There are many studies in many places. Imagine just for a moment that some new drug has no effect. Anyway, 1/20 of them will get a fake statistical significant result. The problem is that this study with a fake good result has a high chance of been published. The other 19 have a high chance of not been published, because the doctors lost interest, because the journal don't like null results, or other similar reasons. So if you make a review of the published studies, you will only find only/mostly the ones that get a good result due to a fluke.
This is not an imaginary problem, this is a known problem. Imagine Big Pharma has a new drug, and they want to it approved. They can hire 100 universities/hospitals to run a clinical study, suggest that only the ones with a good fluke result get published and then they get about 5 good papers about the drug. Perhaps they don't even have to make a suggestion, because it's boring to write a paper about a drug that does not work so the medic just don't write it, and it's boring to read a paper about a drug that does not work so the journal just publish it. It's the same effect, and nobody is lying ;) .
The solution for this is pre-registration. Before starting the trial, you must register it, explain what drugs you are going to use, what are you going to measure. With pre-registration, now the 100 studies are public. If you later find 5 good and 95 inconclusive or bad or disappeared, you know in the meta analysis the 5 good studies are only flukes.