Is she still angry? Intact learning but no updating of facial expressions priors in autism

Autistic people exhibit atypical use of prior information when processing simple perceptual stimuli; yet, it remains unclear whether and how these difficulties in using priors extend to complex social stimuli. Here, we compared autistic people without accompanying intellectual disability and nonautistic people in their ability to acquire an “emotional prior” of a facial expression and update this prior to a different facial expression of the same identity. Participants performed a two‐interval same/different discrimination task between two facial expressions. To study the acquisition of the prior, we examined how discrimination was modified by the contraction of the perceived facial expressions toward the average of presented stimuli (i.e., regression to the mean). At first, facial expressions surrounded one average emotional prior (mostly sad or angry), and then the average switched (to mostly angry or sad, accordingly). Autistic people exhibited challenges in facial discrimination, and yet acquired the first prior, demonstrating typical regression‐to‐the‐mean effects. However, unlike nonautistic people, autistic people did not update their perception to the second prior, suggesting they are less flexible in updating an acquired prior of emotional expressions. Our findings shed light on the perception of emotional expressions, one of the most pressing challenges in autism.


Lay Summary
The ability to flexibly adapt to changing emotional context is critical for social interactions and is one of the most pressing challenges in social functioning of autistic people.For instance, a mother of an autistic child, who is angry at her son, even after the anger has passed, may hear her son asking, "Are you still angry?",reflecting the importance of understanding, adapting, and conveying her emotional state and the level of uncertainty in the child's own ability to read out the change of her emotional state.Although the perception of facial expressions is difficult for autistic people, it is unclear how this difficulty may impact the ability to adapt to changing emotional context.Here, we examined this ability in autistic and nonautistic people, specifically testing how discrimination was modified by the emotional context formed by the presented facial expressions.First, participants were exposed to expressions surrounding one emotion (e.g., mostly angry expressions).Then, the facial expressions shifted to surround a different emotion (e.g., to mostly sad expressions).We show that although autistic people exhibit difficulties in discriminating between facial expressions, the perception of both groups was influenced to a similar degree by the exposure to the first emotional

