EDIT: Not quite that many, more like around 9, 18, 27 or so, see below.
Humans aren't like weighted dice. For example, it's certainly conceivable that a few pairs of these humans misunderstand the goal such that they have a 0% chance of ever choosing numbers that change the respective ordering. Add to that the higher probability that many of these pairs of humans will just toggle their orders each round.
Edit: clarification-- the humans don't even have to misunderstand the rule, just the upshot of the process.
Thus, I recant "around 20, 40, 60 or more" and replace it by "around 9, 18, 27 or more" numbers. Probability that you have to reject any resulting hex digit is obviously 3/8.
With b bits you are subdividing the (0,1) interval into 2^b regions, and only need more bits if one of the 9 multiples of 0.1 land in the interval you've picked. As b increases this probability drops exponentially.
It's a fair point that the number of "numbers" depends on how low the entropy of the source is, but in the link the probability of collision wasn't massive.
The OP is silly though; if you know the distribution there are way better techniques. They don't seem to worry about how they measure it.
Well, if you're really interested in generating random numbers using a group of people, then yes, this may be a problem.
But you should take this as an illustration of the more general problem of generating uniform random numbers from a distribution that is not.
If you think about it this way, author's solution can be applied in a number of practical scenarios. For example, if you want to generate a hash values for objects using properties that are not uniformly distributed. Or if you want to generate random numbers from a physical artifact that is not perfectly uniform.
We could test by asking people to pick numbers less than 100. I bet people would focus on odd numbers, especially those greater than 50, not ending in 5 or not having both digits the same.
Data and light analysis is here: https://docs.google.com/spreadsheets/d/1Dh0wiTCRkBhckWGXtjZg...
I definitely overpaid on Mechanical Turk per response, given how lightning quickly the data came in. (I decided to pay $0.10/HIT for 50 responses and got 68 responses in 8m26s.) I suspect that offering a nickel would have gotten the survey filled in under half an hour still...