A test runs an action, then compares what it gets with what is expected. The oracle is that source of truth: a value written in the test, a specification, an earlier version of the software, or a rule (“the total can’t be negative”).
The oracle problem becomes central with AI: when an answer can be worded in a hundred correct ways, strict equality is no longer enough. You then use properties, annotated golden datasets or an LLM-as-a-judge.