Nowadays, datasets are an essential asset used to train, validate, and test stress detection systems based on machine learning. In this paper, we used two sets of FAIR metrics for evaluating five public datasets for stress detection. Results indicate that all these datasets comply to some extent with the (F)indable, (A)ccessible, and (R)eusable principles, but none with the (I)nteroperable principle these findings contribute to raising awareness on (i) the need for the FAIRness development and improvement of stress datasets, and (ii) the importance of promoting open science in the affective computing community.