Autism is classically associated with difficulties in pragmatic inferences, resulting in an over-literal interpretation of language. This has mostly been observed with figurative language (e.g., metaphors). In contrast, more recent investigations of another type of inference, scalar implicatures, have mostly failed to spot any difference between autistic and neurotypical individuals, raising concerns about any general claim of pragmatic difficulties in autism. However, both lines of research face issues: language demands rather than pragmatic competence might actually explain group differences on metaphor tasks, and scalar implicatures have mostly been assessed with truth judgment tasks, which might bias their results. This work aims to assess whether this contrast between metaphors and scalars can be observed within a single group of autistic children. A group of autistic children (N = 23) was compared to a larger sample of neurotypical children (N = 237), using innovative scalar implicatures and metaphors tablet tasks that address the methodological concerns raised in the literature. The autistic group showed a reverse contrast from what was expected, with poorer scalar implicature but similar metaphor comprehension, consistently at accuracy and response times levels. We discuss the possibility that, complementary to previous accounts, a dimension opposing guided to spontaneous pragmatic processes might explain this result and the challenges faced by autistic individuals in daily situations.
Opmerkingen
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Some (but not all) Pragmatic Inferences are Difficult for Autistic Children
Autism spectrum disorder (ASD, or from now on: autism) is a heterogeneous neurodevelopmental condition characterized by difficulties in communication and social interactions, and atypical or restricted behaviors, activities or interests (American Psychiatric Association, 2022). Autism is frequently associated with other neurodevelopmental conditions, especially language disorders or intellectual development disorders (Levy et al., 2010). Nevertheless, even when it is not the case, autistic children1 experience communication issues, among which they are typically thought to struggle with pragmatic inferences, resulting in an excessively literal interpretation of language. This hallmark of autism has crossed the ages from the early descriptions of Kanner (1946) or Asperger (1944) to the current diagnostic criteria (APA, 2022). A particularly literal processing of language is also reported by autistic individuals themselves (Müller et al., 2008; Wilson, 2022) and has shaped public representations of autism in mainstream culture (see, e.g., Semino, 2014).
This difference is classically observed with figurative language such as metaphors, starting from Happé's (1993) seminal study which has been replicated in several studies (e.g. Rundblad & Annaz, 2010). Recent meta-analyses confirm this difference with both children and adults and with both accuracy measures or, less frequently, response times (Kalandadze et al., 2018, 2019; Morsanyi et al., 2020). Effect sizes, however, largely vary and are influenced by group-matching strategies and task properties (Kalandadze et al., 2018; Morsanyi et al., 2020). Verbal explanation tasks typically yield larger effect sizes. Still, they are criticized for their verbal and metalinguistic demand (Pouscoulous, 2011), which confounds whether group-differences can be attributed to non-autistic differences such as variation in formal language skills (see Norbury, 2005; Kalandadze et al., 2019).
Moreover, the historical claim of a pragmatic impairment in autism (see also Surian, 1996) is challenged by the more recent investigation of other types of pragmatic inferences in autism, such as scalar implicatures. Scalar implicatures refer to the use of terms embedded in informativity scales, such that a weaker term implies the negation of a more informative term (Horn, 1989). The most stereotypical example is the use of some to deny the more informative all, although some is semantically compatible with all. Hence, when presented with underinformative sentences (as in “Some parrots are birds”), adults (more than children) are typically inclined to reject those statements (as implying that notall parrots are birds). Since the enrichment of some as meaning some-but-not-all is arguably pragmatic in nature and not logical, autistic individuals were expected to present differences with their neurotypical pairs. Yet, most attempts to observe such effect have failed, be it in children, adolescents or adults (Andrés-Roqueta & Katsos, 2020; Chevallier et al., 2010; Hochstein et al., 2017; Pijnacker et al., 2009; Schaeken et al., 2018; Su & Su, 2015). Importantly, these studies employed sentence judgement paradigms, asking participants to judge whether they agree with underinformative utterances. However, these metalinguistic tasks should not be considered as direct measures of implicatures derivations, as alternative pathways can lead to rejecting underinformative utterances (see Katsos & Bishop, 2011; Katsos, 2014; or Kissine & De Brabanter, 2023), but rather as a proxy, or as an indirect reflection of scalar implicatures. Interestingly, one study also employed a so-called ternary-judgement task (Schaeken et al., 2018), in which participants are given the possibility of an intermediate judgement between “agree” (in the form of a huge strawberry) and “disagree” (in the form of a small strawberry): they can select a medium-sized strawberry, meaning “I agree a bit”. This revealed a clear group difference: typically developing children largely preferred this intermediate response for underinformative statements, while autistic children were more likely to reject completely such sentences. Two other studies employed multiple-choice tasks and tended to observe group differences (Mazzaggio et al., 2021; Pastor-Cerezuela et al., 2018), but methodological issues raise caution on the interpretation of their results. For instance, Pastor-Cerezuela et al. (2018) mixed scalar and non-scalar inferences and did not include control items, which limits the conclusion that can be drawn on scalar implicatures specifically, or on implied meaning as opposed to language processing in general.
