We can increase our confidence by trying a few examples, but only a little, since the scope for failure is so huge.
Hammering a system with thousands of edge-cases, interleavings, previous regressions, etc. can give us a bit more confidence; but takes a huge amount of time for a person to do manually. It's pretty easy to automate though, and there are a load of libraries and frameworks that make it even easier.
If we want even more confidence in a system, we can verify it against a machine-checkable specification. That's still too hard to automate in most cases.
In any particular situation we need to find an appropriate balance between the cost of increasing our confidence in the system (i.e. how much effort we're willing to spend on prove ourselves wrong) versus the cost of failure.
Automated tests (especially property checks) are so low-cost and high-yield that I find them always worth doing, unless the consequence of failure is near-zero.