INTRODUCTION
The ability to adapt to changes in the emotional expressions of others is at the core of social communication skills and is likely challenging for autistic people, who are characterized by difficulties in social interactions.Although people often change their mood and corresponding facial expressions, whether and how people, and especially autistic people, perceive these ongoing changes in facial expressions remains unclear.
Prior research has demonstrated difficulties in face processing for autistic people, a robust phenomenon that is likely related to core social difficulties in autism (Dawson et al., 2005;Gepner et al., 1996;Hartston et al., 2023).Autistic people have atypical processing of faces (Ciaramidaro et al., 2018), with more pronounced autistic traits associated with difficulties in processing face identity and facial expression (Kovarski et al., 2019;Kuusikko et al., 2009;Neuhaus et al., 2016).While there are no significant differences in performance between autistic and nonautistic people in simple facial expression identification tasks (Homer & Rutherford, 2008;Humphreys et al., 2007), differences arise for more complex settings such as for ambiguous and complex facial expressions (e.g., blended emotion consists of surprise and fear (Fridenson-Hayo et al., 2016;Humphreys et al., 2007;Jolliffe & Baron-Cohen, 1997;Kuusikko et al., 2009)).Autistic people experience difficulties at recognizing negative facial expressions (Ashwin et al., 2006;Corden et al., 2008), tend to confuse them (Lacroix et al., 2009), and show over-attribution of negative emotions to neutral expressions (Eack et al., 2015).
The difficulties in processing faces may be related to broader changes in perception in autism.A growing body of research points to alterations in autistic perception across different sensory modalities and tasks (Dakin & Frith, 2005;Hadad & Yashar, 2022).Under the Bayesian framework, perception entails an inference process given uncertain sensory information.The inference combines the likelihood, the probability distribution of observing a certain set of sensory data given a hypothesis (e.g., it quantifies how likely different stimuli are, given the observed sensory evidence), and the prior, a probability distribution of one's initial assumptions or expectations about the state of the world.Thus, perception is shaped both by one's current sensory information (i.e., likelihood), and by one's predictions about the probability of a stimulus in the environment based on previous experiences, learned knowledge, or even innate biases (i.e., prior).Research has considered the forming of priors in different ways, such as throughout a lifetime exposure, preexisting priors (Hadad & Schwartz, 2019;Langer & Bülthoff, 2001;Sun & Perona, 1996), or quickly as one adjusts their priors to adapt to ongoing changes in the environment (Raviv et al., 2012).
Early accounts that examined perceptual alteration in autism under the Bayesian framework have proposed that autistic people rely more on sensory information (i.e., likelihood) because their internal priors are wider (Pellicano & Burr, 2012).These have been thought to lead to various effects such as attenuated priors, which bring perception closer to the actual input (Lawson et al., 2017), a sense of overflow of sensory information (Pellicano, 2013), and reduced adaptation (Pellicano et al., 2007;Ropar & Mitchell, 2002).Another option is that reduced effects of the priors may arise from narrower likelihoods (i.e., more precise sensory representations; see illustration Brock, 2012).However, autistic people are shown to use both preexisting priors and learned priors in a manner similar to that of nonautistic people in several perceptual tasks involving discrimination of simple features such as size and weight (Binur et al., 2022;Hadad & Schwartz, 2019), suggesting no attenuated priors or enhanced sensory sensitivity in autism (Hadad & Yashar, 2022).Instead, recent studies have proposed that autistic people can form priors and integrate them with sensory information; yet, they tend to rely less on recent experiences, and are less flexible in adjusting their priors (Lieder et al., 2019;Sapey-Triomphe et al., 2021;Vishne et al., 2021).There may also be differences in the use of preexisting priors versus learned priors in autism (see Angeletos Chrysaitis & Seriès, 2023 for a recent review).
In several studies focusing on how priors are formed and used, researchers have examined the effect of regression to the mean-the extent to which perceptual judgments are biased toward the mean stimulus value (e.g., the prior) given a range of values presented to the observer (Petzschner & Glasauer, 2011).When examining pitch discrimination, although regression to the mean was similar for autistic people and nonautistic people, nonautistic people had a stronger bias toward recently presented features (i.e., information presented one to four trials back) than autistic people (Lieder et al., 2019).These results have been suggested to demonstrate intact learning of "long-term" priors formed throughout the experiment in autism (e.g., regression to the overall mean values), but less so for "short-term" priors, which are formed based on the last few trials (Lieder et al., 2019).Similarly, in a study that examined how observers synchronized their finger tapping to the tempo of a metronome, autistic people showed reduced error corrections across consecutive beats, suggesting a slower update rate to the presented tempo (Vishne et al., 2021).In another study, autistic people demonstrated the ability to learn priors in an associative learning task in which there was a contingency between an auditory stimulus (a beep) and a visual stimulus (orientation), yet the magnitude of this prior was larger in nonautistic than autistic people (Sapey-Triomphe et al., 2021).When the association was reversed, the autistic participants showed a reduced prediction effect compared with the nonautistic participants (Sapey-Triomphe et al., 2021).These results suggest an inflexibility in updating priors in autism in basic perceptual processes.
However, most studies have examined priors in simple perceptual stimuli (e.g., the perception of tones or time, houses vs. faces).Few studies have examined the use of priors in the context of socially related stimuli and findings are mixed.For example, in the perception of gaze, the preexisting prior (i.e., the expectation that a gaze is often directed toward the observer) is typical in autistic people and in those with higher autistic traits in the nonautistic population (Pell et al., 2016).Yet, learned associations between gaze direction and reward were different in nonautistic individuals with high autistic traits (Sevgi et al., 2020).In the perception of motor actions, the use of contextual learned priors in predicting others' action (e.g., associations between color and action or the overall probability of an action in the experiment), is reduced in autism (Amoruso et al., 2019), and vary with the severity of the clinical characteristics (Chambon et al., 2017).These studies tested the integration of nonsocial priors (e.g., color cues or general probabilities) and social relevant information.Social functioning, however, often relies on priors formed directly based on the perception of dynamic social information such as intonation, verbal information, gestures, and facial expressions.The ability to perceive and adjust to continuously changing social information such as facial expressions is especially central to our understanding of social-related behavior.
Here, we examined whether and how autistic people acquire priors in the perception of facial expressions of a single person (i.e., "emotional prior", e.g., mostly angry expressions), and the extent to which they can adjust this acquired prior to a new prior (mostly sad expressions).This allowed us to isolate and examine how implicit priors are formed based on the exposure to changes in facial expressions, where, as is the case in many natural social scenes, no explicit instructions or cues are given.
To uncover the mechanisms that underlie the forming and updating of "emotional priors", we focus on processing negative emotional expressions, which has been demonstrated as especially difficult for autistic people (Ashwin et al., 2006;Corden et al., 2008;Eack et al., 2015;Lacroix et al., 2009).Participants performed same versus different judgments in a serial discrimination task of two consecutive faces.The two faces were drawn from a distribution spanning two morphed facial expressions of the same identity.We examined utilization of priors using the effect of regression to the mean, the tendency of perception to be biased toward the mean value of a range of values the observer experienced (Petzschner & Glasauer, 2011;Lulav-Bash et al., 2023).We exposed participants to different statistical distributions in two segments.To test prior acquisition, in the first segment of the study, we presented a distribution of expressions surrounding a particular facial emotion (e.g., mostly sad).To test prior updating, in the second segment of the study, the distribution shifted to a different expression of the same facial identity (e.g., mostly angry).Learning and updating the prior according to the accumulated distribution would be manifested in regression toward the first mean (e.g., the perception of facial expressions biased toward sad) and updating it to the second mean (e.g., biases toward angry).
This experimental design allowed us to address two main questions.First, whether autistic people can acquire priors of facial expression based on information accumulated throughout the experiment.On the one hand, there may be reduced utilization of prior information in autism, based on the attenuated prior account of atypical perception in autism (Lawson et al., 2017;Pellicano & Burr, 2012).On the other hand, there may be typical or even stronger reliance on priors in autistic people because of their difficulty in processing facial expressions (e.g., Fridenson-Hayo et al., 2016), and of negative expression in particular (Ashwin et al., 2006;Corden et al., 2008;Eack et al., 2015;Lacroix et al., 2009).This may lead to stronger reliance on priors, consistent with general Bayesian principles of perception (Petzschner et al., 2015).
Second, we ask whether autistic people can update the acquired prior to the new prior presented in the second segment of the experiment, when the mean of the emotional expressions shifted (e.g., from mostly angry to mostly sad).If autistic people are inflexible in updating to new information, this would be manifested in a difficulty updating to the second prior, consistent with evidence demonstrating inflexibility in adjusting to changes in the environment in basic sensory perception (Lieder et al., 2019).

