Stimulus dependencies—rather than next-word prediction—can explain pre-onset brain encoding during natural listening

The human brain is thought to constantly predict future words during language processing. Recently, a new approach to investigating linguistic predictions emerged which aims to capture predictive pre-activation directly by using neural network representations of words to predict brain activity prior to word onset. However, it is unclear what exactly is driving the predictability of pre-stimulus brain activity. Here we show, across two datasets, that both proposed hallmarks of neural pre-activation—i.e. (i) pre-onset brain response predictability and (ii) its modulation by word expectedness—is not only observed in brain responses, but also in representations of the stimulus material itself. We show that various structural and incidental dependencies existing in natural language can explain previously reported hallmarks of pre-diction without assuming any pre-activation in the neural data. This suggests that pre-onset prediction of brain activity might only reflect dependencies within the stimulus material rather than predictive computations, and questions the extent to which this new prediction-based method can be used to study prediction in the brain.