that in each trial of the PDP, out of the 7 possible response options, only one could regulate the avatar’s facial expression to the target state, our chance level is 0.143. According to the guessing criterion, if participants acquired unconscious knowledge of the rule, we would expect them to show above-chance performance for the trials on which they indicated that they based their responses on guesses. Indeed, congruent with the analysis above, a one-sample t test indicated that even when indicating guessing as a response attribution, participants generated significantly more On-target trials (M_prop. = 0.201, SD = 0.287) than what would be expected at the chance level (i.e., 0.143), t(88) = 2.26, p = 0.0013, Cohen’s d = 0.239.

Collectively, the analyses above clearly indicate that participants had accurate judgement knowledge on trials where they reported relying on unconscious structural knowledge—when they reported basing their answers on an intuition or, even when they indicated that they had chosen them at random.

First, people frequently (i.e., in 49.65% of the trials, averaged over inclusion and exclusion tasks) reported that their decisions were based on implicit decision strategies, indicating that they did not feel that they were aware of the structure that guided their responses in the PDP. We note that initial IL studies operated on the assumption that an accurate PDP performance indicates conscious knowledge of the learned structure; that is, participants would be able to respond according to the learned structure in inclusion, or avoid responding according to it in exclusion, only if they were aware of the structure (e.g., Destrebecqz & Cleeremans, 2001). Nevertheless, a plethora of more recent studies have shown that equating PDP performance with awareness is unwarranted, since subjectively unconscious knowledge can successfully guide participants’ PDP performance. To our knowledge, this is what every study that has supplemented the PDP with subjective response attribution has found (Fu et al., 2012, 2018; Norman et al., 2011, 2016, 2019; Wan et al., 2008), and that is what we have also found: that participants can flexibly follow the requirements of the inclusion and of the exclusion tasks, even when they report not being aware of the structure. Our results also conceptually replicate previous DSC research; similar to Fahey and Dienes (1998), we found that the judgement knowledge is accurate in the sense that, given a situation, participants know what the appropriate response should be. Furthermore, similar to Fahey and Dienes, we found evidence for unconscious knowledge in the sense that subjective access to the memory representation enabling accurate judgement is not required for optimal task performance. More to this point, we consider that this pattern of results resembles natural social phenomena; for instance, situations in which one knows how to behave in a certain context even if one is unable to consciously access the underlying knowledge which substantiated one’s decision/way of action.

Second, as in most other IL tasks, our results do not provide a clear answer regarding the exact nature of the acquired knowledge. For instance, as discussed in the introduction, in the AGL task, participants could learn both specific combinations of letters that make a string grammatical or not, but could also learn more abstract types of knowledge pertaining to the structure of the grammar (such as the global or the local repetition proportion; see, e.g., Scott & Dienes, 2008). Similarly, in the context of DSC, Dienes and Fahey (1995) have found that participants’ performance is not typically guided by the abstract equation embedded in the task, but rather by implicit and explicit memories of interacting with the system on specific trials. Accordingly, we claim that participants have extracted structural knowledge from the task, because they had accurate performance in the PDP and in turn, accuracy in the PDP requires at least knowledge of the mappings between stimuli and responses. However, we do not claim that participants necessarily developed an abstract, mathematical, representation of the equation itself. Rather, it is likely that they have acquired a memory base regarding specific interactions with the task and their outcome. This memory base then informed participants’ ability to conform to the opposing tasks of the PDP while being partly implicit (i.e., the proportion of On-target trials attributed to implicit response bases was 0.198 in the inclusion and 0.13 in the exclusion) and partly explicit (i.e., the proportion of On-target trials attributed to explicit response bases was 0.542 in the inclusion and 0.07 in the exclusion task).

A third aspect regarding the conscious/unconscious status of learning in the present study is that our task has produced a higher proportion of trials attributed to explicit decision strategies compared to the already established IL paradigms. For example, most AGL studies reveal that about 2/3 of the responses are attributed to unconscious structural knowledge (e.g., Dienes & Scott, 2005). In our study, participants attributed their responses to strategies based on unconscious structural knowledge on only about half the trials (see Table 3 above). Thus, it seems that our participants felt that they were applying explicit knowledge to a larger extent than participants have done in other IL experiments. This apparent enhanced acquisition of explicit knowledge in our task affords several candidate explanations. According to the first these candidate accounts, as in all DSC studies, our instructions asked participants to control the state of a system—this instruction by itself is likely to have triggered a more analytical processing style (Dienes & Scott, 2005). A second explanation might be that, in contrast to previous DSC research, we did not use noise in our task. Specifically, in each trial of Berry and Broadbent’s (1984) acquisition phase, the novel state of the system was determined by running an equation to which the computer added 1, 0, or − 1 units on a random basis. Crucially, responses were correct if they were either on target or one unit around the target. We consider that this liberal method to quantify task performance has the potential to be contaminated by both false-positives as well as false-negatives. Therefore, we chose not to introduce the random factor.³ As a third explanation, previous research argued that different implementations of the PDP offer different estimates of the implicit and explicit judgement knowledge. For instance, in the context of the SRTT, by using a PDP with free generation tasks, Destrebecqz and Cleeremans (2001) showed that participants acquired unconscious judgement knowledge of a sequence; whereas by using a PDP with cued generation tasks, Wilkinson and Shanks (2004) indicated that participants acquired explicit judgement knowledge of a sequence. Our version of the PDP used pseudo-cued generation tasks. Specifically, instead of being exposed to a succession of stimuli, participants were exposed to only one facial expression and asked to generate the next to either bring the avatar in the target state (in the inclusion task) or avoid bringing the avatar in the target state (in the exclusion task). As Stahl et al. (2015) showed that the cued generation tasks have the potential to introduce bias in the PD estimates, future research could investigate if the implicit-explicit estimates of the acquired knowledge from our paradigm differ as a function of the generation task (free generation vs cued generation).