The contrast between different types of pragmatic inferences contributed to the fragmentation of pragmatics as a unitary competence, and led some researchers to propose a dichotomy opposing linguistic- vs social-pragmatics (Andrés-Roqueta & Katsos, 2017, 2020). Linguistic-pragmatics designate pragmatic operations which depend on linguistic competence and pragmatic knowledge, while social-pragmatics additionally require reasoning on the speaker’s mental states, i.e. theory of mind (ToM, Premack & Woodruff, 1978). Under this view, autistic individuals with typical linguistic skills are expected to perform similarly to neurotypical controls on linguistic-pragmatics, but poorer on social-pragmatics, given the assumed autistic difference in ToM processes (Happé et al., 2017). Scalars implicatures have thus been exemplified as a linguistic-pragmatics-based ability, leaning on the fact that verbal skills alone were shown to be a good predictor of such ability (Andrés-Roqueta & Katsos, 2020; Chevallier et al., 2010; Pijnacker et al., 2009). In contrast, figurative language is rather considered an instance of social-pragmatics (Andrés-Roqueta & Katsos, 2020). However, this distinction emerged from comparing the results of different studies with different samples. Comparisons of different types of pragmatic inference within a single autistic sample exist but are much more scarce (Andrés-Roqueta & Katsos, 2020; Deliens et al., 2018). Yet, the heterogeneity within the autistic population much weakens the conclusion that can be drawn upon comparison of different tasks in different samples.
School-age appears as a relevant developmental period to study such phenomena, as both scalar implicatures and metaphors undergo important improvements at that age (e.g. Deckert et al., 2019; Katsos et al., 2016; Winner et al., 1976). Such inter-individual variability offers opportunities for between-group comparisons, while also being of practical importance, as differences in these skills have important implications for children’s lives (see, e.g., Del Sette et al., 2021). Indeed, implied meaning is ubiquitous in language use (e.g., Gibbs, 2000) and impaired pragmatic skills have both short and long term negative consequences (e.g., Law et al., 2015).
In this context, the aim of our study is three-fold. First, we wanted to assess whether autistic children’s typical performance with scalar implicatures could be observed in a naturalistic tablet-based task that directly assesses how scalars are interpreted. Second, we wanted to confirm that autism was associated with an atypical processing of metaphors, in a new tablet metaphor task that limits and controls for language demands. Third, we aimed to test whether these tasks confirmed the contrast between scalar implicatures (as linguistic-pragmatics) and metaphor comprehension (as social-pragmatics), in a single clinical sample. Based on the literature, we hypothesized that autistic children with typical language abilities would perform similarly to their neurotypical peers on scalar implicatures, while performing poorer on metaphors. Such refined linguistic consideration is of importance, because it may constitute a window through which gaining a better understanding of how implied meaning in general is processed in autism, in line with what autistic individuals often report (Wilson, 2022). Doing so, it would also shed light on which aspects of language processing may or may not be used by clinicians as diagnostic cues, and on how to adjust our support systems.
Methods
Participants
Twenty-three autistic children (3 girls) were recruited at the hospital Le Vinatier (France) for this study, aged from 6 years and 7 months to 10 years and 8 months old. All of them had received an ASD diagnosis by a trained child psychiatrist. They met the DSM-5-TR diagnostic criteria, which was confirmed by gold-standard measures extracted from their medical files (ADOS-2 [Lord et al., 2012], SCQ [Rutter & Bailey, 2003] and/or ADI-R [Lord et al., 1994]). We report these measures for each participant in supplementary materials (Table S1). Intellectual abilities had been tested with various instruments (see Table S1 in supplementary material, missing = 2) and revealed IQs in the typical range (mean full-scale IQ = 111, 95% CI = [100,122], range = [79, 138], missing = 10; mean verbal IQ = 107, 95% CI = [101, 113], range = [78, 130], missing = 3; mean fluid reasoning = 113, 95% CI = [106, 120], range = [82, 135], missing = 3). All children were additionally screened with a tablet picture sequencing task (General Sequencing Abilities index from Petit et al. [2024]), which was shown to be strongly correlated with IQ (Rajkumar et al., 2008), and which revealed that the group performed typically overall (mean percentile position = 60, SD = 26). Comorbid attention disorder, intellectual development disorder or language disorder were considered as exclusion criteria, to avoid potential confounds. All children attended regular schools and were native French speakers.
To provide control data, we used a data set of 237 typically developing (TD) children from previous studies (Petit, 2023; Petit et al., 2024a), recruited in two French schools, aged between 6 and 11 years old, who were native French speakers and whose parents reported no neurodevelopmental disorder. This sample only included children whose performance did not fall below 2.5 SD under their age groups’ mean on a receptive grammar task and a theory of mind task (Petit et al., 2024a).
