From my understanding, typically when you get random noise from the environment, you aren't exactly sampling the -actual- stuff -- there may not be any sounds, for example. Quiet rooms are not unusual. What you do is what an anonymous coward mentioned - you have a sensor that reads the temperature as 72.7230283037 degrees F. It's not REALLY that accurate, the last few digits are basically random. Maybe the light sensor says 43.584723028 lux, plus or minus 3 lux. It has more -precision- than it has -accuracy-. The decimal part is bullshit, random.
Obviously the device has to convert the analog measurement into a digital number that the CPU can manipulate. It uses an analog-to-digital converter to do this. A typical ADC might convert the reading from a microphone into an 8-bit integer, ranging from 0 (approximately silent) to 255 (very loud). A 16-bit ADC gives a range from 0-65,535. Phone calls traditionally use an 8-bit ADC because that's sufficient for voice. Voice calls don't have to be hi-fi, especially after you remove the highest and lowest frequencies which actually hinder intelligibility.
It would be wasteful for the manufacturer to use a 24-bit ADC on the microphone if they're only going to use 8 bit samples for phone calls, so they may only use an 8-bit ADC. In that case, the range of loudness is only 0-255, with no decimals, meaning any quiet room will register 0. There are no extra extra digits with "random" values. If the manufacturer intended to use it as a source of randomness they could use a 24-bit or 32-bit ADC, knowing that the smallest bits will be roughly random.