He could save power by just using dice and decision tables. There's a long history of civilian and military wargaming that's random but designed vs. real-world situations. Statistical token generation, however, doesn't strike me as a good bet. Hell, during one of the baseball games, Google AI led off their search results with "Seattle is not north of Toronto". It followed by having little real knowledge of historical baseball results. I can see it having problem due to training set with the latter, but damn, where countries and cities are located should be well-defined.