Children’s socioeconomic status (SES) was measured through parental questionnaires of parents’ education level (coded from 0 = no diploma to 7 = PhD) and family resources via the Family Affluence Scale (Currie et al., 2008), standardized in a 0 to 100 score (from the minimum to the maximum possible value), and then averaged in a SES composite. Information on participant’s ethnicity was not collected, in accordance with local laws.
All participants and their parents provided informed written consent. This study was part of a larger project on pragmatic inferences across development, which received ethical approval from the local IRB (Comité de Protection des Personnes Sud-Est I, ID RCB 2019-A01721-56).
Material
Formal Language Skills
Children were tested with two French validated language tests from the BILO-3 C (Khomsi et al., 2007), to confirm the exclusion of language disorder and better characterize the receptive linguistic abilities of autistic participants. More specifically, we assessed receptive grammar (CO subtask) through a computerized image matching task resulting in a raw score ranging from 0 to 27, and receptive vocabulary (JL subtask) through a lexical judgement task resulting in a score ranging from 0 to 43. Higher values indicate more accurate performances. Internal consistency was acceptable to good for the two tasks in this sample (receptive grammar Cronbach ⍺ = 0.68; receptive vocabulary ⍺ = 0.76).
Pragmatic Language Skills
We also used 2 new clinical tasks from the French tablet-based battery TIPi, assessing scalar inferences (Petit, 2023; Petit et al., 2024a) and novel metaphor comprehension [reference masked for review]. Both tasks are designed to reduce language demands, to limit possible confounds in the assessment. Moreover, compared to paper-and-pencil solutions, tablet-tasks tend to be preferred by children (Piatt et al., 2016), especially in autism (Lane & Radesky, 2019). They also allow for reliable group-testing (Bignardi et al., 2021), with perfect standardization and easy response times measurement (see, e.g., Petit et al., 2024b).
The scalar task is an action-based paradigm inspired from Pouscoulous et al. (2007). In this task, a table with 5 plates and fruits are displayed on the screen; children are asked to help set the scene, following oral instructions such as “I want all plates to have a banana”. Participants can afterward modify the setting by sliding fruits on the screen, and then are to validate when they are satisfied with the disposition. Variations are applied to initial fruits disposition (all plates empty, all plates filled, only a subset of the plates filled) as well as the instruction’s quantifier (none, some, all), so as to create all possible conditions, including a critical condition, in which all plates initially contain a fruit, but the instruction is that some of the plates should. A semantic interpretation of the scalar term (‘some-and-possibly-all’) would directly lead to validation. In contrast, a pragmatic interpretation (‘some-but-not-all’) would need at least one plate cleared before validation. This critical underinformative condition thus directly reveals how scalars are interpreted. It represents 7 items, out of a pool of 29 total items when including the other conditions and filler items (where different fruits are used in the same scene, to limit repetitiveness). The items were presented in a fixed pseudorandomized order. Children’s responses at validation and response times were recorded (response times are measured between the disambiguating phoneme of the instruction, i.e. some/all/none, and the validation). The task is preceded by 4 training items, using literal content only, which must be succeeded for the proper task to start. The task is about 5 min long. Internal consistency within the critical condition of the task was acceptable to good (accuracy ⍺ = 0.77, response times ⍺ = 0.80).
The novel metaphor comprehension task is a referential task based on the core principles of Noveck et al.’s (2001) and Van Herwegen et al.’s (2013). It showed good face and structural validity and good developmental sensitivity (Petit et al., 2024a). In this task, oral verbal stories are read to children. A referent is introduced in the first part of the story, and then recalled by a character in a final critical sentence. This reference can be either metaphorical, in the critical condition, or synonymic in a control condition. For example, consider the expression the fountain, which harks back to a previously-mentioned element; in the metaphorical condition, the expression would refer to a crying baby and in the synonymic condition it would refer back to water jets in a park that need repairing (see the structure of the task in Figure S1 of the supplementary materials). Metaphors were designed to be novel. A question on the nature of referent is asked to the child afterwards, to which they are expected to answer by selecting one picture out of four, including a metaphorical interpretation, a literal interpretation, a distractor and an I-don’t-know response option. Questions target what happened concretely in the story, needing the reference to be understood, rather than focusing on the speaker’s intention. Stories are 47 to 63 words long, and the vocabulary is kept simple, especially for the target and the reference nouns which have been carefully selected. For each story, a single picture is displayed on the tablet while the story is uttered, representing its general context, to favor engagement (rather than comprehension of the story per se). Responses as well as response times are recorded. The task includes 6 novel metaphors and 6 control synonymic stories. The linguistic material of each story was carefully selected and controlled to be comparable in the two conditions. To mask the task’s goal and structure to children, the 12 items are mixed with 22 other stories involving other kinds of linguistic or pragmatic phenomena that are not of interest to this paper and are considered as fillers here. Items are presented in a fixed pseudo-randomized order. The task is preceded by 3 training items, using literal content only, which must be succeeded for the proper task to start. This task is about 20 min long. Internal consistency within the critical condition of the task was acceptable to good (accuracy ⍺ = 0.81, response times ⍺ = 0.68).
