Recent research in neural machine translation has explored flexible\ngeneration orders, as an alternative to left-to-right generation. However,\ntraining non-monotonic models brings a new complication: how to search for a\ngood ordering when there is a combinatorial explosion of orderings arriving at\nthe same final result? Also, how do these automatic orderings compare with the\nactual behaviour of human translators? Current models rely on manually built\nbiases or are left to explore all possibilities on their own. In this paper, we\nanalyze the orderings produced by human post-editors and use them to train an\nautomatic post-editing system. We compare the resulting system with those\ntrained with left-to-right and random post-editing orderings. We observe that\nhumans tend to follow a nearly left-to-right order, but with interesting\ndeviations, such as preferring to start by correcting punctuation or verbs.\n