But - and this is key - these errors are usually fixable or the area is not understood well enough and requires more attention to detail. The human intuition is quite powerful.
Usually a lot of previously known errors will crop up despite that the system already works mostly as intended.
>This happens surprisingly rarely. It helps that we are trained to constantly question our entrenched assumptions.
It also helps that finding a counterexample to a widely accepted proof tends to enhance one's career and reputation.
It doesn't need to create the proof, just validate it. See P vs NP.
In some sense we won't really know whether a proof is rock-solid until we have run it through a verifier, using a proven-correct verifier on proven-correct hardware.
If theorem provers grow to be more easy to use, we may require proofs to be verified, but we aren't there yet by a wide stretch.
For an idea about how hard this can be, check out http://www.cs.ru.nl/~freek/comparison/comparison.pdf:
"In 1998, Tom Hales proved the Kepler Conjecture […] with a proof that is in the same category as the Four Color Theorem proof in that it relies on a large amount of computer computation. For this reason the referees of the Annals of Mathematics, where he submitted this proof, did not feel that they could check his work. And then he decided to formalize his proof to force them to admit that it was correct. He calculated that this formalization effort would take around twenty man-years"
(that effort is ongoing at http://code.google.com/p/flyspeck/wiki/FlyspeckFactSheet)
>Wait, so you're saying it's easier for humans to prove complex theorems than computers?
Of course! If proving theorems could be readily automated, this would have been done already. Indeed, automation has succeeded for some kinds of problems (e.g. google Wilf and Zeilberger's "WZ method" in combinatorics for a particularly successful example)
>Doesn't that mean that verifying proofs is then based on a super handwavy process where a bunch of researchers come to a consensus and say "yes, this is correct"?
Basically, yes.
>What if they all overlook some crucial aspect because of the way they were trained, or their entrenched / ingrained assumptions?
This happens surprisingly rarely. It helps that we are trained to constantly question our entrenched assumptions.
>Could we maybe have a common database of well-known proofs, axioms, concepts etc, so that you wouldn't have to rewrite common ideas from scratch?
Various books fill this purpose. But the problem is that theorems are difficult to encapsulate. For example, a theorem will say "Assume X, Y, and Z. Then A and B are true." and you need to use it in some situation where Z isn't true, but some appropriate variant of Z is, and any expert would know how to conclude that A and B are true. This is when the handwaviness comes in.
Mistakes do happen. But surprisingly rarely.