Procedure
Children from the autistic group were tested individually in a quiet office of the hospital. Breaks were proposed between tasks when children showed signs of attentional fatigue or boredom. Typically developing children were tested in collective settings (in their classrooms), yet with individual tablets and headsets (see Petit et al., 2024a, for precise description). The metaphor task, which is longer, was systematically proposed before the scalar task.
Analysis
Analyses were run on R (R Core Team, 2022). Accuracy was analyzed following Thomas et al.’s (2009) developmental trajectory approach, which recommends regressing performance against age for each group and comparing the trajectories, so as to provide richer information than a simple mean comparison, even for small-size samples. We first applied it to the language tests as a control analysis; we used multiple linear regressions with total accuracy as the dependent variable and age, group and their interaction as predictors. Then we turned to our experimental tasks, for which we used generalized mixed effect models with age and group, and their potential interaction, as fixed effect, and by-participants and by-items random intercepts. Mixed effect models have the advantage of being more robust to variations in groups’ sample sizes, allowing to compare unbalanced groups, unlike more traditional approaches like ANOVAs (West et al., 2022). For the scalar task, two separate models were run, one explaining semantic errors on all items (i.e. responses that violate the quantifier’s semantic meaning), as a control analysis, and one explaining pragmatic responses in the critical condition. For the metaphor task, condition (metaphor vs. synonymic) was added as a fixed effect, together with its potential interactions with the other effects. Since gender did not affect performance in the validation studies of our experimental tasks [Petit, 2023; Petit et al., 2024a], it was not included in the analyses.
Response times were then analyzed with mixed effect models, after log-transforming the data to correct for typical skewness, excluding incorrect responses and removing outliers, defined as responses which deviated of more than 2.5 DS from groups and conditions’ means. Here, we did not opt for a developmental trajectory approach, as response times typically yield much more noise. We thus employed the minimally needed predictors to detect our effects of interest. For metaphors, we specified group and reference-condition as fixed effects, with their interaction, and by-item and by-participant random intercepts. Response times analysis for the scalar task imposed using more controls, as the conditions of interest could not be directly compared: we were interested in comparing pragmatic responses (where fruits had to be moved) to semantic responses in critical items (where no fruits had to me moved), and to responses in control situations (responses to control quantifiers, where both situations happened). To do so, we followed (Petit, 2023)’s rational and implemented a model that controls for such parasite variables in addition to the response-type (pragmatic/semantic/control) of interest. The method is detailed in supplementary materials (Box S1).
Models were fit with the lme4 package (Bates et al., 2015) and contrasts estimated with emmeans (Lenth, 2022), adjusting for multiple comparison with Tukey method when necessary. Dummy coding was used for group (TD as baseline) and sum contrasts for the other contrasts.
Community Involvement
This study involved clinical practitioners in direct contact with autistic children and their families at each stage of this research.
Background measures and descriptive performances of both groups on the different tasks are displayed in Table 1. We first analyzed accuracy on the two control language tasks (see Figure S1 in supplementary materials). For the receptive grammar task, the two developmental trajectories were remarkably superimposed, revealing a main effect of age (ß = 1.36, SE = 0.16, p < 0.001), but no main effect of group (ß = 1.8, SE = 4.8, p = 0.70) and no interaction (ß = −0.20, SE = 0.54, p = 0.71). For the receptive vocabulary task, there was a main effect of age (ß = 1.3, SE = 0.17, p < 0.001) and of group (ß = 10.8, SE = 5.35, p < 0.05), while the interaction term did not reach significance level (ß = −1.15, SE = 0.6, p < 0.10): ASD children tended to outperform TD children in the younger section of the sample.
Table 1
Sample characteristics and descriptive statistics for control and experimental tasks
TD
ASD
Demographic information
N
237
23
Girls
51%
13%
X2(1, N = 260) = 0.00, p <.05
Mean age (SD)
8.4 (1.4)
8.8 (1.4)
t(26.6) = −1.34, p =.19
Mean SES (SD)
58.8 (15.7)
53.2 (17.0)
t(25.8) = 1.56, p =.13
Language control tasks
Mean receptive vocabulary (SD)
34.7 (4.2)
35.8 (3.2)
Mean receptive grammar (SD)
15.9 (3.8)
16.5 (3.9)
Scalar task
Mean error rate (SD)
2% (0.6%)
2% (0.5%)
Mean pragmatic response rate (SD)
79% (34%)
65% (43%)
Metaphor task
Mean accuracy, literal condition (SD)
95% (12%)
94% (11%)
Mean accuracy, metaphor condition (SD)
44% (35%)
49% (37%)
Scalar Implicatures
To analyze accuracy in the scalar task, we first looked at semantic errors, which were very rare (2% of the responses). The model showed that they were slightly less likely to occur with higher age (OR = 0.67, 95% CI = [0.48, 0.93], p < 0.05) but there was no effect of group (OR = 0.58, 95% CI = [0.15–2.22], p > 0.05) and no interaction (OR = 2.39, 95% CI = [0.60–9.48], p > 0.05). This confirmed that the task was accessible to all children, independently of group, so we turned to our critical condition, where pragmatic responses appeared more variable (77% of responses overall). The likelihood of pragmatic responses was shown by the model to be influenced by age (OR = 4.87, 95% CI = [2.4, 9.9]) and by group (OR = 0.03, 95% CI = [0.00, 0.38], p < 0.01), with a group by age interaction (OR = 24.7, 95% CI = [2.0, 297.7], p < 0.05): pragmatic responses were clearly less likely to occur in the ASD group, and this difference appeared more pronounced in the younger portion of the sample (see Fig. 1). We computed each participant’s pragmatic interpretation rate in the critical condition, revealing bimodal distributions in both groups, with a subgroup of consistently pragmatic responders, and a subgroup of consistently semantic responders (see Figure S2 in supplementary materials).