The discovery of implicit learning for rules that govern dynamic interactions with a social stimulus advances the existing tools for investigating the involvement of nonconscious learning in social contexts in several important ways, which we will detail in the paragraphs below.

First, as briefly mentioned in the introduction, results obtained in the hidden covariation detection task indicated that participants were able to implicitly learn an arbitrary covariation between a physical feature and personality features, and that the judgement of new avatars is influenced by the learned covariation (Ivanchei et al., 2019; Lewicki, 1986). On a more distal note, Geiger et al. (2018) have shown that social cues (the gaze of an avatar) can guide and facilitate the development of implicit knowledge in a motor task (a variant of the SRTT). While they employed surface stimuli that are relevant for social functioning, the loops through which information is being exchanged between participants and these tasks of these studies are fundamentally different from those in which information is exchanged in social environments. Specifically, while in these tasks (i.e., Geiger et al., 2018; Ivanchei et al., 2019; Lewicki, 1986) information is being exchanged via noninteractive loops, in the real social environment, information is being exchanged via feedback-driven interactive loops. The fact that our task implements such feedback-driven interactive loops increases its external validity for evaluating the function of IL in the social domain.

Second, the present study is the first one to use surface stimuli under the form of realistic emotional facial expressions. The inclusion of socially relevant surface stimuli in our task does not have a mere cosmetic effect but is motivated by two arguments: a first argument, as presented in the introduction, is that the existing studies show that IL is highly sensitive to the stimuli it operates on: the same regularity can be learned to a variable degree (e.g., Jimenez et al., 2020) or can be learned more or less explicitly (Norman & Price, 2012) depending on the characteristics of the surface stimuli. To our knowledge, this is the first study to use kinematic, realistic, facial expressions, bringing our research one step closer to the habitual mode of processing this information in the real social environment. A second argument for the necessity of tailoring the experimental task to the domain of investigation (e.g., social, motor, linguistic) derives from a core principle of current neural and computational theories of IL: that is, IL occurs through changes in the speed, strategy, and efficiency of processing specific stimuli with different processing constraints. For instance, in their influential paper, Frost et al. (2015) present ample evidence that implicit and statistical learning are represented by “a set of domain-general computational principles that operate in different modalities and, therefore, are subject to the specific constraints characteristic of their respective brain regions” (see also Conway, 2020; Reber, 2013).

Third, the strategy for assessing knowledge awareness represents another strong point of the study. Specifically, previous DSC research measured knowledge awareness by direct questioning (i.e., post experimental questionnaires) or verbal report. To the best of our knowledge, this is the first DSC study to employ trial by trial subjective measures of awareness (i.e., response attributions and the PDP); thereby, offering a more precise perspective on the way in which explicit knowledge, as well as implicit knowledge, contribute to task performance.

First, the test phase consisted of precisely the same interaction instances encountered in the acquisition phase. Therefore, on the grounds of this method, we cannot infer transfer of the learned information to novel situations. Future research could use different avatars across the acquisition blocks while maintaining the same rule that structures the interactions to assess the transference of the learned content to novel situations. At this stage, our task cannot provide a definitive answer to whether participants learned the abstract mathematical equation or they learned specific interaction sequences (Dienes & Fahey, 1995). Third, the interaction between participants and the avatar was centered around a limited number of emotional facial expressions that were displayed in exactly the same parameters—however, in the natural social environment, different configurations of a facial expression can express, for instance, an intense level of joy; these two aspects depart from the diversity that characterizes genuine social environments. Hence, future research could employ a larger variety of different surface stimuli that would express similar emotional states.

Despite the importance of IL for social functioning, investigations focusing on the IL of socially relevant information are relatively limited. A potential cause for this may be represented by the challenges involved in the development of methodologically sound experimental paradigms that assess IL in a socially relevant context. Given that the implicit acquisition of regularities unfolds differently depending on the characteristics of the stimuli that follow those regularities (e.g., on their perceptual complexity, motivational relevance, habitual mode of processing those stimuli; see, e.g., Jimenez et al., 2020; Seger, 1994), we suggest that the inclusion of socially relevant surface stimuli could enable our task to reveal untapped, social, facets of this non-homogenous process. We conclude by suggesting that our findings could foster novel research avenues into the role that IL might play in different conditions characterized by atypical social functioning. For instance, it could complement or even replace, in some cases, the AGL and SRTT in studies that investigate the integrity of implicit learning in autism spectrum disorders, depression etc.