Contingency learning involves being sensitive to the statistical relationship between cues and outcomes, which is crucial for decision-making, categorization, and causal reasoning (Baker et al., 1996; Heisz et al., 2011). While much research focuses on neutral stimuli, emotional contingencies are essential for maintaining social understanding and navigating interpersonal dynamics, and where responses often interact with others’ emotional states to moderate social relationships (Thornton and Tamir, 2017). For example, interpreting facial expressions predicts others’ emotions and evokes corresponding responses, reinforcing social interactions. This reciprocal exchange, such as a smile prompting a smile, influences both individuals’ emotional states, fostering social cohesion and mitigating potential discord. Understanding and responding to these emotional cues is fundamental to successful social interactions.

Unlike typical emotion recognition tasks, real-world facial expressions are context-dependent and influenced by reciprocal cues between communicators and emotional expressions (Kunar et al., 2014). Rather than remaining in a constant emotional expression, as in emotion recognition tasks, facial expressions in real-world interactions change and are influenced by reciprocal cues between the sender and receiver, which establish contingent events during social interaction (Pavlova and Sokolov, 2022; Straulino et al., 2023). In such situations, emotional valence and social impact can shape the perception of causality based on pre-existing, often egocentric, mental models of emotional transitions (Thornton and Tamir, 2017; Anderson et al., 2019; Barrick et al., 2024). These models may bias perception of causality, particularly in social relational contexts, such as transitions from sadness to neutrality. This is influenced by both subjective and universal affective patterns, shedding light on how emotions may shape learning in dynamic situations.

To test people’s ability to learn these relationships based on statistical reliability, we presented pairs of faces, as if a transmitter and receiver were engaged in a conversation and manipulated four types of events: (a) when a specific emotion was present in both the transmitter (C) and receiver (O), (b) when the emotion was present in the transmitter (C) but absent in the receiver (~O), (c) when the emotion was absent (neutral) in the transmitter (~C) but present in the receiver (O), and (d) when both transmitter and receiver emotions were neutral (~C and ~O). Allan (1980) introduced the Delta P (∆P) rule, a metric for calculating contingency between these events using a 2×2 contingency matrix (Table 1). Delta P (∆P) is computed as the difference between the conditional probability of the outcome given the presence of the cue, P(O|C), and the conditional probability of the outcome given the absence of the cue, P(O| ~ C). The value of ∆P ranges from −1 to +1, with positive values indicating a positive contingency (increasing as events a and/or d increase) and negative values indicating a negative contingency (increasing as events b and/or c increase) (Allan, 1980; Crump et al., 2007; Hannah et al., 2009; Murphy et al., 2022).

The learning of cue-outcome contingencies more generally seems to involve a diverse set of frontal–parietal and temporal regions, such as inferior frontal gyrus (IFG, BA 44/45), middle frontal gyrus (MFG, BA9), superior frontal gyrus (SFG, BA 10), anterior and posterior cingulate gyrus (AC/PC, BA 24-32-23), inferior parietal lobule (IPL, BA 40) and superior temporal gyrus (STG BA 43) (Fletcher et al., 2001; Tsukiura et al., 2003; Turner et al., 2004; Carter et al., 2006; Tanaka et al., 2008; Klucken et al., 2009; Mullette-Gillman and Huettel, 2009; Liljeholm et al., 2011; Spiers et al., 2017; Morris et al., 2022). For instance, Turner et al. (2004) investigated the role of prefrontal cortices in an associative learning task in which participants were required make judgments about how certain foods (cues) were related to an allergic reaction (outcome) based on being exposed to multiple trials with the same categories of events involved in Delta P. The study showed that contingency judgments involved positive activations in lateral prefrontal regions (e.g., IFG and MFG). Liljeholm et al. (2013) found that contingency learning was associated with significant brain activations in prefrontal regions, including the lateral and medial prefrontal cortex (medPFC) and IFG. It has been suggested that these regions are part of a goal directed system that may rely on computing relative values of cues as possible causal candidates (Turner et al., 2004; Tanaka et al., 2008; Liljeholm et al., 2013).

