Veritable Untruths: Autistic Traits and the Processing of Deception

Even though autism spectrum disorder (ASD) is generally associated with a deficit in social and figurative aspects of communication, many previous studies investigating the abilities of participants with ASD to understand non-literal language have shown that these participants obtain similar comprehension scores to their neurotypical peers (e.g., Happé, 1994, 1995; Hala et al., 2007; Jolliffe & Baron-Cohen, 2000; Loukusa & Molianen, 2009). This raises the question of whether individuals with ASD and those without achieve similar outcomes by employing different processes. Here, we develop the theme of processing differences, with a focus on how the manner of a spoken message is delivered, in particular, disfluent fillers such as “um” and “uh”, influences individuals’ comprehension.

Fillers are often regarded as socially oriented, influencing discourse dynamics by controlling the flow of conversation (e.g., Clark & Fox Tree, 2002), signalling hedges in conversation (Smith & Clark, 1993), or being taken as cues to deception (Loy et al., 2016; Vrij & Semin, 1996; Zuckerman et al., 1981). Previous studies have shown that high-functioning individuals with ASD produce fewer fillers than do the typical population (Lake et al., 2011). Children with autism produce fewer fillers compared to their typically developing peers (Gorman et al., 2016; Heeman et al., 2010; Irvine et al., 2016). This reduced use of fillers, considered as a conversational cue, is consistent with a characterization of speakers with ASD as being less able to take their listeners’ perspectives (Colle et al., 2008; Paul et al., 2009). A decreased attention to social cues may yield difficulty in understanding the implicit social meanings associated with fillers (Irvine et al., 2016).

However, this difficulty may not manifest in comprehension outcomes. Participants who struggle with social cues may nonetheless find ways to interpret fillers in an equivalent way to their peers. By focusing on the outcome of comprehension, usually operationalised as whether the eventual interpretation of an utterance aligns to a norm, it is possible that differences in the processes which underlie these outcomes may be overlooked (McKenna et al., 2015). McKenna et al. (2015) demonstrated a difference in the processing of a type of figurative language, metonymy, in which an attribute of a concept instantiates that concept (such as the use of Dickens when referring to the books written by Charles Dickens rather than to the writer himself). In their study, participants without diagnoses completed an autism spectrum quotient questionnaire (AQ: Baron-Cohen et al., 2001) and read sentences containing target nouns (e.g., Vietnam, Finland), in contexts supporting literal or figurative, metonymic, interpretations (e.g., During my trip, I hitchhiked around Vietnam/Finland; A lot of Americans protested during Vietnam/Finland). Their results showed an “unfamiliar metonym disadvantage” where participants were slower to read a novel metonym (e.g., Finland) in its figurative sense, compared to a metonym which has its figurative meaning established in daily use (e.g., Vietnam referring to the conflict during the 1950s–1970s). Of relevance to the present study, they found that participants with higher scores on the autism quotient measure were relatively slowed down by these novel metonymies, while those with lower numbers of self-reported autistic traits were not. Note that all participants scored 100% on comprehension scores: differences related to AQ could not be found in any measure of interpretation, but were instead only evident when the time course of the processing of the stimuli was investigated.

Although understanding figurative meanings in written language, such as that investigated by McKenna et al. (2015), can be difficult, understanding non-literal meaning in speech adds additional complexity, because artefacts such as fillers can modify literal meaning. Loy et al. (2017) investigated whether interpretations of a sentence (truthful or deceptive) made by listeners in the general population could be influenced by disfluency. Their experiments used a lie-detection game setting, where participants were presented with two images (named image/distractor image) in every trial and were asked to click on the image that they thought concealed some treasure. For each trial, participants heard a pre-recorded fluent or disfluent utterance describing the location of the treasure. Participants’ image clicks revealed a bias to interpret disfluent speech as a cue to speaker deception about the location of the treasure. Importantly, eye movements were also measured, and disfluent speech resulted in more frequent early eye movements towards the distractor item, showing the influence of an early process associating disfluency with deception.

Loy et al. (2017) study provides two essential points of reference for the current study. First, the method allows us to directly observe the time-course of disfluency comprehension, providing a way to investigate how the manner in which something is said might affect its ongoing interpretation. Second, along with other situations which depend on understanding the interlocutor’s intentionality, the recognition of deception has been shown to be difficult for individuals with ASD: for example, it is harder for children with autism to lie to others and to detect when they are being lied to (Ranick et al., 2013). By using a lie detection paradigm in which disfluency can serve as a potential pragmatic cue, we can examine in detail any differences in the processes of comprehension that might arise between individuals. Specifically, we can measure both participants’ eventual interpretations—truth or deception?—and the real-time processes that precede that outcome.

Our experiment replicates experiment 2 from Loy et al. (2017), with a set of participants from the general population who vary in their autism spectrum quotients (AQ). AQ is not a diagnostic test of ASD, but it is used to assess broader autism phenotype (BAP) metrics for understanding subclinical ASD traits within individuals in the general population who are presumed to have typical developmental profiles. The test comprises 50 questions focusing on five BAPs: social skills, attention switching, attention to detail, imagination, and communication (Baron-Cohen et al., 2001). AQ scores from low to high quantify where a given participant is situated on a putative continuum from neurotypical to autism diagnosis (Baron-Cohen et al., 2001). While our findings do not allow us to draw direct conclusions about comprehension processes in ASD, the study emphasizes the utility of real-time measures in comparing processing differences among populations that vary on an AQ spectrum.

