It's important to keep experimental subjects in the dark; this isn't hard when they're rats, but can be challenging when they're humans. An informed human being can do a lot of damage to your experiment. A misinformed one can do even more.
An ideal experiment calls for the experimenter to disappear. Subjects are meant to respond to the conditions of an experiment as if they weren't being watched. Sadly, with the current state of consent laws, subjects have to know they are being watched, so the next best thing, experimentally speaking, is to give them no clue as to why they're being watched. Since the experimenter is exactly as human as the subject, that's not an easy task. Researchers have a goal in mind when they conduct experiments, and they can convey that goal through their voice, their actions, and the conditions they set up during the experiment. The conglomeration of all the cues they give to let the subject in on the hypothesis of the experiment is called the demand characteristics of the experiment.
It's surprisingly easy to let people know what an experiment is meant to test. One study on demand characteristics involved setting up a dummy study. (Yes, a study-within-a-study, Inception style.) The "experimenters" were asked to give the subjects tasks which required them to think of a physical object. In one section of the study the subjects were asked to visualize the object, and in one section they worked with the physical object in front of them. Meanwhile, half the experimenters were given one hypothesis for the study — that subjects with the physical object would do better. The other half were given the hypothesis that people would do better when visualizing the object. The experiment was real, the subjects uncoached, but the two groups of experimenters got opposite results. The subjects did better when the experimenter expected them to do better. When reviewing recordings of the procedure, the actual researchers saw that the experimenters gave slower and clearer instructions for the part of the study during which they expected the participants would do better. The participant, given clearer instructions, more time to think about them, and probably sensing the emphasis of the experimenter, did better.
The Good Subject Effect Versus the Screw You Effect
Once a subject guesses what the psychology experiment is all about, there are all kinds of ways they can skew results. Naturally, they want to make themselves look good. In one study examining the effects of demand characteristics, two groups were asked to measure the relative tapping speed of both of their index fingers. One of the groups was told that students at Yale didn't have much of a gap in performance between their dominant and non-dominant hands. Guess which group ended up with similar tapping speeds for their left and right hand?
Generally, the subjects who guess demand characteristics will separate into two groups. One group wants to help the researcher out. They want to be part of a successful experiment. They will try to play to the researcher's expectations or hopes. This "good subject" effect is also found by people conducting surveys, who will often notice that people will tell them whatever it seems like they want to hear.
On the other side of the coin is the "screw you" effect; it happens when people get discouraged or angry. They try to rile up the experimenter by screwing up their experiment. The Milgram Experiment — during which participants were made to think they were shocking someone to death — quite possibly led to a few cases of the screw you effect. They weren't particularly kind to human nature. One man, asked why he walked out rather than shock someone to death, said he knew he was being "lied to" by the experimenter and wasn't going to stand for it. This, surprisingly, makes the experiment even darker — what we might think shows a conscience is actually a tantrum.
Overall, though, there is one thing that's more likely to produce the scew you effect than any other — dislike of the experimenter. The more obnoxious (or just different) the experimenter was from the subject, the more the subject decided to try to wreck the experimenter's hopes and dreams. So clearly, if you're going to do human research and want to publish a lot of successful studies, try to be nice to your participants. You'll still get crap data, but you'll get publishable crap data.