It seems that one continuously hears about individuals passing or failing the lie detector, and despite many questions regarding its veracity, people still assume that there is a scientific basis for its use.

However, lie detection, or polygraphy is not based on science.  In fact it isn't based on much of anything, except psychological manipulation of the subject under the guise that taking the lie detector may cause them to confess, because they believe it is based on science.

In short ... it's voodoo psychology.

The first problem is in what the polygraph, or lie detector actually measures.  The lie detector measures blood pressure, heart rate, respiration, and galvanic skin response or skin conductivity.  Therefore the first scientific problem that must be addressed is what the correlation with these physiological measurements has to do with deception.  It seems clear that the objective is to measure anxiety, but this presupposes that individuals become anxious for the same reasons and under the same circumstances.  Since there is no context for these measurements by themselves, the thinking is that by asking some control questions, a kind of baseline can be established against which deception can be measured.

However, the problem with that, is that it presumes that individuals react in a predictable manner with specific traits associated with deception.  In other words, the underlying "theory" is that physiological reactions will occur because of "guilty knowledge" in whatever is being investigated.

This brings us to another problem with the polygraph and that's the definition of deception.  In this case we are looking at a specific psychological condition in which the subject must keep two entirely separate but plausible scenarios in their mind during the interrogation.  One must reflect the "truth" while the other is the "lie" or the story that is intended to deceive the interrogator.  Of course, this renders the polygraph totally inappropriate in those situations where an individual believes what they are saying.  So, in instances of UFO witnesses or witnesses to paranormal phenomenon, the use of the polygraph is completely irrelevant unless the sole purpose is to establish whether the subject is simply making up the story.  Truth is entirely subjective and the subject that believes what they are saying is not being deceptive regardless of how wrong their information may be.

Despite some of the obvious difficulties with such assumptions, there has been no scientific testing that establishes that such a correlation exists, let alone exists consistently enough to be interpreted as being deceptive.  People may be anxious because they are lying, but they may also become anxious because they are afraid the examiner thinks they are lying.  Instead of "guilty knowledge", they may experience a guilty reaction over an unrelated event.  More importantly, it fails to consider the issue of a psychopath that may not have any anxiety over their crimes and therefore escape detection.  Essentially the entire premise on which lie detection is built consists of a variety of assumptions which have never been demonstrated to be valid.

Presumably some of these difficulties can be overcome because of the expertise of the examiner.  Not really.

Here is a link to a checklist for finding a qualified examiner.  

Apparently the requirements to become a polygraph examiner are somewhat vague.  Despite the claim that such an examination is a psychology test, there is no requirement that one be a psychologist, or even have a psychology degree.
"Polygraph examiners often have experience in law enforcement or investigations, but a background in psychology or behavioral science is also beneficial. Polygraphists work in the public and private sector, and many are employed by local, state and Federal government agencies."
It would appear that the actual training for becoming a polygraph examiner involves approximately 380 hours (about 9.5 weeks) of training at an accredited Polygraph school.

Claims of accuracy abound, but generally this is promoted by polygraph examiners, so it is a bit of a biased perspective.  Even if we were to accept an accuracy rating of 92% to 98%, this means that of 100 suspects, 2 to 8 would return bogus results.  While the actual guilty party might be in this group, it doesn't bode well for the innocent individuals that will have been accused as well. 

Equally well-known is the case of Gary Ridgway [aka, the Green River Killer] who took and passed a polygraph test in 1984.  As a direct result of this "passing", the DNA results taken from him in 1987 were not analyzed until years later, ultimately leading to his arrest in 2001.  The polygraph had allowed a serial killer to go free for an additional 17 years because of an improper result.
"What this research tells us is that if a polygraph exam is done properly, by a skilled and experienced examiner, using the best methodologies, techniques and equipment, accuracy can be as high as 96%."
"OTA concluded that the available research evidence does not establish the scientific validity of the polygraph test for personnel security screening. OTA was able to identify only four studies directly relevant to personnel security screening use (one by DOD). None of these studies specifically assessed validity of polygraph testing for the purposes proposed by DOD or the administration, and all had serious limitations in study design."
Psychologists have repeatedly told U.S. courts that polygraph tests--popularly thought to reveal a person's truthfulness through assessment of physiological states--are theoretically unsound and not valid in assessing honesty.
Yet, what is the basis for assessing such a high degree of accuracy?  Interestingly, it is based almost exclusively on the "confession criteria".  In other words, an individual that fails the polygraph is further interrogated, ultimately leading to a confession.  This is used as one of the primary basis for establishing that the lie detector was accurate.
"Because polygraphers seldom discover ground truth except as a consequence of post-test confessions, and because diagnoses evaluated in this way are almost invariably verified as correct, the typical experienced examiner will accumulate a personal record of almost unblemished accuracy (p. 202)."
Under these circumstances, the criterion (the confession) and the test outcome (deception indicated) are not independent. The method thus virtually guarantees that the two will match, ensuring 100% “accuracy.”
" It is noted that the validity of polygraph testing has yet to be established. The present review suggests serious problems with both the theoretical rationale underlying use of polygraph tests and the quality of available evidence supporting the validity of such tests."

"Most of the respondents believed that polygraphic lie detection is not theoretically sound, claims of high validity for these procedures cannot be sustained, the lie test can be beaten by easily learned countermeasures, and polygraph test results should not be admitted into evidence in courts of law."
Polygraph is fundamentally an abuse of science, because it is based, at least in part, on the ignorance of the populace in eliciting a confession because they believe it has scientific merit.  This is not a game, but a technology that has the potential to destroy an innocent person's life, while letting the guilty go free.  The notion that some degree of accuracy is all that is necessary to achieve an objective is essentially an argument that it is valid to destroy people's lives in an effort to apprehend the guilty.  Yet, we see that the guilty may just as readily escape such detection, so even this rationalization fails.

Whether anyone likes it or not, the human mind is complex and the simplistic notions that truth and deception are simply binary is the height of psychological naivete.  An individual may be lying, but we can't determine about what.  We can't determine the motivation, nor even whether the lie is relevant.  We may also have the response of individuals that feel guilty at the mere mentioning of various acts, so that they fail the polygraph despite being innocent.

This is an area that law enforcement, security personnel, and many businesses would love to have a crystal ball to manage, but using faulty science to achieve such an objective is not the vehicle.  For those under scrutiny, while it may be tempting to volunteer for a polygraph to establish one's innocence, it is equally important to remember what the consequences could be from a wrongful accusation.  The problem is that the innocent invariably believe that they will prevail.  This is also a lie.
Additional references:

For an alternative view of deception detection, the work of Paul Ekman provides some interesting reading.