Participants
A total of 26 nonautistic participants (14 males; mean age = 26.9,range 19-39, SD = 5.3 yrs.), and 26 autistic people without accompanying intellectual disability (21 males 1 ; mean age 27, range 18-37, SD = 5.1 yrs.) participated in the study.Specific data on socioeconomic status and anxiety levels were not recorded.The participants recruited for the experiment were between the ages of 18-40, with normal or corrected-to-normal vision.Nonautistic participants had no known diagnosis of any neurodevelopmental condition.The experiment was approved by the Ethics Committee of the Faculty of Education at the University, and participants signed an informed consent form.Participants consisted of a convenience sample and were recruited via advertising ads and via ads distributed at autistic housing facilities and associations and were compensated for their time.Six participants were excluded from analyses: one nonautistic and two autistic participants did not show improvement in performance as the range difference between the two faces increased (notice the difference better), and for three other participants (one nonautistic and two autistic participants), performance was below chance across ranges.Diagnosis of autism was determined for our participants during early childhood but was confirmed in the lab using Autism Diagnosis Observation Schedule (ADOS, Lord et al., 2000).Participants in the autistic group underwent a nonverbal IQ test [TONI-4 test] (Brown et al., 2010) in the lab, all scored within the typical range (>80).Participants in both groups filled an autism-spectrum quotient (AQ) test (Baron-Cohen et al., 2001), to measure the expression of autismspectrum traits.The AQ scores were used to screen out those nonautistic participants who may exhibit autistic traits, yet no nonautistic participants were screened out.The average AQ scores were significantly higher in the autistic group (mean = 26.04;SD = 8.81) than in the nonautistic group (mean = 17.04;SD = 6.75), t(50) = 4.13, p = 0.04.Because we began data collection during the Covid-19 pandemic, some participants performed the study (the research task and the AQ questionnaire only) at home (nonautistic group: n = 12; autistic group: n = 1).We did not find any significant differences between participants according to test site (home vs. lab; see the Data S1, test site analysis).A G-power analysis showed that 24 participants were sufficient for revealing a significant regression effect in the autism group (based on the effect size at Hartston et al., 2023) and a power of 0.8.

Stimuli
Face stimuli were created from two parent facial expressions, an angry and a sad facial expression of the same identity from the compound facial expressions of emotion (CFEE) database (Du et al., 2014).Throughout the entire experiment only one identity appeared; another identity appeared in the training part that preceded the experiment.The two identities were female Caucasians.For these two images, 100 steps continuum of morphed faces were created by combining a different percentage of two-parent faces (e.g., 1% sad and 99% angry, and so on; see Figure 1b), using Fanta Morph software.In the lab, the faces were presented on a 27 00 inches monitor with 1920 Â 1080 resolution and subtended approximately 11 Â 9.5 .