Fig. 1
Predicted probability of pragmatic responses in the scalar task based on age and group. Shaded areas represent 95% CI
×
As an exploratory analysis, we assessed children’s tendency to operate changes on the scenes, independently from the instructions. To do so, we isolated control items for which changes were possible but not necessary, i.e., items with the quantifier ‘some’ and a partial initial context (e.g., when the instruction was “I want some plates to have a banana” when 2 plates initially did: even if this was not necessary, children still could add 1 or 2 bananas without changing the answer’s categorization). We excluded errors. Then we fitted a generalized model explaining the probability to make a change, with age, group and their interaction as fixed effects, and random intercepts for participants and items. This revealed that the probability to operate unnecessary changes was important but reliably decreased with age (ß = − 0.54, SE = 0.12, z =− 4.3, p < 0.001), and that it was more important in the autistic group (ß = 1.3, SE = 0.45, z = 2.9, p < 0.01). No interaction was detected (p = 0.79).
Turning to response times, we were interested in the effect of response-type (semantic responses to the critical condition vs pragmatic responses to that condition vs responses to control quantifiers, 3637 data points after excluding 2.7% of outliers), while controlling for the variables described in the methods. This revealed a group by response-type interaction (likelihood ratio test χ2(2) = 41, p < 0.0001, complete model’s output in Table 2). Post-hoc contrasts revealed the same pattern in both groups (Fig. 2): pragmatic responses were slower than both semantic responses and control responses (all ps < 0.05), but this difference was amplified in the ASD group, which was faster on both semantic responses (ß = 0.56, SE = 0.14, p < 0.001) and control responses (ß = 0.48, SE = 0.13, p < 0.001), but was as slow as TD participants when responding pragmatically (ß = 0.05, SE = 0.14, p = 0.70).
Table 2
Outcome of the model explaining (log transformed) response times in the scalar task
Predictors
Estimates
95% CI
p
(Intercept)
8.02
7.95–8.09
< 0.001
DirectValidation
0.06
−0.10 – 0.22
0.435
Moves
0.22
0.20–0.25
< 0.001
Response -type1
− 0.17
− 0.37–0.03
0.102
Response -type2
− 0.16
− 0.40–0.07
0.173
Group [ASD]
− 0.37
− 0.57 – − 0.16
< 0.001
Response -type1 * group [ASD]
− 0.36
− 1.03–0.31
0.296
Response-type2 * group [ASD]
− 0.58
− 1.00 – − 0.16
0.007
Bold lines represent significant effects
Fig. 2
Response-type and group partial effects and interaction on response times in the scalar task Error bars represent 95% CI
×
Metaphors
In the metaphor task, accuracy was very high in the synonymic control condition (95%) and moderate in the metaphor condition (45%; chance level = 25%), where most of the errors resulted from the selection of the literal response option (94% of the errors). The model (Table 3) revealed a clear age by condition interaction but, crucially, no group effect or interaction involving the group. In other terms, the developmental trajectories of both groups were remarkably similar on both conditions (see Fig. 3), with literal stories at ceiling as early as 6 years old, but an important progression with increasing age on metaphors. Exploratory by-item analysis revealed that this pattern applied to each item taken individually.
Table 3
Output from the generalized model explaining accuracy in the metaphor task
Predictors
Odds ratios
95% CI
p
(Intercept)
5.01
3.48–7.20
<.001
Condition
46.7
24.12–90.46
<.001
Age
1.8
1.43–2.18
<.001
Group
0.86
0.43–1.74
.68
Condition * age
0.43
0.32–0.58
<.001
Condition * group
0.70
0.28–1.76
.44
Age * Group
1.18
0.56–2.48
.64
(Condition * age) * group
1.14
0.41–3.14
.81
Bold lines represent significant effects
Fig. 3
Predicted accuracy in the metaphor task for both groups based on age and condition. Shaded areas represent 95% CI
×
Regarding response times, the model was run on 2125 data points, after excluding 2% of outliers. It revealed main effects of condition (ß = −0.16, SE = 0.06, p = 0.02) and group (ß = 0.21, SE = 0.05, p < 0.001), but no interaction (ß = 0.03, SE = 0.06, p = 0.64). Metaphors prompted longer response times. Autistic participants were slower in general (see Fig. 4). That is to say, the temporal cost associated with the metaphoric condition, compared to the control synonymic condition, was of similar amplitude in both groups.
