Natural languages generally use such patterns as ‘A is B’ or ‘A and B are the same’ to mean an identity relation. However, it remains unclear what the cognitive mechanism actually is in using these identity statements. In this study, I embed the identity statements in attitude reports and investigate possible and impossible readings for such attitude reports as ‘John thinks A is B’, ‘John thinks A and B are the same’, etc. Intriguingly, the study reveals that felicitous ‘de dicto’ identity reports have no corresponding ‘de re’ reports. To account for this effect, I propose that the identity relation between A and B as encoded in natural languages means the contextually salient properties coerced from the expression A hold in a certain world (e.g., in the belief worlds of an attitude holder) for the individual named B and vice versa. The current analysis also suggests that natural language users can have access to some expressions in two different ways simultaneously: both as descriptions to describe certain objects and as variable names to refer to certain objects.