This article presents a comparative numerical study of the collocated and distributed massive MIMO deployments at 3.5 GHz in an industrial indoor environment from the point of view of the downlink human Electromagnetic field (EMF) exposure. A collection of environmental models incorporating stochastic geometry elements is generated, in which the EMF propagation is calculated using the Ray-Tracing (RT) method. To evaluate the human exposure, the Finite-Difference Time-Domain (FDTD) method is used, including a realistic human phantom and the user equipment model, into which the excitation is introduced based on the RT results. Single-user Maximum Ratio Transmission and multi-user Zero-Forcing scenarios are studied. Small-scale EMF distributions in proximity of the phantom's head are assessed in FDTD and analysed for different user locations in the environment and the user equipment placement with respect to the head. The massive MIMO hot-spot is characterized in terms of its size, instantaneous and time-averaged EMF enhancement, position with respect to the head and the user equipment. The human exposure is assessed using the peakspatial Specific Absorption Rate averaged over 10 g, referenced to the hot-spot EMF and compared to international guidelines. It is shown that the distributed deployment results in a more accurate and consistent EMF hot-spot around the user equipment with a higher average E-field gain, compared to the collocated deployment. In addition, the distributed configuration produced more compact hot-spots relative to the collocated one, leading to a more than 10-fold average exposure reduction in a multi-user scenario.