Fig. 4
Predicted log-transformed response times in the metaphor task for both groups in both conditions. Error bars represent 95% CI
×
Discussion
In this study, we wanted to test whether contrasting different types of pragmatic inferences was relevant to spot specific differences between autistic and neurotypical children. We did so using two new tablet-based experimental tasks that controlled for language demand, and in a single clinical sample. Conversely to our hypothesis, we observed that autistic children performed poorer than neurotypical peers in scalar implicatures but similarly on metaphors. Response times patterns were consistent with this view, since autistic children showed an increased cost of pragmatic responses compared to control items in the scalar task, but a similar temporal cost associated with metaphors. We discuss below how these results relate to the literatures on scalar implicatures and on metaphors, then examine why previous accounts could not predict the contrast we observed. We finally consider the need for an alternative explanation.
Scalar Implicatures
Our results on scalar implicatures contrast with most studies so far, which observed that autistic participants processed scalar implicatures in a neurotypical fashion (Andrés-Roqueta & Katsos, 2020; Chevallier et al., 2010; Hochstein et al., 2017; Pijnacker et al., 2009; Schaeken et al., 2018; Su & Su, 2015). Our exploratory analysis suggested that the group difference we observed could not be attributed to peculiarities in the management of the task itself, rather than with scalar implicatures, since, 1) no group difference could be observed on control items, and 2) autistic children were more likely to make unnecessary changes, which should artificially inflate the frequency of pragmatic responses, while we observed the very opposite. Moreover, since the autistic children we tested presented typical formal language and intelligence skills, we can rule out the possibility that our results could be attributed to general or verbal intellectual functioning differences in our samples.
Previous studies have indicated that autistic individuals are typically able to judge the appropriateness of a pragmatic utterance when asked to in judgement tasks (although atypically when given the possibility of an intermediate judgement [Schaeken et al., 2018]). However, our results suggest that, at the same time, they may be less likely to spontaneously use informativity to derive speakers’ intentions (Petit, 2023), resulting in fewer pragmatic interpretations, and at a greater cost in terms of response times. In other words, autistic individuals may be able to judge as incorrect a sentence like “I want some plates to have a pear” when explicitly asked to evaluate this sentence in a situation where all plates do; but they also may be less likely to spontaneously interpret it as meaning “not all plates have a pear” in more naturalistic situations. We hypothesized that autistic children would perform similarly to their neurotypical peers on scalar implicatures, but our result support a more nuanced view: between-groups differences seem to depend on how scalar implicatures are tested. Consistently with that view, studies which did not employ binary judgement but multiple-choice response format tended to observe a diminished rate of pragmatic responses in autistic children (Mazzaggio et al., 2021; Pastor-Cerezuela et al., 2018).
Metaphor Comprehension
Our results also contrast with previous evidence that reported an impaired comprehension of metaphors in autistic compared to neurotypical individuals (Kalandadze et al., 2018, 2019; Morsanyi et al., 2020), since we observed very similar developmental trends in our groups. Previous evidence, however, showed considerable variation in effect sizes which has been related, notably, to group matching strategies (Kalandadze et al., 2018) and experimental variables (Kalandadze et al., 2019; Morsanyi et al., 2020): effects sizes were typically lower when groups were matched on language abilities and when using multiple-choice instead of verbal explanations. Since our study meets both of those criteria, a lower effect size could have been expected, even though we did not expect an absence of group effect. Anyway, our results suggest that autistic individuals can interpret metaphors typically, in such conditions.
Response times might shed light on the mechanisms underlying children’s responses. Few previous studies have used this kind of measure with autistic participants, and did report group differences, even in cases where accuracy was comparable between groups (Chahboun et al., 2016; Morsanyi et al., 2022). However, while those studies reported slower responses to metaphorical material in autistic compared to neurotypical participants, they did not include control material involving literal meaning only, raising doubts about whether autistic participants were slower with metaphors specifically, or with these tasks in general. In our study, autistic participants did respond slower to metaphors, but so did they with literal control items, with no interaction between group and condition. In other words, the temporal cost associated with metaphors comprehension, compared to literal comprehension, was similar in both groups. This suggests atypical processes involved by the task in general (say, in the literal processing of the utterances, or in selecting an appropriate response option), rather than with metaphorical expressions specifically, which might also apply to previous evidence. Accordingly, the other studies with autistic adults which included control conditions did not observe clear group by condition interactions (Gold & Faust, 2010; Gold et al., 2010).
