In science, mistakes happen: bad experiment design, noisy measurements, incomplete data, expensive experimentations, and other factors. Sometimes theories are wrong because we don't know enough.
In mathematics, mistakes can be found right in the paper or traced back to the original, if that paper employs some erroneous results. We have complete information, and the number of mistakes depend mostly on our desire to find them.
The question is: What approximate percentage of published papers contains mistakes? Maybe some data available. And if not, what does your experience say in this respect?
P.S.: By mistakes, I mean mistakes that invalidate proofs by (1) being incomplete, e.g., not all cases considered, (2) proving a correct conjecture with incorrect reasoning, (3) proving an incorrect conjecture.
By "published," I mean papers published in peer-reviewed math journals.
As in @HaraldHanche-Olsen's comment, the question itself is more ambiguous than the questioner may realize. But this meta-point itself deserves to be considered. That is, "correctness/incorrectness" is not objective, _in_practice_, because (as @Harald H-O notes) arguments perceived as incomplete by outsiders may be seen as completely adequate by experts. To elaborate slightly: even the issue of what's omitted may itself be tacit, thus adding a layer of impenetrability for novices. Similarly, statements that are false-as-stated, but irrelevantly so since they are only applied in a sub-case or variant that is correct, and easily verifiably by an expert, will not receive erratum treatment. At best, but also "worst", the virtue of a paper can be in its over-all narrative, which, if sufficiently intelligible to an expert, can be considered "correct" by experts, even if the thing is riddled with literal "errors", as many expert-perceptions have substantial "error-correcting" tendencies built in. And, of course, line-by-line "correctness" does not necessarily prevent other sorts of "mistakes", especially those that are not "localized" in writing.