Trial sequence
The experiment was designed in E-Prime 3.0 software.Each trial began with a 1000 ms fixation cross at the center of the screen.Then, the first face appeared for 750 ms followed by a 500 ms blank screen and a second face appearing for 750 ms.Participants' task was to determine whether the two faces were the same or different by pressing a button (see Figure 1a).Half of the trials consisted of identical stimuli (e.g., same) and half consisted of different stimulus (i.e., different).The trial ended when the participants responded or after 5000 ms had elapsed.In half of the trials, the two faces were identical, and in the other half, the two faces differed by either 9%, 13%, 17%, 20%, or 25% morph steps from each other with an equal probability (see Figure 1b).The order of presentation of the pairs of faces was randomized within and between participants (i.e., the combinations of same and different-regression facilitates/hinders, and all ranges were random).There were two segments of trials in the experiment that were counterbalanced in order.The segments consisted of two blocks of 240 trials each.Each segment surrounded a mean morph corresponding to either mostly sad (30% morph) or mostly angry (70% morph).See details below ("Prior manipulationssegment I and segment II").

Task
On each trial the participant's task was to indicate with a button press whether the two stimuli in the sequence were same or different.No instructions were given regarding the identification of the emotional content of the facial expressions.There was equal probability for the two faces to be same (50%) or different (50%).Instructions were given before the experiment (i.e., same instructions throughout the experiment).
Regression to the mean Among the "different" trials, there were two types of equally likely trials that consisted of the same pair 1 Analysis of only males in each group demonstrated the same pattern of results as in the main findings (see Data S1).
of faces in two different presentation orders (see Figure 1c).In the "Bias+" trials (regression facilitates), the value of the first face in the sequence was closer to the distribution average than the second face (e.g., for a distribution average of 30%, a first face with a value of 23 and a second with a value of 10), and in the "BiasÀ" (regression hinders) trials the order was reversed (e.g., the first face is 10 and the second is 23).Thus, each pair of "different" faces appeared twice, once as a "Bias+" trial and once as a "BiasÀ" trial."Bias+" and "BiasÀ" trials in different ranges were presented in random order.Learning the distribution of expressions and utilizing the average as a prior should manifest in the attraction of the first face in the sequence to the mean (Ashourian & Loewenstein, 2011).This is believed to occur because the first stimulus in the sequence must be held in memory for longer than the second stimulus before making the decision and thus it is more prone to being integrated with prior expectations (i.e., the average) compared with the second stimulus.This effect of regression to the mean is implicit and does not depend on instructions (e.g., Lieder et al., 2019).When regression to the mean occurs, performance is expected to be higher for "Bias+" trials compared with "BiasÀ" trials because the perceived distance between the two faces is larger for "Bias+" trials (i.e., the first face is perceived as closer to the average and thus further away from the second face; see Figure 1c).In contrast, comparable performance in "BiasÀ" trials and "Bias+" trials suggests no regression to the mean (i.e., no effect of the prior on perceptual judgments).The pairs of faces were selected such that the two faces never crossed the distribution mean (e.g., 30 for the sad prior or 70 for the angry prior).In sum, half of the trials consisted of same stimuli (50% of overall trials) Regression to the mean for different trials is manipulated by the order of presentation of two sequentially presented faces.The first face in the trial sequence is attracted to the mean (i.e., the prior; horizontal line).Thus, the perceived difference between the two faces may either increase ("Bias+") or decrease ("BiasÀ"), resulting in facilitating or hindering performance, respectively.(d).The two segments of the experiment with the different prior distributions.In segment I, participants were exposed to expressions distributed around the first prior (sad or angry; counterbalanced).In the second part, the distribution of expressions shifted to a second prior (angry or sad, respectively).Examining performance to stimuli at the center, the "switch stimuli," enables testing the acquisition of the prior (segment I) and the flexibility in updating the prior (segment II) because these trials should switch in the direction of regression in each segment.and half were the different stimuli; among the different stimuli the face pairs were equally likely to consist of "Bias+" trials (25% of the overall trials) and "BiasÀ" trials (25% of overall trials).

