It's amazing to see this on the front page of HN as it came up in a discussion with my partner early in our relationship. I was saying something about how a lot of people don't understand the robustness of peer review and replication. I was gushing about how it's the most perfect system of knowledge advancement and she replied, "I mean, it's not perfect though," and then said, pretty much verbatim, the title of this article.

The problem specifically isn't so much that null results don't get published, it's that they get published as a positive result in something the researchers weren't studying - they have to change their hypothesis retroactively to make it look like a positive result. Worse, this leads to studies that are designed to attempt to study as many things as possible, to hedge their bets. These studies suffer from quality problems because of course you can't really study something deeply if you're controlling a lot of variables.

I won't say it doesn't happen. I'll just say that I haven't seen it happen in my science career on any of the projects I have worked on.

That's crazy, the serendipity.-

PS. Good point about the "shotgun" studies.-