A fair die has six faces, so each roll gives 1 to 6, each with probability 61, and no roll tells you anything about the next. A computer cannot roll a die, so it fakes one with a formula. A tiny example: start from a seed, say 7. At each step, multiply by 5, add 3, and keep the remainder after dividing by 16. From 7: 7 × 5 + 3 = 38, remainder 6. Then 6 × 5 + 3 = 33, remainder 1. Then 8, then 11, then 10. The numbers jump around with no obvious pattern, yet they are completely fixed by the seed: start at 7 again and you get 6, 1, 8, 11, 10 again.
That is a pseudo-random number generator. Real ones use far bigger numbers, so the sequence runs for billions of steps before repeating and passes statistical tests for randomness. To turn its output into a die roll, a program maps it onto 1 to 6, and the mapping has to keep every face equally likely; a careless one can favor some faces.
Random does not mean evenly spread in the short run. Roll a fair die 12 times and getting no 6 at all happens about 11% of the time, and a run of high numbers does not make a low one 'due', because each roll is independent. Random numbers are also not the same as a simulation. A random number is one ingredient, such as whether one customer arrives this minute; the simulation is the whole model that uses thousands of them.
Think of it like this. A pseudo-random generator is like a very long shuffled playlist: the songs seem to come in no order, but the order was fixed the moment the shuffle was made. Start the same playlist from the same point, the seed, and you hear exactly the same sequence. A truly random source, such as radioactive decay, has no playlist at all.
Core idea. Random numbers help models include uncertainty, variation, and chance.