Prior manipulations-Segment I and segment II
To examine the acquisition and updating of priors, the study consisted of two segments that differed in the means of the distribution of faces presented across trials in each segment (see Figure 1d).The two segments were separated by a short break.Manipulating the means in each segment allowed us to examine how perceptual judgments are affected by regression to the mean in each segment.In the first segment, the mean morph value was 30 (or 70) and face morphs ranged from 1 to 70, whereas in the second segment the mean morph value was and 70 (or 30, respectively), and face morphs ranged from 30% to 100%.The order of means was counterbalanced between participants.The distribution of faces was approximately Gaussian; thus, half of the trials were below the average and half were above the average.The exact mean (i.e., the mean of the values of the faces presented) for the sad prior was 30.849 (SD = 16.835), and for the angry prior it was 69.242 (SD = 16.913).Each segment consisted of two blocks of 240 trials each and there was a short rest period between segments ($5 min).
To examine the acquisition and updating of the "emotional prior", we analyzed "different" trials at the center of the distribution (i.e., with values between 30 and 70; Figure 1d), which we term "switch stimuli."These faces, between 30 and 70, appeared in both segments of the experiment for all participants.Critically, for these trials, the regression effect is expected to switch due to a change of the mean between segments, allowing us to examine both acquisition of the prior in the first segment, and updating of the prior in the second segment.In contrast, trials with face morphs at the edges (i.e., 0-30 or 70-100; "non-switch stimuli"), are expected to be attracted to the same direction in the first and second segment (see Figure 1d).

Data analyses
Data were summarized and analyzed using IBM SPSS Statistics Version 21 (SPSS, 2013) and R-4.1.0for Windows.The main analysis focused on the regression effects, manifested in differences in accuracy between the "Bias+" and the "BiasÀ" different trials.To examine prior manipulations, we focus the analysis on "switch stimuli" (see "Methods-Prior manipulations").Repeated measures ANOVA was carried out on accuracy measures (i.e., accuracy and d-prime) and reaction times, with segment (first vs. second), regression ("BiasÀ" vs. "Bias+"), and morph range (9%, 13%, 17%, 20%, 25%) as within-subject factors, and group (autistic vs. nonautistic) as a between-subject factor.We present the analysis for accuracy rates (see Data S1 for d-prime and criterion analysis).

Community involvement
We collaborate with our autistic participants in studying autism through workshops conducted at our "Special Populations Advanced Research and Clinic Center" (SPARC; https://sparc.haifa.ac.il/en/).Questions arising at these workshops inform inquiries in the lab.Results and their implications are discussed with our autistic collaborators to further the translational implications of lab findings to the field.
One possible explanation for the difference between the autistic group and the nonautistic group in the second segment is that the autistic group acquired a prior of an overall mean that included all faces presented in the experiment (e.g., angry and sad; an overall mean of 50), rather than not updating to the second prior.To examine this possibility, we inspected the regression effect in the second segment, for trials in which the direction of regression should have switched if the acquired prior was the overall mean rather than the mean of the second segment (e.g., if the second prior was 30 and an overall prior of 50 was acquired, trials with face morphs between 30 and 50 should have been attracted to 50 rather than to 30, thus "Bias+" and "BiasÀ" should have the opposite effect).Yet, we did not find any significant difference between "Bias+" and "BiasÀ" for the autistic group (Figure 2d).In sum, there was no evidence that the autistic group acquired the second prior, acquired an overall prior, or remained with the first prior.
To rule out the possibility that fatigue or lapses of attention in autistic individuals limited their ability to demonstrate any type of regression effects in the second segment, we examined the nonswitch stimuli in the second segment.For these trials, the direction of the regression effect should be similar to that of the first segment (that is the regression should not switch when the means of the distribution switch; Figure 1d), in contrast to switch stimuli (see "Methods").If participants were unable to exhibit any regression effect in the second segment, this should be apparent also for the nonswitch stimuli.Thus, we performed the same analysis as for switch stimuli for nonswitch stimuli.We analyzed accuracy using a repeated measures ANOVA and found a significant interaction between group, segment, regression, range, and switch type, F(1,100) = 12.093, p = 0.001, η 2 p = 0.11.As expected, in the first segment, the interaction between switch type, regression, and group did not reach significance, F(1,50) = 3.48, p = 0.068, η 2 p = 0.06, suggesting no significant differences between switch and nonswitch stimuli.A significant main effect of regression F(1,50) = 60.09,p = 0.001, η 2 p = 0.55, indicated regression effects across switch types for both groups.In contrast, in the second segment, there was a significant interaction between switch type, regression, and group, (F(1,50) = 9.44, p < 0.001 = 0.003, η 2 p = 0.16).For the nonautistic group, regression effects were significant F(1,25) = 20.79,p < 0.001, η 2 p = 0.45, and did not interact with switch type, F(1,25) = 2.625, p = 0.118, η 2 p = 0.09.In contrast, for the autistic group, significant differences in regression effects were found between switch and nonswitch stimuli, (F(1,25) = 7.06, p = 0.014, η 2 p = 0.22), that were driven by a significant regression effects for nonswitch (F(1,25) = 18.01, p < 0.001, η 2 p = 0.4), but not for switch stimuli (F(1,25) = 0.014, p = 0.907, η 2 p = 0.00, see above).In sum, the significant regression effect for the nonswitch stimuli indicates that priors that did not require updating were used in the autistic group in the second segment.