Evidence from contingency judgment studies further refines this understanding by demonstrating that attentional resource allocation varies with contingency type (Maldonado et al., 2006; Saylik et al., 2022). First, zero-contingency tasks, characterized by the absence of predictive relationships between cue and outcome, represent a form of uncertainty that may elicit stress and demand significant cognitive and neural resources (Behrens et al., 2007; Le Pelley et al., 2010; Murphy et al., 2011; Matute et al., 2015; Peters et al., 2017; Stolyarova et al., 2019). The dorsomedial prefrontal cortex (DMPFC), lateral prefrontal cortex (LPFC), and temporal regions are critical for identifying shifts in environmental contingencies and updating cognitive models to accommodate the absence of clear patterns and formation of causal biases (Mullette-Gillman and Huettel, 2009; Spiers et al., 2017). Complementing these regions, the anterior cingulate cortex (ACC) plays a pivotal role in managing uncertainty by integrating volatility signals and modulating learning rates, enabling individuals to adaptively respond to uncorrelated outcomes (Behrens et al., 2007). Second, negative contingencies are also considered demanding because it is more difficult when the presence of a cue signals the absence of an outcome (Maldonado et al., 2006; Saylik et al., 2022). For instance, Heisz et al. (2011) employed a streaming paradigm with emoticons, revealing that negative contingencies elicit stronger activity in attention-related brain areas compared to positive contingencies. Extending this, our functional magnetic resonance imaging (FMRI) study used geometric shapes (triangles and hexagons) to examine contingency judgments. We observed that negative contingencies led to higher activation in areas such as the IFG and MFG, whereas a common neural response was found in the posterior cingulate cortex (PCC), supramarginal gyrus, and STG across all contingencies (Saylik et al., 2022). Finally, positive contingencies may involve more automatic processes, as it is easier to follow co-presence and co-absence of cues and outcomes, and thus elicit less activation in cognitive control areas (Maldonado et al., 2006; Saylik et al., 2022). Taken together, zero and negative contingencies demand greater cognitive and neural resources due to their inherent uncertainty and complexity, engaging regions like the DMPFC, LPFC, and ACC compared with positive contingencies.

In addition, emotions interact with higher-level cognitive processes, including decision-making and judgment, by altering how individuals perceive, interpret, and evaluate available information (Blanchette and Richards, 2010; Brosch et al., 2013). The effect of emotional stimuli may become more pronounced in uncertain and demanding conditions (e.g., zero or negative contingencies) due to the increased levels of uncertainty and cognitive difficulty involved. Two potential mechanisms may explain these effects: the availability heuristic and the narrowed attention hypothesis. The availability heuristic suggests that emotions influence judgments by affecting memory retrieval, with mood-congruent memories being more accessible (Blanchette and Richards, 2010). The attention-narrowing hypothesis posits that negative emotions may elevate arousal levels, leading individuals to focus on certain aspects of information, such as saliency, during decision-making and judgment (Easterbrook, 1959; Blanchette and Richards, 2010; Wichary et al., 2016). In this regard, sad emotional stimuli may lead to more conservative and critical judgments by priming sadness related negative exemplars through the availability heuristic or by narrowing attention due to elevated arousal associated with the uncertainty and stress inherent in sadness (Tiedens and Linton, 2001; Blanchette and Richards, 2010; Wichary et al., 2016; Anderson et al., 2019). Happy emotional stimuli, on the other hand, may enhance the retrieval of positive associations, promoting more optimistic and liberal judgments by priming positive exemplars through the availability heuristic and broadening the scope of attention, as they do not elicit stress-related arousal levels (Blanchette and Richards, 2010; Evans et al., 2010). Together, these mechanisms demonstrate how emotions dynamically modulate cognitive strategies, influencing perception and judgment based on the emotional context (Anderson et al., 2019).

