The do-operator erases all the arrows that come into X, and in this way it prevents any information about X from flowing in the noncausal direction. Randomization has the same effect. So does statistical adjustment, if we pick the right variables to adjust.