Type of emotion
We examined whether the regression effect differed between groups according to the type of emotion (i.e., sad vs. angry).We found that the interaction between group, regression, and emotion in the first segment and in the second segment were not significant (both Fs(1,48) < 1), indicating that the type of emotion did not significantly modulate the regression effect.

Bias magnitude in segment I versus II
We set to examine whether the magnitude of the regression effect of the first acquired prior is related to the magnitude of the regression effect in the second, updated prior.We calculated the magnitude of the regression effect by calculating the mean difference between "BiasÀ" and "Bias+" across ranges for each participant in each segment.A Pearson correlation coefficient was computed to assess the linear relationship between the magnitude of regression in the first segment versus the second segment and revealed a significant negative correlation, r = À0.3 and p = 0.03 (Figure 2c); a Fisher test revealed that there was no significant difference in the correlation between groups (p = 0.37).This suggests that a larger prior in the first segment is related to a smaller prior in the second segment.We performed an ANOVA to show that group differences in regression bias in the second segment were not accounted for by the magnitude of the regression bias in segment 1 (F(1,49) = 7.51, p = 0.009, η 2 p = 0.13).

DISCUSSION
We show reduced sensitivity to negative facial expressions in autism; however, most importantly, we show that despite the difficulties in the ability to discriminate facial expressions, autistic people (A) form facial expression priors based on accumulated experience, but (B) have difficulty updating this prior to a new one, suggesting it is hard to adjust a prior once it has been acquired.Our first finding demonstrated that like nonautistic participants, autistic participants integrated the accumulated information and were sensitive to the distribution from which the facial expressions were drawn.These findings suggest that autistic people can learn the statistics of the environment and implement them as prior knowledge.This is consistent with results from previous research that showed intact construction and utilization of prior in autism using simple visual and auditory stimuli such as visual judgment of size (Sapey-Triomphe et al., 2021;Binur et al., 2022) and weight (Hadad & Schwartz, 2019), duration judgment (Karaminis F I G U R E 2 (a).Accuracy rates (mean ± SEM) for "switch stimuli" across groups in the two segments of the study.Autistic participants showed overall lower accuracy rates in the first segment versus the nonautistic participants.Importantly, significant effects of regression were obtained for both groups in the first segment, but only for the nonautistic participants in the second segment.(b).Accuracy rates (mean ± SEM) for "switch stimuli" as a function of morph range in the two segments of the study.(c).Correlation between the magnitude of the regression bias (the difference between "Bias+" and "BiasÀ"), in the first segment compared to the second.(d).Accuracy rates (mean ± SEM) across groups in the second segment of the experiment, for the switch trials between 50 and the mean in the second segment (i.e., for a mean of 70, accuracy was calculated for stimuli between 50 and 70).The lack of the difference in the autistic group suggests that they did not learn an overall mean prior (e.g., of 50).There was no evidence for the nonautistic group acquiring the overall mean prior either, as the "Bias+" trials did not reduce performance compared with the "BiasÀ" trials, as would have been expected in such a case.et al., 2016), and pitch discrimination (Lieder et al., 2019).The acquisition of priors in the present study with facial expressions extends previous research on social priors and show priors can be implicitly formed based on accumulated social information without explicit knowledge.This suggests that overall "weak" or uninformative priors (Lawson et al., 2017;Pellicano & Burr, 2012) cannot explain perceptual alterations in autism.
The similarity between autistic and nonautistic participants in the acquisition of the first prior may seem surprising given that autistic participants exhibited difficulties discriminating facial expressions compared to nonautistic participants.Based on Bayesian principles of perception (Petzschner et al., 2015;Rahnev & Denison, 2018), reduced sensitivity (i.e., likelihood) is expected to lead to stronger reliance on priors.However, we did not observe differences between groups in the magnitude of the regression bias of the first prior, implying autistic people did not increase their reliance on the prior to compensate for noisier likelihood.This result is consistent with a recent study showing autistic people do not scale priors to sensory noise and to changes in the uncertainty of incoming input as nonautistic people do (Binur et al., 2022).
Our second finding reveals that autistic people exhibit difficulties updating the prior to the changing context.While the nonautistic group updated the formed priors to the new mean, the autistic group did not show such adjustments.These results were found regardless of emotion type (e.g., sad or angry).Notably, differences in response criteria between the groups cannot explain our findings either: analysis of d-prime and criterion showed the same pattern of results as accuracy (see Data S1).
Why were autistic people able to acquire a first prior but unable to update it?One explanation for the lack of updating may be increased lapses of attention or fatigue in the autistic group in the second segment of the study.However, results suggest that in the second segment autistic participants were engaged in the task.First, performance was modulated by task difficulty.Second, the autistic group demonstrated regression effects in the second segment of the experiment for "non-switch" stimuli that did not require updating.These findings reduce concerns regarding the impact of fatigue or lapses of attention in the autistic group might have played in the second segment.Future studies should control and possibly modulate social attention and eye movements to examine the role these might play in prior updating of facial expressions.Interestingly, the demonstration of regression effects for the nonswitch stimuli also indicates that participants did not default to learn an extreme expression prior (e.g., 100% angry), but were able to learn more nuanced priors of expressions (e.g., 70% angry).
It is possible that low-level features of the facial expressions were used during the task to discern whether the faces are the same or different.For example, adaptation to simple low-level features can impact the perception of faces (Xu et al., 2008).However, in the current study both segments had similar low-level feature configuration, but the autistic group only had a difficulty in the second segment, in updating.Our recent findings showing that regression effects are only seen for upright but not for inverted faces (Lulav-Bash et al., 2023) indicate low-level features, that are identical for upright and inverted faces, are less likely to underlie these effects for nonautistic people, but future studies should examine this in autistic people.Another thing to consider is that difference in regression effects between "switch" to "nonswitch" stimuli in the second segment may be related to the fact that "switch" stimuli are drawn from the center of the distribution between angry and sad expressions, and thus contain more ambivalent expressions.Notably, this did not prevent prior acquisition for these stimuli in the first segment of the study.
