A lock allows you to force multiple threads to access a resource one at a time, rather than all of them trying to access the resource simultaneously.
As you note, usually you do want threads to execute simultaneously. However, imagine that you have two threads and they are both writing to the same file. If they try to write to the same file at the same time, their output is going to get intermingled and neither thread will actually succeed in putting into the file what it wanted to.
Now maybe this problem won't come up all the time. Most of the time, the threads won't try to write to the file all at once. But sometimes, maybe once in a thousand runs, they do. So maybe you have a bug that occurs seemingly at random and is hard to reproduce and therefore hard to fix. Ugh!
Or maybe... and this has happened at a company I worked for... you have such bugs but don't know they're there because they are extremely infrequent if your computer has only a few CPUs, and hardly any of your customers have more than 4. Then they all start buying 16-CPU boxes... and your software runs as many threads as there are CPU cores, so suddenly you're crashing a lot or getting the wrong results.
So anyway, back to the file. To prevent the the threads from stepping on each other, each thread must acquire a lock on the file before writing to it. Only one thread can hold the lock at a time, so only one thread can write to the file at a time. The thread holds the lock until it is done writing to the file, then releases the lock so another thread can use the file.
If the threads are writing to different files, this problem never arises. So that's one solution: have your threads write to different files, and combine them afterward if necessary. But this isn't always possible; sometimes, there's only one of something.
It doesn't have to be files. Suppose you are trying to simply count the number of occurrences of the letter "A" in a bunch of different files, one thread per file. You think, well, obviously, I'll just have all the threads increment the same memory location each time they see an "A." But! When you go to increment the variable that's keeping the count, the computer reads the variable into a register, increments the register, and then stores the value back out. What if two threads read the value at the same time, increment it at the same time, and store it back at the same time? They both start at, say, 10, increment it to 11, store 11 back. So the counter's 11 when it should be 12: you have lost one count.
Acquiring locks can be expensive, since you have to wait until whoever else is using the resource is done with it. This is why Python's Global Interpreter Lock is a performance bottleneck. So you may decide to avoid using shared resources at all. Instead of using a single memory location to hold the number of "A"s in your files, each thread keeps its own count, and you add them all up at the end (similar to the solution I suggested with the files, funnily enough).
