Challenging the predictive mind

Abstract

In everyday life, language comprehension rarely unfolds under ideal circumstances. Individuals may encounter reduced intelligibility of linguistic input, may be required to divide their attention between comprehension and a concurrent task, or may experience diminished sensory or executive capacities as a consequence of healthy ageing or illness. Additionally, the ways in which individuals engage with language across the lifespan are highly idiosyncratic, shaped not only by the quantity and quality of linguistic experience but possibly also by inter-individual differences in cognitive-perceptual style. Against this backdrop, it is crucial to understand the mechanisms that enable successful language comprehension, as well as the compensatory strategies that allow comprehension to remain robust when situational demands or individual predispositions are less than optimal. Language comprehension is supported by predictive processing, in which the brain continuously anticipates upcoming linguistic input based on prior context and knowledge of the speaker or situation. However, despite its central role in language comprehension, the cognitive cost of language prediction remains under debate. In particular, it is unclear whether predictive processing is an automatic consequence of exposure to linguistic input or whether it draws on domain-general executive resources, and how this potential relationship may be shaped by inter-individual differences in age as well as in the propensity to weight prior knowledge versus incoming sensory evidence.

The aim of the present thesis was to address these questions within two behavioural pilot studies as well as two large-scale behavioural studies (Studies 1 and 2) and one EEG study (Study 3), employing a novel dual-task paradigm combining self-paced reading of naturalistic texts with an orthogonal, non-linguistic n-back task. This design allows cognitive load to be manipulated independently of linguistic processing, enabling a dissociation of prediction-related effects from general task difficulty. Across studies, word predictability was quantified employing word surprisal derived from a large language model (GPT-2). Predictive processing was evaluated through responses to variations in word predictability, measured both in reading times (across studies) and in neural tracking (in Study 3).

In Study 1, results revealed a robust facilitation of reading by word predictability under low cognitive load. Crucially, increasing cognitive load systematically attenuated these effects, indicating that predictive language processing depends on the availability of executive resources. However, this dependency was modulated by age: older adults showed stronger predictability effects under low load but were also more susceptible to their reduction under increasing load, consistent with earlier exhaustion of cognitive capacity. In contrast, the youngest participants exhibited a qualitatively different pattern, with weak or absent predictability effects under minimal load and an increase under higher load, raising the possibility of differences in how language prediction manifests behaviourally in young and old adulthood.

In Study 2, we first proposed an improved, psychometrically sound measure of cognitive- perceptual style based on responses in the Autism-Spectrum Quotient (AQ) and the Schizotypal Personality Questionnaire (SPQ), capturing the relative weighting of prior expectations versus sensory evidence. This measure revealed systematic age-related shifts toward a more sensory-driven profile, and showed that cognitive-perceptual style modulates sensitivity to word predictability, such that individuals with a more sensory-driven style exhibited greater susceptibility to variations in word predictability, particularly in older age.

In Study 3, reading times reproduced the graded sensitivity to word surprisal observed in Study 1: older adults showed pronounced predictability effects overall, whereas younger adults displayed little sensitivity under low cognitive load but marked increases when executive resources were constrained. At the neural level, however, a qualitatively different pattern emerged. Results revealed age-related divergences in the cortical tracking of word predictability, with younger adults exhibiting stronger neural sensitivity to variations in word predictability than older adults, within a spatiotemporal region aligned with the canonical N400 component. Moreover, when cognitive demands were high, younger adults showed attenuated updating of higher-level semantic predictive models – reflected in decreased neural tracking of word predictability within a spatiotemporal region corresponding to the Post-N400 ERP component – whereas older adults relied on sustaining and updating semantic predictions more robustly under increased task demands. Collectively, these results point to an age-dependent dissociation between overt performance and underlying neural indices of predictive processing, and indicate that executive resource availability shapes the updating of semantic predictions in qualitatively different ways across the adult lifespan.

Taken together, this thesis provides converging behavioural and neural evidence that predictive language processing is not cost-free but depends on the availability of domain-general executive resources, with a central role of executive resources especially in the updating of higher-level semantic predictive models. Moreover, the findings demonstrate that when executive resources are inherently constrained, the neural mechanisms supporting predictive language processing undergo functional reorganisation. The present work thus portrays predictive processing as a sophisticated, context-sensitive mechanism – malleable to both task demands and individual differences – that dynamically guides real-time language comprehension across the adult lifespan.