The effect of negative emotional stimuli on cognitive processes appears to depend on task demands and the balance between bottom-up emotional processing and top-down cognitive control. One line of research suggests that negative emotions enhance memory-guided attention by increasing activity in frontoparietal, insular, limbic and parahippocampal regions, indicating aheightened allocation of attentional resources to task-relevant stimuli in visual search tasks (Brosch et al., 2013; Pedale et al., 2019; Salsano et al., 2024). In this context, the interference from negative emotional stimuli acting as distractors can be resolved through enhanced connectivity between the insular cortex and prefrontal regions, along with increased heartbeat-evoked responses, which have been shown to be modulated by emotional arousal (Dolcos and McCarthy, 2006; Anticevic et al., 2010; Kim et al., 2019; Pedale et al., 2019). However, this cognitive effort, along with increased fronto-parietal activity, may be accompanied with more conservative judgments and reasoning in the conditions with negative emotional stimuli (e.g., sad faces) compared to those with positive emotional stimuli, which tend to facilitate cognitive control processing (Calvo and Nummenmaa, 2008; Yuan et al., 2023). While sad stimuli weaken causal perception and reduce valence ratings in contingency judgments, leading to more conservative judgments (Young et al., 2020), happy faces offer an advantage in tasks like visual search (Calvo and Nummenmaa, 2008) and associative learning (Evans et al., 2010), including probabilistic tasks where they enhance reward processing and motivation, encouraging more exploratory behaviour, even when no clear contingency exists (Averbeck and Duchaine, 2009; Saylik et al., 2021). The processing of happy faces may enhance contingency judgment processes and reduce interference in frontoparietal activation, while negative emotional stimuli often disrupt attention allocation, leading to stronger activations (Dolcos and McCarthy, 2006; Anticevic et al., 2010; Pedale et al., 2019). As the processing of zero and negative contingencies requires more attentional resources, this effect could be more pronounced in such conditions. Taken together, these findings suggest that emotion-related brain regions interact with traditional associative learning networks, with emotional valence modulating attention allocation and contingency evaluations (Furl et al., 2012; Keskin-Gokcelli et al., 2024; Wang et al., 2024).

The current study aims to extend the investigation of contingency learning by examining how different types of contingencies (negative, zero, and positive) are processed when associated with happy and sad facial expressions. Utilizing a streaming trial procedure adapted from Crump et al. (2007) and employing functional magnetic resonance imaging (fMRI), we investigate the neural correlates underlying the processing of these emotional contingencies. In our study, we use two distinct types of stimuli to evoke emotional responses and explore their effects on learning processes. Negative stimuli are represented by sad facial expressions, which may be associated with uncertainty and increased arousal levels (Tiedens and Linton, 2001; Anderson et al., 2019). In contrast, positive stimuli are represented by happy facial expressions, which may be linked to certainty and approachability (Tiedens and Linton, 2001). Emotional expressions appear to affect judgment, reasoning and decision-making process due to various factors, including mood, psychological disorders, pre-existing mental models of emotional transitions leading biases, as well as valence of emotions (Blanchette and Richards, 2010; Brosch et al., 2013; Thornton and Tamir, 2017; Matute et al., 2019; Barrick et al., 2024). These happy and sad stimuli are integrated with three types of contingency relations (i.e., positive, zero, negative) to examine the functional neuroanatomical correlates of contingency judgment in neurotypical participants. We hypothesize that contingencies involving sad facial stimuli will lead to a weaker perception of causality and will more robustly engage frontal brain regions in negative and zero contingency tasks compared to positive contingencies. This reflects the increased cognitive demands required for processing negative and zero contingencies during the judgment process.

