A suitable operational definition of the subjective acceptability of an information source to a human user is shown to be the "probability of acceptance in a multiple-choice test." It is shown that acceptance probability relates directly to the user's statistical dependence on a given source. The notion of subjective cost of information is introduced as a concise way of defining such acceptance probabilities and a general statistical model of decision behavior used to establish the relation between expected cost and probability of acceptance. Distortion is then defined as the marginal cost of accepting a replication over that of the original source. It is shown that this leads to a way of determining distortion functions from observation of acceptance decisions. The method is illustrated with an example of image noise evaluation.