[IggleSniggle]

This bypasses the desired capability, that of getting info from the hard disk.

If you connect input in a feedback loop with an output and train on a particular output, of course you’re going to get the output you’re training for. It’s just not going to bear much resemblance to the data that’s on the hard disk without cooperation from the host system.

[alboy]

It depends on how much control the host system has over the output (can consciously drown out the signal). Suppose you just show YES and NO in big flashing letters while asking your subject the questions you already know the answers to and measure the output when training the device. Then it boils down to whether it can pick up what you "really think" over what you "try to think" better than the current generation of polygraphs (which is very bad at its job). So these technical specifics would decide where exactly it falls within the range from "comically unreliable" to "dystopian nightmare".