[Swordmaker]

That makes sense. The best way to get a cryptographically sound random key is to pull in high entropy sources like noise in microphones or noise in camera CMOS sensors. Basically make it impossible for an adversary to recreate. If they store that as a seed for a pseudorandom number generator that would be a sound way of restoring the AES key when it is needed.

The problem with any computer is finding a way to generate a truly random number when one needs one that is not some how infected with a human's biases toward something. Even this method has a built in bias against low random numbers simply because it uses three or more inputs and combines the sampled "noise" which will result in a seed that inherently will result in a random number seed higher than what a truly random number could be if you accept that a truly random number could include "1" in its set. That choice was human based in its bias toward larger random number seeds, but it's certainly an acceptable bias.

[palmer]

The probability distribution of those sources may not be even, and you want to have as close to white noise (uniform distribution) as possible or a completely even chance of any number in the range being picked. There are a few tricks to make distributions more even, the main one is hashing since hash output is uniform.

Now I see wikipedia lists hashes as one of the "Software whitening" tricks to make random numbers more uniform. But it is certainly a nontrivial issue and hashing is not the first choice.