The difference is deterministic systems vs. probablistic systems. In a deterministic system - what developers have historically coded (non-ML), we know the expected output or behaviour, the developer(s) code to that and the system is tested before being released to have VERY HIGH confidence that it behaves as expected. If there are some edge cases they missed, they'll likely show up as bugs, be fixed and the system is even more reliable.

In a probabilistic system, the system can behave perfectly normally and still be wrong and you still need ways to trap those "errors" etc. There are definitely use cases for this kind of architecture but we will have to learn where we should and where we shouldn't use them.

There's a lot of unknowns right now.