Our experiment measures the influence of manner of delivery on pragmatic comprehension in two ways. First, we measure the outcome of the comprehension process by recording which object people click on. This gives us a direct indication of whether a participant has interpreted a given utterance as truthful, in which case they are likely to click on the named (target) object, or deceptive, in which case they are likely to click on the other (distractor) image. If the results of Loy et al. (2017) are replicable, we expect to see an effect of the presence of disfluency, such that participants are less likely to click on the target following a disfluent utterance. We further assess the process of comprehension. By measuring participants’ eye movements time-locked to the speaker’s mention of the target object, we can test the time-course of comprehenders’ responses to the speaker’s manner of delivery. If the results replicate those of Loy et al., we should see a decrease of bias towards the named object in the disfluent condition. Of critical interest in the present study is whether the outcome or process measures interact with AQ. An interaction at the outcome level (item clicked) would suggest that the levels of broader autism phenotypes measured by the AQ affect the eventual understanding of what is said. Regardless of whether such an interaction is found or not, an interaction at the process level (eye movements) would suggest differences in the processes underlying non-literal interpretations, opening the door to research with formally-diagnosed groups of participants.

Data and analysis scripts are available at https://​osf.​io/​jwhfr/​. Prior to analysis, practice and filler trials were removed, as well as the trials in which participants failed to click on either object within 5000 ms of audio onset (0.5% of the critical data).

The results reported here confirm that listeners’ pragmatic judgements are influenced by manner of speech, and that disfluent instructions are associated with deception. We took the greater frequency of clicks on the distractor object after hearing an instruction such as “The treasure is behind, thee, uh, hat” as evidence that participants were less inclined to believe the location of the treasure named by the speaker when the utterance was disfluent. Moreover, eye movements made during the early time-course of comprehension showed that these judgements were made quickly. These findings replicate the results of Loy et al. (2017).

Note that in these studies, participants were explicitly told that they might be deceived, and the utterances to which they responded were recorded. Whether there are general conclusions to be drawn about deception is unclear; here we are concerned with the processing of language that is ‘known’ not to have a literal interpretation. We can however draw some ecological comfort from a study by Loy et al. (2018), in which participants’ judgements were affected in a similar way by disfluency in a two-player version of the current deception game, in which all utterances were spontaneous.

Critically, the present study shows that the general effects of disfluency described above vary by AQ. When hearing disfluent instructions, participants with higher AQ scores, and therefore higher numbers of traits typically associated with autism, are more likely to make early fixations on the target. The evidence for these effects is manifest as a series of interactions with a continuous AQ score predictor, and therefore not subject to criticisms typically applied to median splits (e.g., McClelland et al., 2015). However, the differences emerge as differences of processing but not of comprehension: our data do not suggest that participants’ final interpretations vary with AQ, but they do suggest that there is variation in the mental processes which underlie those interpretations.

A potential account of these findings is that as the number of autistic-like traits associated with higher AQ scores increases, participants have more difficulty in utilizing contextual information. This may lead to a weaker association between disfluency and deception, attenuating the early tendency to fixate the distractor following a disfluent utterance. A variant of this account is that it may be more difficult to override more basic effects of filler disfluencies, such as their tendency to focus listeners’ attention on the upcoming message (Collard et al., 2008; Fox Tree, 2002). According to the latter view, participants with higher AQs may show an initial tendency to fixate the target quickly following a disfluency, before this is overridden by slower pragmatic reasoning.

It is important to stress again that these findings are only relevant to people diagnosed with autism to the extent that AQ measures can be considered a proxy for diagnosis. The critical aim of the paper is not to identify a processing difference in the clinical population, but to point to a likely candidate which merits further investigation. Moreover, because of the continuous nature of the AQ, some statistical issues which can arise from group comparisons (especially when “controlling” for other varying factors: Miller & Chapman, 2001) are avoided. We can be confident (to the usual limits of statistical inference) that the effect is real: in the subclinical population, comprehension processes may result in the same outcomes, even though the processes by which those outcomes are achieved may vary.

In line with the study by McKenna et al. (2015), the implication of the current study is that by merely focusing on the outcomes of cognitive processing, it is possible to miss the real differences in detailed processes that vary across the population. Previous studies have used tasks in which participants are asked to provide context-appropriate explanations for non-literal speech, such as Happé’s Strange Stories Test (Happé, 1994), in exploring disfluency comprehension in ASD children. However, many of these studies represent cases in which ASD- and non-ASD individuals achieved similar results, despite the ASD participants being expected to show less developed ability (Hala et al., 2007; Happé, 1995; Jolliffe & Baron-Cohen, 2000). Such results may not be conclusive without exploring the detailed cognitive processes underlying any parity in behavioural outcomes.

The present study suggests that there are differences in the processing of disfluency among individuals varying in numbers of self-identified autistic traits. This may, by extension, have implications for the comprehension processes of people with ASD diagnoses.