If we take a null hypothesis that a medication has no effect, but its actual effect is very small, we would need a very large sample in order to have any chance of dependably giving the true positive result it's possible that the required sample size is too large to be theoretically possible. We also often get false negatives because the test doesn't have good enough resolution. Similarly, it's possible that we got exactly 500 heads with a coin that is weighted so that the chance of a heads is only $0.02$, an outcome that should occur only about 31 out of every googol times. If we got only 5 heads, it's still technically possible that it is a fair coin and we just got an outcome that statistically should occur less than once every googol times we run the test. Suppose we want to test if a coin is fair by flipping it 1 thousand times. We can get any type of error by bad luck. There can be a few causes, even if the test is properly run. However, the coin could still be biased and flip heads once. Since this is bigger than 0.05 I would accept the null hypothesis and conclude that the coin is not biased. In this universe I would calculate the P-value to be 0.21875. In an alternative universe I might have done the the same experiment, but instead have flipped 5 tails and a single head. This would be an example of a type 1 error. What is important to note however is that the coin may not have been biased, but it could just be random chance. So in this case we would reject the null hypothesis and conclude that the coin is biased. How the standard cutoff for P-values is $P<0.05$. Say I was doing a statistical test to determine whether A coin was biased or not.

Type 1 error stats