That's not strictly true. There's some recent work (as fascinating as it is incomprehensible) on generating datasets that share most aggregate properties with the actual dataset (measured through joint probability distributions), but do not reveal more than some epsilon of information about any individual contained in the original data set.
These have the potential to revolutionize private computation and analysis, as they provide provable hard (theoretical) limits on the amount of information you can learn about individuals regardless of the type of analysis performed on the proxy dataset.
