How accurate is a polygraph?
I think that really depends how you're defining accuracy.
If the question is "how accurate is a polygraph at detecting the physiological signals that it's programmed to detect?" then with modern technology I'd be inclined to believe the "close to 100%" statistic quoted in the OP.
However, I suspect that the question is really "how accurate is a polygraph at detecting a lie?" and I personally wouldn't put a lot of faith in it.
As I recall, polygraphs are based on the premise that we're all hardwired to tell the truth and lying prompts a subconscious stress reaction which manifests itself through certain physiological changes such as sweating, changes in heart rate and respiration etc.
The problem for me is that there are so many variables to take into account. Some people are more predisposed to showing physiological reactions than others (I know a couple of people who visibly blush when under pressure - absolutely no machine needed there to see what's going on) so the degree of variation in vital signs that indicates stress varies from person to person. It's not an exact science where a deviation from the baseline of x is stressed, y is not stressed and z is indeterminate.
To further muddy the waters, put yourself in the position of an innocent person being wired up to a machine and asked "did you commit this crime?" - I think most of us would show some stress reaction in these circumstances. Ambiguous questions where there's not a straight "yes" or "no" answer can also cause stress while the person makes up his/her mind on what the correct response is.
So even if a respondent shows a definite stress reaction the interpretation of the results is subjective and dependent on the examiner deciding what's a "normal" stress reaction vs a lie and on their phrasing of the questions.
If I was falsely accused of something, it's certainly not something that I'd be willing to stake my future on.
I'll admit to not having read up on it, but in principle I'm not convinced by the MRI method mentioned above either.
I'll accept that it can identify the area of the brain engaged in a complete fabrication but I have my doubts as to whether it could differentiate between recall of a real event and recall of a script / story that had been memorised in advance. IMO that distinction would very likely also be highly subjective.