There was a pretty funny paper in The American Statistician a few years ago: You can load a die, but you cannot bias a coin. As far as I can recall, they flipped beer bottle caps or some other obvious non-coins, still producing results close to 50-50.
Given the publication bias towards significant results, of any 1000 studies that tossed a coin 1,000,000 times, the 50 that found a significant difference from 0.50 will be published. Meta-analysis can uncover though that 25 would find a positive bias, and 25, a negative bias.
Read about John Kerrich for the real reasons one would want to toss a coin for a few months.
As a class activity, I had my undergrad students sand-paper a few cubes, roll them and prove to me, using Pearson $\chi^2$ test, that they indeed produced a biased die. For the time limits they had (50 to 100 rolls), you had to basically reduce one of the sides to a half to see significant results.