This is at least partially a failure in publication. Once a paper is published, it's usually left up in the same state forever. If it fails to replicate, that data is published somewhere else. So when someone references the paper, and the diligent reader follows up and reads the reference, it looks convincing, just as it did when first published. It's not reasonable to expect the reader, or even the writer, to be so well versed in all the thousands and thousands of papers published that they know when something has failed to be replicated.

What we need is for every paper to be published alongside a stats card that is kept up to date. How many times it's been cited, how many times people tried to replicate it, and how many times they failed.