Imagine testing a new headache medicine by giving it to a hundred people with headaches and asking how they feel an hour later. Most of them will report feeling better. The trouble is that most headaches fade within an hour anyway, and many people feel better simply because they have taken something they expect to help. A test designed this way cannot separate the medicine’s effect from the effects of time and expectation.

Researchers solve this problem with a comparison group. In a controlled trial, participants are assigned at random to one of two groups. One group receives the new treatment, and the other receives a placebo: a pill or injection that looks identical but contains nothing active. Because chance alone decides who goes where, the two groups should be alike in age, health and every other respect except the treatment itself. If the treatment group improves more than the placebo group, the difference can reasonably be credited to the treatment.

Random assignment is not enough on its own. If participants knew which group they were in, those receiving the placebo might expect little and report little, which would exaggerate the medicine’s apparent benefit. For this reason, trials are usually blind: participants are not told which pill they received. In a double-blind trial, the researchers who measure the results are not told either, because a researcher who knows may, without meaning to, judge the two groups differently.

Placebos also raise an ethical question. When a proven treatment for a serious illness already exists, giving some patients an inactive pill would mean withholding care that is known to work. In those cases, researchers generally compare the new treatment with the existing one instead of with a placebo. The question then changes from “Does this work?” to “Does this work better than what we already have?”

To people outside medicine, this careful design can seem excessive. It exists because expectations are powerful enough to change what people report, and a trial that cannot separate expectation from effect cannot answer the one question it was designed to ask.