We consider an agent who chooses from a set of options after receiving some private information. This information however is unobserved by an analyst, so from the latter’s perspective, choice is probabilistic or random. We provide a theory in which information can be fully identified from random choice. In ad- dition, the analyst can perform the following inferences even when information is unobservable: (1) directly compute ex-ante valuations of option sets from ran- dom choice and vice-versa, (2) assess which agent has better information by using choice dispersion as a measure of informativeness, (3) determine if the agent’s beliefs about information are dynamically consistent, and (4) test to see if these beliefs are well-calibrated or rational.