Flip a fair coin and track the share of heads. After 10 flips you might have 7 heads, a share of 0.70. Keep going. After 100 flips the share is very likely to be between about 0.40 and 0.60; after 10,000 flips, between about 0.49 and 0.51. The running proportion jumps around at first and then flattens toward 0.5. The law of large numbers says this flattening happens for any sequence of independent trials: Xˉn→μ as n grows.
The mechanism is dilution. Suppose those first 10 flips gave 7 heads and 3 tails, 4 more heads than tails. The next 990 flips are fair and give about 495 heads, so the total is about 7+495=502 heads in 1,000 flips, a share of 0.502. The early surplus did not go away; heads still lead by about 4. It just matters far less once it is spread over 1,000 flips. Nothing pushes the coin toward tails to restore the balance.
Two ideas get confused with this law. The gambler's fallacy claims that after five heads, tails is due; the law says no such thing, because each flip stays at 21. And the law is about averages, not counts. The typical gap between the number of heads and tails grows from about 2 or 3 after 10 flips to about 80 after 10,000, even though the proportion ends up far closer to 0.5. The average settles; the count gap does not have to.
Think of it like this. Picture a few drops of food colouring in a glass, with clear water poured in and in. The colour fades as the volume grows, even though the drops are never taken out. An early lucky run in a long series of trials behaves the same way: it is diluted, not undone. The picture leaves out one thing: each new trial adds a little fresh randomness of its own.
Core idea. As trials grow, the sample mean converges to the true population mean — randomness averages out.