We consider the amount of randomness necessary in information-theoretic private protocols. We prove that at least Ω(log n) random bits are necessary for the t-private computation of the function xor by n players, for any t ≥ 2. In view of the upper bound of O(t 2 log(n/t)) [23], this bound is tight, up to constant factors, for any fixed t. For a class of protocols obeying certain restrictions, we give a stronger lower bound of Ω(t log(n/t)). We note that all known randomness efficient private protocols designed specifically for xor belong to this class. In fact we prove slightly stronger statements: we prove that on every input there is a run where the number of random bits used is large, rather than only proving that on some input there is a run where the number of random bits used is large. All our lower bounds hold for the "trusted dealer" model as well, and the Ω(t log(n/t)) lower bound for restricted protocols is tight, up to constant factors, for any t ≥ 2 in this model. In comparison, the previous lower bounds on the amount of randomness required by t-private computation of explicit functions did not grow with n for constant values of t, and our results improve the previous lower bounds for xor for any 2 ≤ t = o(log n). Our results also show that already for t = 2, Ω(log n) random bits are necessary, while it is known that for the case of t = 1 a single random bit is sufficient for privately computing xor for any number of players. Our proofs use novel techniques by which we extract random variables from a t-private protocol, and then use the t-privacy property of the protocol to prove properties of these random variables. These properties in turn imply that the number of random bits used by the players is large.