Contrasting Scalars and Metaphors
We can now confront these two pieces of evidence to turn to our more global research question, the contrast between scalar implicatures and metaphor comprehension. Based on the linguistic- vs social-pragmatics distinction (Andrés-Roqueta & Katsos, 2017), we expected our autistic participants to succeed the linguistic-pragmatics task (scalar implicatures) but to perform poorer in the metaphors task. And yet, as discussed above, we observe a contrast in the opposite direction. Following the mental-blindness account of autistic functioning (Baron-Cohen, 1997), one might argue that the metaphor task we used did not require theory of mind abilities to be solved, which would explain 1) that it could be passed by autistic children, and 2) that this task would thus make a poor instance of social-pragmatics. We do not believe this explanation to be likely. First, we follow Marocchini (2023) and acknowledge that it is not because autistic participants succeed in a task that this task does not involve social reasoning or, more specifically, theory of mind, despite the heritage of such account of autism. Second, we showed in previous work that performance in the very metaphor task we used was indeed associated with theory of mind performance [Petit ae al., 2024a], making it an acceptable instance of social-pragmatics.
Moreover, the literature reveals that lowering linguistic and interaction demands is usually a factor facilitating access to implied meaning for autistic individuals (Kalandadze et al., 2019; Morsanyi et al., 2020). Yet, both the scalar task and the metaphor task we used are designed to limit those demands; so this aspect alone cannot either explain the contrast we observed. We thus argue that the contrast we observed calls for an additional explanation.
A Need for a Novel Account?
An a posteriori examination of the tasks we employed revealed that another factor may distinguish them: the scalar task demands the spontaneous use of information to derive the speaker’s intention (Petit, 2023), as discussed above, while the metaphor task could be described as guiding the pragmatic reasoning, through explicit questions and response propositions. This raises the possibility that autistic individuals may show typical pragmatic abilities, as measured with guiding paradigms, which provide scaffolded support; on the other hand they could struggle to recruit those skills spontaneously, resulting in poorer performances in tasks which do not guide pragmatic reasoning, and in everyday life interactions with (mostly non-autistic) speakers, where a high level of social cognitive engagement is needed.
Such distinction might contribute to explain the larger effect sizes observed in metaphor studies which employed verbal explanation tasks, which are currently attributed to language demands only. Indeed, in addition to increasing language demands, these tasks arguably guide less pragmatic reasoning than judgement tasks or multiple-choice response formats. However, very different tasks are gathered under the label “verbal explanation tasks”, and different levels of language demands appear to be conflated. Language demands are arguably very different in a proper verbal explanation task (where participants are expected to explain the meaning of a metaphor, as in Melogno et al., 2012) and in an open question task (where single words referring to concrete material may be sufficient). For example, in Kalandadze et al.’s (2019) meta-analysis, language demand was used to explain why Rundblad and Annaz’ (2010) study found the largest effect. Interestingly, this study used a task that is in some respects very similar to ours, except children had to respond orally (without any response proposition), with a language demand that did not appear much higher than in our task, as a single word response could be sufficient. However, we argue that open questions provide much less scaffolding for pragmatic reasoning than multiple choice, and this factor is likely to play an important role in explaining the large discrepancy between our results, rather than language demand (although language demand may still be an important factor to consider in general).
Accordingly, previous studies have directly shown that subtle modifications in experimental settings could greatly impact group-differences in autism studies. For example, Wilson and Bishop (2021) showed that when presented with questions on implied meanings, autistic adults were much more likely to select an “I don’t know” response when available. But when this proposition was removed, they responded similarly to non-autistic adults in most cases. This reveals that autistic participants’ management of pragmatic inferences is somewhat related to a general preference for certainty and that response propositions could guide pragmatic reasoning and force participants to reach a conclusion, while they might spontaneously remain on an uncertain outcome, if given the possibility to do so.
Conversely, compared to classical binary judgement tasks, our action-based scalar task reduces metalinguistic demands and boosts typically developing children’s performances (Petit, 2023; Pouscoulous et al., 2007). Yet, while it should also facilitate the reasoning of autistic children, it showed the opposing effect on them: their performances were relatively hindered by such modification. This indicates that metalinguistic demand can’t account for the effect we observed, contrary to a dimension opposing guided (i.e., environmentally structured) to spontaneous pragmatic reasoning.
This could also help understand the somewhat small effect sizes observed in pragmatic research, compared to the important challenges that autistic individuals and their relatives report in clinical practice. Developing experimental paradigms that are sensitive to spontaneous pragmatic reasoning could thus help in detecting the difficulties that autistic individuals can experience in real-life interactions. Indeed, real-life situations demand to keep track of others’ minds in much noisier situations, compared to lab- or clinic-controlled environment, and to apply various competences outside a planned task or event. Interestingly, such distinction between guided and spontaneous pragmatics processes echoes a proposal in the field of theory-of-mind. Indeed, the historical claim of a ToM deficit in autism (Baron-Cohen, 1997) has evolved to a subtler proposition of relatively preserved competencies, which may be less spontaneously recruited in autism (Senju, 2012).
