99% polygraph accuracy is marketing. What the report everyone cites actually says

If someone promises you that the polygraph is 99% accurate, you are being sold to. The largest review of the research, carried out by the American Polygraph Association (APA) itself, gives 85–89%. And even that figure comes with caveats that are rarely mentioned.

I am a polygraph examiner. I have been in practice since 2012 and have conducted more than 6,000 tests. I read all 113 pages of the original report, and below I show what it actually says — with page numbers, so that you can check me.

Need an actual test rather than theory? Take the VerifEye online test — from 2,500 UAH (≈ €50), with no office visit, or calculate the cost of a polygraph test.

In brief

• Polygraph accuracy is 85–89%: 89% when a single event is tested, 85% when several issues are tested at once (APA meta-analysis, 2011, p. 203).
• This figure leaves out tests that end without a decision. Roughly every eighth or ninth test ends as “inconclusive” (11–13%, p. 203).
• The polygraph is wrong about truthful people more often than about deceptive ones. Out of 100 truthful people, about 12 are wrongly judged deceptive; out of 100 deceptive people, about 8 are wrongly judged truthful (calculated from Table 7, p. 246).
• 99% accuracy comes from studies run by the developers of the techniques themselves. The APA called those results statistical outliers (pp. 248–250).
• There is no “most accurate technique”: no difference in accuracy was found between the twelve techniques reviewed (p. 257).

Where does the 85–89% figure come from?

In 2011 an APA committee gathered 38 studies in which it was known for certain who had told the truth and who had not. In total, 295 scorers evaluated the recordings of those tests and produced 11,737 scored results (pp. 199, 203). It is the largest review carried out by the profession itself.

Source: Polygraph, 2011, 40(4), p. 203 and Table 7 on p. 246.

What does “89%” really mean?

89% is the share of correct decisions among the tests where a decision was made. Tests that ended as “inconclusive” are not part of that figure. The report says so directly (p. 203, footnote 1).

If everything is counted, the picture for 100 people tested about a single event looks like this.

My calculation from Table 7 of the APA report (p. 246). How I did it is explained at the end of the article.

So the honest answer is this: out of ten tests, roughly eight give a correct decision, one gives a wrong one, and one gives none.

Can a polygraph “accuse” an honest person?

Yes. According to the report, the polygraph wrongly judges about 12 out of 100 truthful people to be deceptive. It wrongly “clears” deceptive people less often — about 8 out of 100. In other words, the error more often goes against the honest person.

For several traditional scoring systems the report admits even more: their ability to recognise a truthful person is not statistically different from chance (pp. 242, 244, 251). Those systems nevertheless stayed on the list of “validated” techniques.

That is exactly why an examiner’s conclusion is an argument, not a verdict. And it is why the conclusion is not, on its own, evidence in court.

Why do some say 99% and others 85%?

The report includes two proprietary techniques — IZCT and MQTZCT — with an accuracy of 99.4%. The APA committee called these results statistical outliers and explained why (pp. 248–250):

• the studies were run by the developers themselves, who have a financial interest in the techniques;
• in one study a participant whose result would have spoiled the perfect accuracy was removed;
• for another, the committee was at first given scores only for the tests in which the scorer had not made a mistake.

And in footnote 59 (p. 259) the authors wrote the most candid sentence of the whole report. In my own words: every lie-detection method of the past hundred years that was studied by its own developer showed accuracy of around 99% — and each time independent checks were either never done or gave much lower figures.

Which polygraph technique is the most accurate?

None. The report compared twelve techniques and found no statistically significant difference between them. The authors explicitly advise against ranking them and suggest paying less attention to names (p. 257).

The phrase “we use the most accurate technique” is marketing. What affects the result is how the examiner worded the questions, how the pre-test interview was conducted and how the recording was scored.

Can the report itself be trusted?

In part. The authors do not hide the weak points, but they do not put them on the first page either.

• Some of the “tests” are computer simulations. Eight samples consist of 100 simulated cases each, not real ones (pp. 214–221, Appendix A on p. 269).
• The authors assessed their own work. The principal investigator of the meta-analysis is a co-author of 18 of the 38 included studies. The report acknowledges this and adds that without those studies some techniques would not have made the list (pp. 256–257).
• Real cases were confirmed mostly by confessions. If a person did not confess, the case often never enters a study. The authors write that this may inflate accuracy (pp. 197, 253).
• The figures apply to healthy adults who do not use countermeasures. Studies involving fatigue, intoxication or trained countermeasures were left out; there were no juveniles among the examinees (pp. 198, 254, 256).
• Study quality was not weighted, and laboratory and field cases were counted together (pp. 213, 255).
• The report contains arithmetical discrepancies. For example, the abstract gives 3,723 examinations, while Table 2 gives 4,367 (pp. 203 and 237).

For these reasons I treat 85–89% as an estimate under good conditions, not as a guarantee. The APA report itself notes that, according to the US National Research Council (2003), no new technology is ready to replace the polygraph, yet the polygraph remains far from perfect (p. 259).

What should you do with this?

If you are ordering a test. Ask the examiner three things: what accuracy they quote and where the figure comes from; what they do when the result is inconclusive; and whether they will write “inconclusive” in the report. Anyone who answers “I get 99% and never have inconclusives” either has not read the research or is counting on you not having read it.

If you are the one being tested and you are telling the truth. There is a risk of error. It is reduced by a good night’s sleep, a calm pre-test interview and an honest conversation about illnesses and medication. More in the article on how to pass a polygraph.

What I do when the result is inconclusive. I say so plainly and do not “stretch” the conclusion towards a convenient one. Then I suggest a repeat test or a different method — for example EyeDetect, which works on a different principle. Two different methods with the same result are more convincing than one.

I quote 85–89% to my clients and will not quote more until an independent study shows different figures. Methods and prices are on the prices page.

FAQ

How accurate is a polygraph in per cent?

85–89% among the tests where a decision is made: 89% for a single event, 85% for several issues. A further 11–13% of tests end without a decision. The source is the 2011 meta-analysis of the American Polygraph Association.

Is a lie detector ever 100% accurate?

No. No independent study has shown either 100% or 99%. Such figures appear only in studies by the developers of the techniques, and the APA called them statistical outliers.

How often is the polygraph wrong?

In roughly one case in ten. Out of 100 truthful people, about 12 are wrongly judged deceptive; out of 100 deceptive people, about 8 are wrongly judged truthful. The rest are correct decisions or “inconclusive”.

What is an inconclusive polygraph result?

It is when the person’s reactions give no grounds to say either “telling the truth” or “lying”. This happens in roughly every eighth or ninth test. An honest examiner writes exactly that in the report.

How I calculated

The “out of 100 people” table is my conversion of the rows of Table 7 of the report (p. 246) for techniques used to test a single event: sensitivity 0.833, specificity 0.765, false negatives 0.078, false positives 0.115, inconclusives 0.093 for deceptive and 0.127 for truthful examinees. I multiplied these proportions by 100 and rounded to whole numbers, which is why the truthful column adds up to 101. The outliers (IZCT and MQTZCT) are not included in the calculation.

Source

American Polygraph Association. Meta-Analytic Survey of Criterion Accuracy of Validated Polygraph Techniques. Polygraph, 2011, 40(4), pp. 194–305 — full text on the APA website (PDF).

Author: Andriy Maryniak, polygraph examiner, Lviv. In practice since 2012, more than 6,000 tests. This article is a review of a primary source, not legal advice.