We explored functional neuroanatomical correlates of emotional facial expressions (happy and sad) during different contingencies (negative, zero, and positive). Overall, our results revealed specific effects where emotional stimuli modulate the perception of causality, particularly in zero contingency conditions. Behavioural results indicated that participants perceived slightly negative sense of causality when judging zero contingency conditions with sad faces, while they perceived a slightly positive sense of causality happy faces were involved. This pattern of results was consistent across both early and late trials of the zero contingency task. Furthermore, participants perceived a weaker sense of causality in negative contingency conditions with sad faces during early trials, with this effect diminishing in the late trials.

These behavioural findings were accompanied with increased activity in bilateral lateral frontal regions, ACC and additional subcortical areas. In the early trials, sad faces during zero contingencies, compared to happy faces, led to heightened activation in broader lateral frontal regions, including the anterior cingulate and temporal regions (such as the superior temporal gyrus (STG)), which extended into the insula, thalamus, and putamen. These activations continued into the late trials, involving deeper structures like the posterior cingulate cortex (PCC), thalamus, and precuneus. Similarly, negative contingencies with sad faces resulted in increased activation in medial frontal regions, extending into the superior frontal gyrus and motor-related areas during early trials. In contrast, participants’ ratings in positive contingency conditions remained consistent across trials, with similar behavioural and neural responses showing no significant differences between happy and sad face conditions. Our findings align with previous research addressing these anatomical regions involved in contingency learning (Tsukiura et al., 2003; Liljeholm et al., 2011; Morris et al., 2022; Saylik et al., 2022) and the processing of negative emotional stimuli (Bush et al., 2000; Kanske and Kotz, 2011; Pedale et al., 2019; Salsano et al., 2024). Moreover, our study provides further insights into how emotional stimuli modulate social interactions and judgment processes across contingency learning tasks.

The activation of the ACC and IFG is well-documented, highlighting their roles in conflict resolution, adaptive decision-making, the processing of negative emotional valence, and the handling of environmental dynamics like volatility and uncertainty (Aron et al., 2004; Aron, 2007; Behrens et al., 2007; Peters et al., 2017; Stolyarova et al., 2019; Klein-Flügge et al., 2022). The ACC, as part of a distributed network that includes lateral frontal areas, is particularly active during the modulation of causal perception in uncertain situations, such as zero contingency tasks and in response to negative emotional stimuli (Bryden et al., 2011; Palomero-Gallagher et al., 2019; Stolyarova et al., 2019). The ACC connects to lateral and medial frontal regions (e.g., IFG, SFG, MedFG), forming cognitive networks with subcortical areas (e.g., putamen, amygdala, hippocampus, and thalamus) anchored in the entorhinal cortex (Reznik et al., 2023; Syversen et al., 2024). This network, with the temporal gyrus acting as a hub, integrates and updates information from both cognitive and emotional domains (Krug and Carter, 2011; Spiers et al., 2017; Stolyarova et al., 2019; Horibe, 2024). In this context, regions such as the insula, which is involved in emotional salience, and the putamen, which is sensitive to the accumulation of various types of information, may play a crucial role in bridging the gap for making perceptual decisions in tasks where learning associations between stimuli is crucial (Behrens et al., 2007; Li et al., 2020; Salsano et al., 2024). Thus, the broader activation observed in the IFG and ACC, extending into temporal and subcortical regions, may reflect the increased cognitive demands of processing uncertain or ambiguous contingencies, particularly those involving negative emotional stimuli in zero contingency tasks.

Understanding the contingency-specific effects of emotional stimuli on causal perception requires examining both the characteristics of the emotional stimuli and the type of contingency involved. The altered perception of causality in zero contingency tasks with sad faces, compared to happy faces, likely arises from the uncertainty inherent in both sad faces and zero contingency conditions. Uncertain situations heighten stress and arousal, demanding significant cognitive resources related to attention and learning in order to resolve ambiguity (Peters et al., 2017). Zero contingency tasks, characterized by the absence of predictive relationships between cues and outcomes, represent a form of uncertainty that can induce stress and require substantial cognitive and neural resources to manage (Behrens et al., 2007; Matute et al., 2015, 2019; Msetfi et al., 2015; Stolyarova et al., 2019). This inherent ambiguity makes the judgment of zero contingency tasks particularly challenging, often triggering cognitive biases and mental shortcuts that hinder the accurate perception of cue-outcome relationships (Tiedens and Linton, 2001; Blanchette and Richards, 2010; Matute et al., 2015; Stolyarova et al., 2019).

