For example, if you run an experiment 4 times, and three of those times it works in a similar way, how do you treat that fourth data set? Can you write it off to some sort of experimental mistake or oversight? Did you mix up the concentrations, or were the cells somehow different to begin with? When you get funded or get fired, you drop that data set, and hope the other three are meaningful. There is no time or reward for figuring it out. Actually, you are punished for doing so. If you include that data set, idiot reviewers will say that your data is not 'significant' and not let you publish your results.

Yup. Null results are almost never publishable, which is ridiculous. Every lab I've been in has dozens of old experiments that were never published because they "didn't work", which makes you wonder about any paper you read. If there are plenty of unpublished versions of one study with a null result, then a published study of the same experiment with a result probably found a difference by chance. But we have no way of knowing whether an identical experiment has been done before and by how many people and how many times unless it's published. But you can't publish null results and the cycle continues.