It 1) picks a random topic (of all the cs., econ., math.* etc. topics) 2) finds the maximum amount of papers in that topic, and 3) queries for a random paper in that topic.
Note that this skews the distribution heavily in favor of topics that are less common, but it should get the job done. Suggestions for improvements are welcome.