Autism spectrum disorder (ASD) is a complex neurodevelopmental condition characterized by atypical brain function that significantly impacts social communication, interaction, and induces restricted or repetitive behaviors and interests. Historically, theoretical frameworks focused narrowly on isolated symptom clusters, leading to incomplete explanations and difficulties establishing correlations between these symptoms and underlying neural mechanisms. As research has advanced, more recent hypotheses suggest that abnormalities in predictive processing may play a crucial role in the core symptoms of ASD. In this study, we conduct a comprehensive review of three prominent theories of prediction processing: the predictive impairment in autism hypothesis (PIA), the Bayesian perceptual theory, and the predictive coding theory. We synthesize relevant empirical evidence from diverse methodological approaches and domains. Each theory offers unique insights into the nature of predictive deficits in individuals with ASD.

The PIA hypothesis posits that individuals with ASD exhibit domain-general predictive processing deficits due to impaired estimation of the dynamics of temporally unfolding Markov systems, which reflects failures in learning and utilizing conditional probabilities. Research testing this framework often employs probabilistic learning paradigms to determine the presence of generalized predictive impairments. While much of this research has focused on the ability to form predictions, mixed findings raise questions about the nature of predictive abilities within this population. Bayesian perceptual theory posits that individuals with ASD face difficulties in the formation and application of priors, resulting in perceptual experiences that rely more heavily on immediate sensory input and are less influenced by prior knowledge. Empirical evaluation of this account frequently involves tasks demonstrably modulated by prior knowledge, such as susceptibility to visual illusions or perceptual closure tasks. Findings consistently indicate that individuals with ASD are less likely to use priori knowledge to shape perception. However, it is crucial to clearly define and distinguish between different types of prior knowledge in order to explore core deficits in individuals with ASD in the framework of Bayesian perceptual theory. Predictive coding theory emphasizes abnormalities in processing prediction errors when sensory input violates prior expectations, impairing context-based discrimination between relevant and irrelevant errors. Consequently, research aligned with predictive coding theory has predominantly focused on examining prediction error processing dynamics rather than prediction formation itself. Specifically, studies investigate whether individuals with ASD exhibit atypical weighting of prediction errors when sensory input violates expectations, particularly in relation to situational context. Relevant findings suggest that atypical processing of prediction errors in individuals with ASD is characterized by overweighting of prediction errors, and is potentially linked to low contextual sensitivity.

Synthesizing these perspectives, we propose that predictive differences in ASD manifest in a fundamentally context-dependent manner, rather than constituting a pervasive, domain-general impairment. Individuals with ASD may demonstrate competence in generating predictions within deterministic, rule-based environments, yet exhibit significant difficulties in flexibly adjusting predictions within uncertain, ambiguous, or dynamically changing contexts. This context-dependent deficit profile may arise through at least two potentially dissociable pathways: (1) Neuromodulation Mechanisms: Directly impacting prediction precision estimation, potentially influencing how individuals with ASD process and respond to complex social information, which is inherently dynamic and probabilistic. (2) Information Extraction Challenges: Difficulties in efficiently extracting and utilizing relevant information from past experiences, which is crucial for estimating prediction precision and adjusting priors in novel or uncertain situations.

Future research could address the limitations of prediction processing theories in accounting for the pervasive core deficit in individuals with ASD via simulation of computational models. Furthermore, appropriate paradigms could be explored in the fields of sensorimotor learning and sensorimotor adaptation to test the plausibility of different prediction processing theories. Finally, the three prediction processing theories should be systematically validated by considering contextual factors and examining whether the anticipations formed by individuals with ASD can be generalized and adapted to different situational demands.