Neural Machine Translation (NMT) currently exhibits biases such as producing\ntranslations that are too short and overgenerating frequent words, and shows\npoor robustness to copy noise in training data or domain shift. Recent work has\ntied these shortcomings to beam search -- the de facto standard inference\nalgorithm in NMT -- and Eikema & Aziz (2020) propose to use Minimum Bayes Risk\n(MBR) decoding on unbiased samples instead.\n In this paper, we empirically investigate the properties of MBR decoding on a\nnumber of previously reported biases and failure cases of beam search. We find\nthat MBR still exhibits a length and token frequency bias, owing to the MT\nmetrics used as utility functions, but that MBR also increases robustness\nagainst copy noise in the training data and domain shift.\n