It is also possible that for autistic participants the amount of time and/or exposure that was enough to learn the first prior (and both priors for the nonautistic group), may not be enough to acquire the second prior.Our paradigm examined priors in a short time (within $10 min), similar to other studies (Raviv et al., 2012).However, prior acquisition has been investigated under different time scales varying from short time frames to long-term development and across different contexts.For example, different designs have been used to investigate prior acquisition in highly volatile environments with priors manipulated every few minutes (Lieder et al., 2019;Sapey-Triomphe et al., 2021;Vishne et al., 2021), stable environnements (Pellicano & Burr, 2012), or priors acquired throughout life (e.g., perceptual illusions (Binur et al., 2022;Ganel & Goodale, 2003;Hadad, Russo, et al., 2019;Langer & Bülthoff, 2001;Sun & Perona, 1996)).It is likely that the process of learning and updating priors differs according to the time scale and context.Future research should examine differences between a prior acquired in a short time frame versus a prior acquired over days or years and the ability to flexibly update this prior.
Our results may be related to a general inflexibility in autistic people (Hadad & Yashar, 2022).The negative correlation we observed, for autistic and nonautistic participants, between the magnitude of the regression effect in the first segment of the experiment compared to its magnitude in the second segment, suggests that flexibility can somewhat mediate the extent to which a formed prior is updated.Importantly, however, it cannot account for the differences we found between autistic and nonautistic participants; when considering the magnitude of the first prior, the difference between groups is still evident.Our results, then raise questions about how prior updating relates to the rigidity behaviors of autistic people.Inability to adjust priors has been demonstrated for simple stimuli (Lieder et al., 2019;Sapey-Triomphe et al., 2021).The tension between updating versus rigidity is evident in many real-world behaviors in social settings.For example, it has been suggested that an inability to fluidly process faces during social interactions, the social information processing speed hypothesis, may underlie the challenges in responding in real time to social information in autism (Gates et al., 2023).Inappropriate responses to social situations in autism (Baltaxe, 1977;De Marchena & Eigsti, 2016), may stem from longer adherence to prior expectations.For example, autistic people often exhibit difficulties understanding humor (Asperger, 1944), which entails multiple, incongruous meanings and sources of information (e.g., facial expressions, gestures, and context), and involves violations of expectations, which is particularly challenging in autism.Thus, prior inflexibility may be a general mechanism that impacts social interactions in autism.
Our study's design was inspired by real-world difficulties in social interactions between autistic and nonautistic people: a caretaker who might be angry at some point, may change her emotional state and expression after a while; yet she might be asked whether she is still angry by an autistic child.Both the caretaker and the child may experience difficulties interacting (i.e., the caretaker not being able to effectively convey the change in her emotion; and the child not being able to discern the current updated emotion of the caretaker).These and related interactions suggest inflexibility in behavior may impact social interactions.To control for the various factors that affect perception and processing in such scenarios, the present study examined prior acquisition and updating in lab settings and was geared toward scrutinizing the specific roles of facial expression perception and priors in autism, showing that social priors are learned but not updated.
Our study has the potential to inform autistic and nonautistic people, family members, friends, educators, and caretakers, about the difficulties autistic people might experience in processing changes in emotional expressions.Nonautistic people may consider providing more explicit cues about their emotional state, and specifically changes in emotional state (i.e., "prior change") given that these changes are especially challenging; for example, by informing partners about a change explicitly, rather than implicitly assuming the partner will notice the change (e.g., "I am not angry anymore").Previous studies have shown that explicit cues modify social behaviors in autism (Hudson et al., 2021;Keifer et al., 2020).For example, providing explicit information about social intentions is helpful for modifying perceptual processing of actions in autism (Hudson et al., 2021).Yet, it is unclear how explicit social cues modify updating in particular, and how this interacts with implicitly acquired priors.In the context of priors in facial expressions, this could be tested in a lab experiment by comparing updating when providing explicit information about the facial expression.Nevertheless, explicit cues may not always be appropriate or needed.Thus, nuance in understanding the factors that influence the processing of emotional changes in facial expressions of conversation partners is an important topic for future research.For example, explicit cues may be helpful in ambivalent expression with similar valence as in the current study, but not when transitioning to a different valence (e.g., angry to happy), which can occur in many real-life scenarios.It is likely that the nature of the change and the real-life context impacts "prior updating." Another aspect to consider in social communication is possible emotional prior "carry-over": an acquired prior for one identity (e.g., sad Mary), may carry over to a different conversation partner (e.g., Sarah who is neutral, is now perceived as sad).This could hinder social relationships and group interactions where emotions and expressions vary across individuals (Gates et al., 2023).The present study examined prior updating for a specific identity and future studies should examine how prior updating interacts across identifies and how context may influence this process.
Our findings raise questions for further research about the nature of the "emotional prior" and how it relates to social interactions and whether it is specific to facial expressions or extends to different emotional stimuli, such as speech intonations, prosody, or body gestures.Moreover, the paradigm we developed here can be used to examine processing of "emotional priors" in populations other than autism who may have difficulties in social interactions such as people with anxiety or depression.

