A smoke alarm makes a decision every second, with H0: "no fire." It can be wrong in two ways. If it shrieks over burnt toast, it rejected a true H0: a Type I error, a false alarm. If it stays silent during a real fire, it failed to reject a false H0: a Type II error, a miss. When it is right, it is either quiet with no fire or sounding during a fire; the chance of that second correct outcome is called power.
Lay the four cases out as a grid, with the truth (H0 true or false) across the top and the decision (reject or not) down the side. The two errors sit in opposite corners. An error's name depends on the truth and the decision together: "reject" is a mistake only if H0 was true, and "fail to reject" is a mistake only if H0 was false.
The two error rates pull against each other. A more sensitive alarm catches more fires (smaller β) but goes off for toast more often (larger α). In a test, lowering α from 0.05 to 0.01 makes false alarms rarer but demands stronger evidence, so real effects are missed more often. The way to lower both at once is better information: a larger sample shrinks the standard error, so true and false nulls produce clearly different data.
A p-value is not an error rate. One study with p=0.03 has not made "a 3% error"; it either made an error or it didn't. α is the long-run rate of Type I errors when H0 is true, a property of the rule, not of one result.
Think of it like this. A referee's whistle faces the same two errors. Calling a foul that never happened is a Type I error; letting a real foul go is a Type II error. A strict referee misses fewer fouls but makes more false calls. Only a better view of the play, like video replay, cuts both, just as a larger sample does for a test.
Core idea. A Type I error rejects a true null (false positive); a Type II error fails to reject a false null (false negative).