Response conflicts hamper goal-directed behavior and may be evoked by both consciously and subliminally (unconsciously) processed information. Yet, not much is known about the mechanisms and brain regions driving the size of subliminally induced conflicts. We hence combined a response conflict paradigm featuring subliminal primes and conscious flankers with in-depth neurophysiological (EEG) analyses, including source localization in a sample of N = 243 healthy subjects. Intra-individual differences in the size of subliminal conflicts were reflected both during early attentional stimulus processing (prime-associated N1 and target-associated P1 and N1 amplitudes) and conflict monitoring (N2 amplitudes). On the neuroanatomical level, this was reflected by activity modulations in the TPJ (BA39, BA40) and V2 (BA18), which are known to be involved in attentional stimulus processing and task set maintenance. In addition to a “standard” analysis of event-related potentials, we also conducted a purely data-driven machine learning approach using support vector machines (SVM) in order to identify neurophysiological features which do not only reflect the size of subliminal conflict, but actually allow to classify/predict it. This showed that only extremely early information processing (about 65 ms after the onset of the prime) was predictive of subliminal conflict size. Importantly, this predictive feature occurred before target information could even be processed and was reflected by activity in the left middle frontal gyrus (BA6) and insula (BA13). We conclude that differences in task set maintenance and potentially also in subliminal attentional processing of task-relevant features, but not conflict monitoring, determine the size of subliminally induced response conflicts.