This is when subsequent studies show association between the the SNP and trait, but for different alleles. Put another way, an increased risk turns into a reduced risk, or vice versa. Rather than falsify the whole concept, it is argued these cases are frequently proof of actual causal variant or variants within vicinity, or LD of the SNP pinpointed in the study. This is actually quite expected at current state, because the genotyping chips are designed to only test for common variants thousands of bases apart. As such, the association could even disappear entirely when more diverse, admixed populations are studied.
For example, for the HumanOmniExpress-24 beadchip currently used, they are expected to capture 73% of SNP's with MAF > 5% within LD r^2 < 0.8 for CEU population. With most harmful variations expected to be rare, it's clear that most of them won't be directly tested. Even most of the common variants will only be tagged within LD, suggesting they are most commonly inherited together but the exact alleles will depend on population. Next generation sequencing will be alleviating thris problem, both because more and more people having their whole sequence and thus each SNP, and because the results will allow imputing the missing SNP's from beadchip results with greater accuracy.
Another problem is continuing view of genetics by public and even some researchers through the Mendelian framework, where a single or at most couple of mutations are expected to account for majority of a trait. Biology has much redundancy, where other mechanisms can take over from a failed one, and complex inter-locked dependencies. Certain variation can be harmful only in the presence of one or more other variations (which may not have been tested), or only in certain environmental conditions, such as a specific diet or age. These are mostly factored in to the risk probabilities (which is why they are only probabilities), but it is not unusual to see the risk probability change or even reverse in different populations even when it's the right SNP. Computational advances will allow better taking into account effects of other SNP's and environment.
Some possible flip-flops:
[PMID 17273975] No gene is an island: the flip-flop phenomenon