If you do a trial in a large population of a drug, device, or clinical practice that does nothing- a perfect placebo- you'll see a variety of effects: statistical noise.
If you do a trial in a large population of a drug, device, or clinical practice that has an effect-- you'll measure that effect, plus the statistical noise.
You can't generally tell for any individual whether the drug helped or hurt. But you can tell that more people did well (or badly) in group A than group B.
You can't even really know exactly how big the effect is precisely: just a range of likely effect sizes.
The more things you try to measure to more precisely zero in on an effect, the greater the chance that statistical noise spuriously makes one of these look important (and the larger the effect must be to be reliably measured). https://xkcd.com/882/
