The question is a quote from the Microsoft documentation. but has a second purpose here. See the article written by two Microsoft researchers. http://msdn.microsoft.com/en-us/library/cc441928.aspx
Based on the fact that RAND is deterministic when used with a seed, they are able to get a speedy sample from huge tables using the NEWID() function in conjunction with the RAND function. If it weren't deterministic, the sample would not work. Attempt it on a table with several million records. The sampling is extremely fast. The algorithm is faster if the keycol1 refers to a clustered ID table.
SELECT * FROM Table1
WHERE (ABS(CAST(
(BINARY_CHECKSUM
(keycol1, NEWID())) as int))
% 100) < 10
Jamie