Limitations
Our study has several limitations.The stimuli consisted of morphed facial expressions rather than naturalistic expressions.Although the current generated stimuli allowed controlled levels of discrimination, future research should consider using more dynamically and naturally elicited facial expressions (e.g., Watson et al., 2020).Another limitation of the study is the imbalance between males and females in the autistic group, however, when analyzing the male sub-groups, a similar pattern of the findings emerged (see Data S1).For nonautistic participants nonverbal IQ was not measured.Nevertheless, it seems unlikely that IQ played a role in our study, given that all participants in the autistic group were above 80.We did not collect race information directly from participants, however, to avoid any confounds with other-race effects (Goodman et al., 2007;Hadad, Schwartz & Binur, 2019), we adjusted the race of the identities in the experiment to the most frequent race in Israel (i.e., Caucasians), given that exposure to specific races underlies these effects (e.g., Wang & Han, 2021).
Given the heterogeneity among autistic people, it is difficult to draw conclusions on the entire autistic population.The study includes autistic people with ADHD and anxiety-related characteristics; it is complex to determine whether co-occurrence with these neurodevelopmental conditions, which are common in autism (Rong et al., 2021), have an effect on the pattern of results observed in the study.Future studies should examine whether and how anxiety may interact and impact the ability to learn and update facial expressions priors.

F
I G U R E 1 Experimental procedure.(a).Trial sequence.Participants viewed two faces consecutively and judged whether the faces were the same or different.(b).Examples of facial expressions from the sequence of 100 morphed faces that were created by combining the two-parent faces of the same identity, sad and angry.Below, a breakdown of the types of trials in each segment.(c).