VSA vs. Polygraph Accuracy: Evidence and Limitations

Can VSA and polygraph accuracy be compared with one percentage?

Not responsibly without examining what each study measured. A laboratory deception task, a specific-incident examination, and broad screening are different applications. Comparing headline percentages from different settings can create a misleading impression of a direct contest.

For trained investigative professionals, agencies, and examiners, the useful question is narrower: how did this particular system and examination procedure perform against independently established truth, in circumstances comparable to the intended use?

What the NIJ voice stress field study actually tested

The National Institute of Justice published a 2008 summary of a field evaluation of Computer Voice Stress Analyzer (CVSA) and Layered Voice Analysis (LVA). Researchers questioned more than 300 arrestees about recent drug use and compared their responses with urine drug tests.

NIJ reported approximately 50% overall deception-detection accuracy in that setting. The summary also reported that the programs identified 15% of deceptive respondents and incorrectly classified 8.5% of truthful respondents as deceptive. Those are different measures, not interchangeable accuracy scores.

The findings are substantive negative evidence for the programs and task evaluated. They should neither be dismissed because of their age nor presented as a direct performance test of every voice stress product. AISIS and VASIS were not the systems evaluated in that study. Their absence does not establish that they perform better.

What the polygraph evidence supports

The National Research Council’s 2003 review concluded that specific-incident polygraph tests could distinguish deception above chance but well below perfection. It also identified serious limitations in applying available evidence to personnel security screening.

That distinction matters. A result from one examination technique or study population is not a universal polygraph accuracy rate. Nor does a favorable finding for polygraph validate voice stress analysis, or an unfavorable voice stress finding establish that every polygraph application is reliable.

Read the errors, not just the headline

A useful evaluation makes its denominator and decision rules explicit. Ask how the following were measured and reported:

A high percentage among conclusive examinations does not mean the same percentage of all examinations produced a correct decision. Results also need uncertainty estimates, transparent sampling, and enough detail for independent scrutiny.

• Sensitivity: the proportion of deceptive cases correctly identified as deceptive.
• Specificity: the proportion of truthful cases correctly identified as truthful.
• False positives and false negatives: truthful cases classified as deceptive, and deceptive cases classified as truthful.
• Inconclusive results: how often no decision was reached, and whether those examinations were excluded from the advertised accuracy figure.
• Ground truth and base rates: how truth was established independently, and how common deception was in the tested population.

Stress identification is not proof of deception

A stress response and a lie are not interchangeable. A truthful person may be anxious, and a deceptive person may not show a distinctive response. Detecting a physiological or vocal change does not, by itself, establish why that change occurred.

For either method, investigative usefulness, admissions obtained, examiner agreement, and verified deception-detection accuracy are separate outcomes. Each needs evidence appropriate to the claim. An examination result should inform corroborative investigation, not replace it.

The same evidence standard applies to AISIS and VASIS

VIPRE Technology Group publishes this reference and supplies AISIS, its AI-assisted platform, and VASIS, its examiner-led voice stress analysis system. This page is not an independent validation of either product.

A different architecture, AI assistance, professional experience, or a newer release does not establish a superior deception-detection accuracy rate. Agencies evaluating AISIS, VASIS, or any competing system should request independent, product-specific research covering the relevant version, protocol, population, error rates, and limitations.

• Identify the exact hardware, software version, scoring method, and examiner procedure tested.
• Request independently established ground truth, blinded assessment where feasible, and disclosure of funding and conflicts of interest.
• Separate controlled-study performance from operational claims, and ask whether findings have been independently replicated.

Further reading from Robert Martin

For the fuller discussion of accuracy claims, research quality, and transparent validation, read Robert Martin’s article. This reference provides the concise comparison; the article develops the broader argument.