Recent results obtained with a neural-network model of the language cortex suggest that the memory circuits developing for words are both distributed and functionally discrete. This model makes testable predictions about brain responses to words and pseudowords under variable availability of attentional resources. In particular, due to their strong internal connections, the action-perception circuits for words that the network spontaneously developed exhibit functionally discrete activation dynamics, which are only marginally affected by attentional variations. At the same time, network responses to unfamiliar items – pseudowords – that have not been previously learned (and, therefore, lack corresponding memory representations) exhibit (and predict) strong attention dependence, explained by the different amounts of attentional resources available and, therefore, different degrees of competition between multiple memory circuits partially activated by items lacking lexical traces. We tested these predictions in a novel magnetoencephalography experiment and presented subjects with familiar words and matched unfamiliar pseudowords during attention demanding tasks and under distraction. The magnetic mismatch negativity (MMN) response to words showed relative immunity to attention variations, whereas the MMN to pseudowords exhibited profound variability: when subjects attended the stimuli, the brain response to pseudowords was larger than that to words (as typically observed in the N400); when attention was withdrawn, the opposite pattern emerged, with the response to pseudowords reduced below the response to words. Main cortical sources of these activations were localized to superior-temporal cortex. These results confirm the model’s predictions and provide evidence in support of the hypothesis that words are represented in the brain as action-perception circuits that are both discrete and distributed.