I previously read an article about random number generation manipulation which was quite fascinating:
"Count population profiles in engineering anomalies experiments" by Jahn et al. 1991,
in which they showed a very slight deviation of electronic coinflips towards the desired outcome of an operator focusing on the machine.

Since something like this is easy to test yourself I wrote a simple python script to do so, here it is:
$<br /> \text{\#\!/usr/bin/env python}\\\text{import sys}\\\text{import numpy as np}\\\\\text{def main(argv):}\\\text{\quad score=0}\\\text{\quad it = 0}\\\text{\quad while(1):}\\\text{\quad it += 1}\\\text{\quad\quad score += np.random.randint(2)}\\\text{\quad\quad if (it \% 200000 == 0):}\\\text{\quad\quad\quad print(score / it)}\\\\\text{if (\_\_name\_\_ == &quot;\_\_main\_\_&quot;):}\\\text{\quad\quad main(sys.argv)}<br />$

The script keeps flipping coins (casting RNG values to 0 or 1) and outputs the average over all coins flipped so far every 100000 throws. I was consistently able to make the result go above or below 0.5 by a deviation of about 0.001-0.002 for extended periods of time while concentrating on it. Give it a try.