Consistently, sad emotional stimuli are associated with uncertainty, negativity, and stress-related arousal (Tiedens and Linton, 2001; Calvo and Nummenmaa, 2008; Yuan et al., 2023). These features of negative emotional stimuli, such as sad faces, can trigger task-irrelevant mental activities that interfere with task-related processes, requiring greater effort during challenging cognitive tasks (Bush et al., 2000; Dolcos and McCarthy, 2006; Kanske and Kotz, 2011). In contrast, happy faces are the only positive expression among the six basic emotions, making them rewarding, pleasant, and uniquely distinguishable both in physical appearance and psychological impact when compared to other basic expressions (Calvo et al., 2018; Barros et al., 2023). Various studies highlight the pronounced salience of happy faces, demonstrating higher visual contrast in tasks like search and detection, speed-accuracy response systems, and machine learning analyses, which facilitate the allocation of attentional resources (Calvo et al., 2016, 2017, 2018; Calvo and Marrero, 2009; Calvo and Nummenmaa, 2008; Stuit et al., 2023). Consistent with our results, the distinct features of emotional expressions suggest that emotional content can lead to an overestimation or underestimation of the associative value between a cue and an outcome, particularly in zero-contingency situations (Averbeck and Duchaine, 2009; Evans et al., 2010).

We can speculate possible reasons why facial expressions differentially impact learning in zero contingency tasks, drawing on contingency learning and emotion-related theoretical frameworks. We have discussed how the characteristics of the faces and zero-contingency tasks might influence causal perception. One perspective might reflect on the previous predictability of different facial expressions interacting with saliency and attentional system. Mackintosh (1975) proposed a theory of selective attention, suggesting that cue validity depends on both physical salience and reliability of cues as predictors of outcomes. According to this theory, we focus on task-relevant cues while ignoring irrelevant ones, with more salient cues drawing attention and being learned more easily (Aisbitt and Murphy, 2016). This idea aligns with the attention-narrowing hypothesis, which suggests that negative emotions interfere with cognitive processes by heightening arousal (which leads to task-irrelevant mental activity) and narrowing attention to specific details during decision-making (Easterbrook, 1959; Wichary et al., 2016; Saylik, 2017).

An alternative view suggests that perception of causality may be driven by predictive value than actual value in zero contingency situations, particularly when affective features are involved (Yarritu and Matute, 2015; Matute et al., 2019). As a result, individuals are more likely to perceive connections between unrelated events based on emotions, prior beliefs, expectations, mood, or emotional cues (Thornton and Tamir, 2017; Anderson et al., 2019; Matute et al., 2019; Barrick et al., 2024). This aligns with the availability heuristic hypothesis, which proposes that emotions shape judgments by enhancing access to mood-congruent information (Blanchette and Richards, 2010). Recent research supports this idea, showing that participants are more likely to judge zero-contingency tasks negatively when exposed to undesirable stimuli, such as sad faces. In contrast, desirable stimuli, like happy faces, lead participants to overestimate outcomes and provide more positive ratings (Fulcher et al., 2001).

In summary, applying this to our findings, pairing sad stimuli with zero-contingency tasks may increase uncertainty and stress-related arousal, requiring more cognitive resources and strategies, which in turn increases activity in cognitive control areas. In contrast, happy expressions-associated with approachability and stronger saliency-may serve as more reliable predictors of causal outcomes, thereby facilitating associative processes.

