I vaguely remember the same advice, it's pretty old. How you use the randomness is test specific, for example in math_add() it'd be something like:

  jitter = random(5)
  assertEqual(3 + jitter, math_add(1, 2 + jitter))

If it was math_multiply(), then adding the jitter would fail - that would have to be multiplied in.

Nowadays I think this would be done with fuzzing/constraint tests, where you define "this relation must hold true" in a more structured way so the framework can choose random values, test more at once, and give better failure messages.