Instead of asking for example "Did you rob the bank?" and interpret the response as "lying" or "not lying" they can be asked "Was the getaway car white?" "Was the getaway car red?" "Was the getaway car blue?". By asking several of these you can statistically prove the suspect is guilty if they responded differently to several detail questions about information only they knew. Tthis doesn't help prove innocence (because people can beat the machine) but it reduces false positives and can therefor prove guilt (because it's unlikely a nervous person would randomly be more nervous when the true details are being asked vs. the decoys).
Of course to be done right the operator must be "blind" (not know the answers and subconsciously tip the suspect off) and other statistical safe-guards need to be in place, so how useful in the real-world this is I don't know.