The assumption of linear time-invariance (LTI) in the human primary somatosensory cortex (SI) is assessed for fMRI signals generated by variable-duration vibrotactile stimuli. Predictions based on time-shifted summation (TSS) of responses to 2 s stimuli overestimate observed BOLD signal amplitudes in response to longer-duration stimuli, in agreement with previous findings in other primary sensory cortices. To interpret these results, we undertook an alternative approach for LTI assessment by characterizing BOLD signals using two biophysical models. The first model assumes that the input stimulus envelope is proportional to neural activity. The second assumes that neural activity exhibits both transient and steady-state components, consistent with extensive electrophysiological data, and fits the experimental data better. Although nonlinearity remains evident for short stimulus durations, the latter model shows that the TSS procedure to assess LTI overestimates the BOLD signal because the temporal characteristics of neural activity have not been considered adequately. Further research to investigate the BOLD response to time-varying neural activity is required.