> humans are notoriously bad at generating random numbers

Shannon was able to model humans (choosing "heads" or "tails") effectively with only a 3-bit history, which he encoded in relays, having done so around the middle of last century.

Note that recombination is a population's way of balancing exploitation and exploration.

Did Shannon use it for something or just as an experiment?

I don't know what he did with it; try starting with https://ieeexplore.ieee.org/document/5311579 and working forward for cites?