Communication is fast and full of ambiguity. Thus, comprehenders must keep up with the speed of language and at the same time determine the intended meaning of a sentence (Crocker, 2005). Generating predictions are fundamental herein. Predictions about upcoming language speed up processing and thus help comprehenders to keep up with the speed of spoken language (e.g., Corps et al., 2018, 2019; Fitz and Chang, 2019; Kochari and Flecken, 2019; Shain et al., 2020).

It has been well-established that linguistic predictions are generated when particular linguistic information is activated in language users, even before the input that carries this information becomes available (Pickering and Gambi, 2018). For example, when hearing “John wants salt and pepper on his steak” a comprehender would be highly likely to predict lexical information, i.e., the word “pepper,” after hearing “salt and” before actually hearing “pepper,” because “salt and pepper” is a pair of words frequently used together in a fixed order. Preactivation of linguistic information is also empirically demonstrated by, for example, eye-tracking studies in which participants heard a sentence like “The boy eats the cake” while looking at pictures. Adults and children already look at the correct picture of a cake instead of competing pictures of other objects before they actually hear the word “cake” (e.g., Altmann and Kamide, 1999; Nation et al., 2003; Borovsky et al., 2012; Mani and Huettig, 2012). Prediction of the word “cake” comes from the verb “eats” which requires an object that is edible. This shows that comprehenders use lexical information to predict upcoming language.

Comprehenders also predict upcoming language on the basis of syntactic information. For example, Lukyanenko and Fisher (2016) showed that 3-year-old children are already faster and more likely to shift their gaze from an incorrect picture displaying only one thing to a correct picture displaying multiple things upon hearing a verb that requires a plural noun (e.g., “Where are the cookies?”) compared to a verb that is uninformative about the number of the noun (“Do you see the cookies?”). Thus, toddlers use agreement between the number marking (singular or plural) of a verb and a noun to predict features of an upcoming noun, resulting in a faster identification of the correct picture. The studies described above show that preactivation of lexical as well as syntactical information helps to do some of the processing ahead of time so that comprehenders can process language fast, despite the speed with which sentences are produced and despite the pervasive ambiguity of language (Pickering and Gambi, 2018).

Given the ambiguity in language and the need for fast processing, predictions can sometimes also steer comprehenders in the wrong direction. This is the case if a comprehender's initial prediction turns out to be false. For example, in a sentence like “The horse raced past the barn fell” (Bever, 1970), comprehenders will predict that “the horse” is the subject of the sentence, because there is a tendency to interpret the first noun phrase in a sentence as the subject, and subjects as agents. Subsequently, they will predict that “raced” is the sentence's main verb (referring to what the horse did) and that “past the barn” is the direction in which the horse raced. Thus, after hearing “the horse raced past the barn,” comprehenders will predict that the sentence is complete. However, upon hearing the final word of the sentence (the verb “fell”) comprehenders will discover that the sentence was not yet complete, and that “raced past the barn” was a specification of “the horse” (i.e., the horse that was raced past the barn). That is, comprehenders will discover that they were led up the garden path. These so-called garden path sentences thus require comprehenders to update their initial incorrect prediction by replacing it with a new prediction, to arrive at the intended interpretation of the sentence.

Autistic individuals often have difficulties with language and communication. It is conceivable that these difficulties are a consequence of problems with predictions, and that the difficulties with ToM and EF often observed in autistic persons are linked to these prediction problems. ToM enables comprehenders to deduce why a person acted in a certain way or to anticipate how a person is likely to act. In this sense, ToM tasks are inherently prediction tasks, as comprehenders need to predict the actions of a person following a certain observation. Therefore, difficulties with ToM may be caused by a reduced ability to predict another person's behavior. EF is related to predictions as well, especially in situations where an initial prediction turns out to be incorrect. In such situations, comprehenders are required to inhibit their initial but false prediction and switch to an alternative interpretation while holding all information active in their working memory. Therefore, updating an incorrect predictions requires EF, in particular, cognitive inhibition (MacLeod, 2008), cognitive flexibility (Miyake and Friedman, 2012) and working memory (Miyake and Shah, 1999). This thus shows how ToM and EF are related to generating and updating predictions. However, as said earlier, the language difficulties seen in autism are broader than the hypotheses of ToM and EF can explain individually or together. In contrast, prediction impairments during language processing may explain the pragmatic and non-pragmatic language difficulties seen in autism and subsume the hypotheses of ToM and EF. Hence, ToM and EF are related to predictive abilities, but the process of generating and updating predictions is broader and has more explanatory power.

Recent studies have illustrated abnormalities in predictive abilities in autistic individuals outside the domains of language (see Cannon et al., 2021, for a recent review of empirical evidence). Neuroimaging and eye-tracking studies using tasks in which participants are presented with predictable repetitive stimuli that are infrequently interrupted by an unpredictable deviant stimulus found that autistic individuals showed an altered response compared to neurotypical controls (Jeste et al., 2015; Balsters et al., 2017; Lawson et al., 2017; Goris et al., 2018). The results of these studies suggest that autistic individuals may be less surprised when their predictions are being violated, as indicated by reduced brain responses (e.g., Lawson et al., 2017). This finding has been taken to mean that their predictions are less strong compared to those of neurotypical individuals (Pellicano and Burr, 2012), potentially because they struggle to generate precise internal prior beliefs about the world (both social and non-social; Pellicano and Burr, 2012; Friston et al., 2013; van de Cruys et al., 2014; Lawson et al., 2017), especially in temporally demanding environments where the environment changes fast in time (e.g., Pellicano and Burr, 2012; Hohwy et al., 2016; Vogel et al., 2019). Being less surprised after encountering a violation to a prediction could, in turn, mean that it is also harder to update a prediction, because the main function of surprise is to interrupt an ongoing action and reorient attention to the new, possibly significant stimulus (Kalat, 2015). This fits well with the finding in the language domain that predictions allow for faster processing. Indeed, the abnormalities in predictive abilities within autistic individuals found outside the language domain triggered our question if these abnormalities also occur within the language domain.