Although negative contingency generally did not show significant behavioural or neural differences across the stimuli manipulation, an analysis of the early trials revealed that sad faces elicited a weaker perception of causality compared to happy faces. This was accompanied by increased activity in frontal areas (MFG, SFG and ACC), which was similar to the activation patterns observed in zero-contingency tasks. It appears that forming predictions was easier when contingencies were paired with happy faces, but more demanding when paired with sad faces, particularly during the early trials. One explanation for this is that forming predictions in negative contingencies is more challenging than positive ones. This is because it is harder to detect a cue signal when it is not followed by an outcome, and vice versa. Therefore, the observed divergence in early trials between sad and happy faces is likely not due to uncertainty effects in zero-contingency tasks but rather the increased cognitive load associated with processing negative emotional stimuli, combined with the challenging nature of negative contingencies (Maldonado et al., 2006; Heisz et al., 2011; Saylik et al., 2022). The diminished effect observed in the late trials seems to be related to increased certainty, as reliable information accumulates over time. This suggests an adaptation or learning process, where participants adjust their judgments as they gather more information (Mackintosh, 1975; Behrens et al., 2007; Murphy and Castiello, 2024). Therefore, the divergence in the early trials and the convergence in the late trials indicates that inferring cue-outcome relationship is based on the total accumulated information (Mackintosh, 1975; Behrens et al., 2007). Thus, the biased perception in happy and sad face conditions reduces after experiencing sufficient number of trials (Murphy et al., 2011). That is to say, in the early trials of a contingency participants are still forming their judgment along with limited information to infer cue-outcome relationships while sad faces may interfere and happy faces facilitate the process, however as the trials progresses to the end, they receive more information and adjust their judgments (Murphy et al., 2011).

Finally, our findings integrate well with theoretical models of contingency learning and emotional processing. The Delta P (∆P) rule, as introduced by Allan (1980), provides a robust framework for understanding how individuals learn and discriminate between different types of contingencies. Our results extend this model by incorporating the influence of emotional valence, demonstrating that emotional stimuli engage distinct neural circuits and modulate the learning and evaluation of contingencies.

The current study has several limitations that should be addressed in future research. First, we used fixed Delta p values (0.50, 0, −0.50) with a limited number of trials and ratings (16 trials per condition with 8 repetitions/ratings) and only two emotional expressions (happy and sad). Previous research indicates that biased perceptions of causality tend to disappear after experiencing a sufficient number of trials, with early biases diminishing over time (Murphy et al., 2011; Matute et al., 2019). While we observed this effect in the negative contingency, the divergence persisted in zero-contingency tasks through stimuli manipulation. Second, although we applied some exclusion criteria (e.g., participants’ age, gender, and psychiatric history), certain individual differences, such as alexithymia, were not considered. Future studies should explore varied contingency settings (e.g., longer trials, increased ratings per condition) and a broader range of emotional expressions (e.g., anger, fear) while examining the long-term effects of emotional contingencies on social interactions and decision-making, taking individual traits like alexithymia and neuroticism into account (Saylik et al., 2018; Cuve et al., 2021). Third, in our analysis, we applied an FHWE correction threshold of 0.05 but when this is not significant, we examined uncorrected significant thresholds at the cluster level (e.g., uncorrected at 0.001). Particularly, for the analysis of the first and second trials, we adopted more liberal thresholds due to the reduced number of trials. Future research should take this into account when interpreting our findings.

In conclusion, the findings suggest that sad stimuli heighten cognitive demands in tasks with zero or partially negative contingencies eliciting weaker perception of causality and activating contingency-related brain areas more than happy stimuli. This underscores the intricate relationship between emotional processing and causal learning, with facial expressions influencing causal perception, potentially overriding more rational judgments (Furl et al., 2012). The results suggest that certain emotional experiences have a stronger influence on shaping the judgments. Our findings integrate well with theoretical models of contingency learning and emotional processing. The Delta P (∆P) rule, as introduced by Allan (1980), provides a robust framework for understanding how individuals learn and discriminate between different types of contingencies. Our results extend this model by incorporating the influence of emotional valence, demonstrating that emotional stimuli engage distinct neural circuits and modulate the learning and evaluation of contingencies.