The proposed opposition could be operationalized through a set of criteria for what makes a pragmatic reasoning more guided (or less spontaneous), such as 1) being in a testing environment (such as in a lab or hospital office with an experimenter), 2) being in a distractor-/noise-free situation, 3) being asked explicit questions, 4) being offered response propositions. Real life usually meets none of these criteria (although it is likely to involve more or less guiding situations), while our scalar task meets the first two, and our metaphor task meets all four (see Fig. 5). This hypothesis is not meant to overwrite others that have previously been discussed to explain the variability in the evidence on autistic pragmatic comprehension, such as language demand (Kalandadze et al., 2019) or theory-of-mind-demand (Andrés-Roqueta & Katsos, 2020): we remain agnostic as to whether or not the spontaneousness of a pragmatic operation may vary orthogonally to such factors.
Fig. 5
Schematic representation of a dimension opposing guided to spontaneous pragmatic reasoning, on which different situations are theoretically positioned. Following our a posteriori hypothesis, pragmatic reasoning situations (a fortiori, pragmatic experimental tasks) could be positioned on a dimension opposing guided to spontaneous pragmatic reasoning (horizontal axis), for different pragmatic phenomena (vertical axis), which may contribute to explain the magnitude of the difference between autistic and neurotypical individuals, from minimal (in highly guided reasoning) to larger (in highly spontaneous reasoning). Real world situations are positioned on the right of this axis, as they are expected to demand more spontaneous reasoning in general, but they may also include more or less guiding situations
×
This hypothesis may have promising clinical implications, but it remains an a posteriori explanation, which should now be specifically tested, for instance by comparing how autistic individuals perform in a guided vs spontaneous measure of a single pragmatic phenomena, using carefully controlled materials, that especially controls for language demands. More generally, our study presents some general limitations that should be addressed in future research. Using a control group with a large sample size allowed us to increase the statistical power of the group comparison, while we meticulously selected autistic participants (in terms of intellectual and verbal abilities, as well as co-occurring conditions) to control for a series of potential confounds, which limited recruitment. Such an imbalance in sample sizes should not be problematic with the statistical tools we employed, but the limited sample size of the autistic group could have hindered the precision of the effect described within this clinical group. Moreover, increasing sample size may also reveal subtler effects, such as non-linear developmental patterns, or slighter effects on response times, which dependent variable is typically associated with important noise. Furthermore, different from our careful selection of autistic participants, in real-life settings autism is most frequently associated with other psychiatric or neurodevelopmental conditions (Levy et al., 2010). The generalizability of our results to other subgroups of autistic individuals thus still needs to be assessed.
Conclusion
In this study, we tested a group of autistic children with unimpaired language and intellectual functioning on two different pragmatic tasks that involved metaphors, which are classically considered as difficult for autistic individuals, and scalar implicatures, which autistic participants have mostly been found to understand as well as neurotypical individuals. Yet, compared to typically developing children, our clinical sample showed a reversed contrast, with poorer scalar implicatures comprehension but typical metaphor comprehension. Accuracy and response times pattern were consistent with that view. Complementary to previously introduced accounts, we suggest that this reversed contrast could be explained by an opposition of guided to spontaneous pragmatic processes. Paralleling theory of mind hypothesis, autistic individuals might show typical pragmatic abilities, as measured with guiding paradigms, but might struggle to recruit those skills spontaneously. This would result in poorer performances in tasks which do not guide pragmatic reasoning, and difficulties in everyday life interactions, which additionally demand to keep track of others’ minds and communicative intentions in a much noisier environment.
Acknowledgements
We thank Ira Noveck and Jerome Prado for their contributions and support throughout this project. We are also to thank all the children and families who gave their time for this study, as well as Le Vinatier’s teams who helped with the recruitment of the participants. We extend our gratitude to the associate editor, Patricia Ann Prelock, and the two anonymous reviewers for their insightful comments and suggestions, which significantly enhanced the manuscript
Declarations
Competing interest
The authors have no competing interests to disclose.
Ethical approval
All participants and their parents provided informed written consent. This study was part of a project which received ethical approval from the local IRB (Comité de Protection des Personnes Sud-Est I, ID RCB 2019-A01721-56).
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
share
DELEN
Deel dit onderdeel of sectie (kopieer de link)
Optie A:
Klik op de rechtermuisknop op de link en selecteer de optie “linkadres kopiëren”
Met BSL Psychologie Totaal blijf je als professional steeds op de hoogte van de nieuwste ontwikkelingen binnen jouw vak. Met het online abonnement heb je toegang tot een groot aantal boeken, protocollen, vaktijdschriften en e-learnings op het gebied van psychologie en psychiatrie. Zo kun je op je gemak en wanneer het jou het beste uitkomt verdiepen in jouw vakgebied.
Regarding the way to refer to autistic individuals, which has important real-life consequences (Vivanti, 2020), and in absence of a clear consensus of the preferred terms (although a consensus emerges on the least preferred), we followed Botha et al.'s (2021) rationale and used identity-first formulation, considered as less offensive and less stigmatizing.