GAO reported a 66 percent average polygraph pass rate for FBI applicants from fiscal years 2018 through 2023. That is not an agency-wide FBI failure rate: it covers applicants, not the people already working there, and GAO reported pass rates rather than a direct failure-rate measure.

What does a polygraph failure rate actually measure?

It measures only what its source defined. The numerator might contain technical significant-response results, unresolved charts, countermeasure calls, admissions, or final disqualifications. The denominator might contain examinations, unique applicants, completed examinations, or everyone who entered a hiring pipeline.

Start with the table title and footnotes, not the headline. “Failed tests divided by tests conducted” answers a different question from “applicants removed after the polygraph stage divided by applicants reaching that stage.” Both can be correct. They still answer different questions, and their rates are incompatible.

The National Academies review of polygraph evidence distinguishes event-specific investigations, employee screening, and preemployment screening. Those uses seek different information and carry different implications. A statistic from one setting should not be imported into another without evidence that the method and population match.

A candidate should also ask what “failure” means operationally. Did the agency stop the application? Did an examiner record a technical result? Did a reviewer uphold it? Did an applicant make an admission that led to a separate suitability decision? One word can hide several events.

Before copying a rate, determine whether the original data can identify people across repeat appointments and connect the examination stage to final dispositions. Sometimes the source cannot do either.

Why are examinations, applicants, and completed cases not interchangeable?

One applicant may generate more than one examination. Another may leave before testing. A third may receive an inconclusive result and return. Counting each examination enlarges the denominator differently from counting each person once, while completed-case data may exclude withdrawals and unfinished retests.

A 2012 GAO review of CBP data from January 2008 through August 2012 illustrates the problem. Its table counted 11,149 examinations, including 859 retests. The failed-test category included significant-response and no-opinion outcomes; “no opinion” included cases in which no exam was administered or no viable exam was collected, usually because an applicant did not appear. That table cannot be read as a clean share of unique people who lied.

Even “applicant” needs a boundary. The word alone tells you little. Is the population everyone who applied online, everyone found minimally qualified, everyone scheduled, everyone who appeared, or everyone with a final quality-controlled result? Attrition before the exam can change the composition of each later group.

When a source lacks a person-level identifier, you may be unable to remove repeat tests. When it reports only completed charts, you may be unable to recover no-shows or withdrawals. Mark those as unknown. Do not repair missing data with an assumption.

Is a technical polygraph result the same as a final hiring outcome?

No. A technical result belongs to the examination process; a hiring outcome can incorporate admissions, background information, suitability criteria, qualifications, medical or drug-screen results, and other requirements. The relevant agency must explain how its process connects those stages.

The same 2012 GAO review described the reverse mismatch too: an applicant could technically pass yet make an admission that might affect suitability when the agency reviewed the complete record. The report also found that the agency then lacked a mechanism to identify which screening tool supplied the information behind every unsuitable determination.

A September 2024 GAO review reported a related data limit: CBP could not state the exact share of applicants who made disqualifying admissions because it recorded those admissions only when an examiner used the on-call adjudication process. That is an unknown in the dataset, not permission to treat all technical failures as admissions.

How should inconclusives, retests, withdrawals, and no-shows be handled?

They should appear as separate categories unless the source fully explains another treatment. Combining them with failures changes the numerator. Excluding them changes the denominator. Carrying only the final result hides the path and can erase the burden of repeated testing.

No-opinion language deserves special attention. It may refer to insufficient data, inability to render a result, suspected countermeasures, or an administrative circumstance, depending on the program and period. Never assign one universal meaning without the source’s definition.

The practical candidate lesson is not to predict your own result from a web percentage. Understand the agency’s stated process, prepare to answer accurately, and follow its instructions. If you receive an inconclusive or adverse result, ask the agency what the status means and whether its published process provides a next step.

How do applicant self-selection and job-specific pools affect comparison?

The people who reach testing are selected by earlier steps and by their own choices. Qualification screens, background forms, scheduling delays, agency reputation, job requirements, and withdrawals can all alter the tested pool. A high or low observed rate may partly describe who remained, not just the examination.

Federal law generally requires polygraphs before hire for CBP law-enforcement applicants, subject to waiver provisions. The current U.S. Code provision for that population does not make it equivalent to every intelligence, law-enforcement, contractor, or incumbent-employee population.

Test scope matters as well. A counterintelligence-focused examination and a broader-scope examination do not necessarily ask the same categories of questions. The explanation of counterintelligence versus full-scope polygraphs helps identify that mismatch before you compare outcomes.

Agency pass rates are not a quality ranking. GAO cautioned that differences can reflect applicant demographics, where the examination occurs in the hiring process, and agency-specific protocols. Those factors have to be addressed before rates can support a comparison.

Why do fiscal year, policy change, and data source matter?

A rate is a snapshot of a defined period. Hiring surges, statutory changes, examiner capacity, retest rules, reciprocity, question formats, quality-control practices, and reporting systems can change over time. A historical figure should remain labeled with its period and definitions.

Separate agency administrative data from a scientific accuracy study. Administrative data show how that program classified events under its operating rules. They do not reveal ground truth for every examinee and therefore do not, by themselves, establish sensitivity, specificity, or a false-positive rate.

The National Academies executive summary warns that results from particular research settings or populations are difficult to generalize to other settings and populations. It also explains the tradeoff between catching more targets and generating more false alarms when the target condition is rare.

Source lineage matters. A search snippet is not enough. Prefer the actual agency table, audit report, statute, or study over a copied number without footnotes. Record publication period, data owner, extraction date, revisions, and whether the source labels the table preliminary or complete.

What should a responsible agency comparison show?

It should expose definitions and missing fields before displaying a rate. If the rows cannot be aligned, stop short of ranking. A table that says “not reported” is more accurate than a polished league table built from incompatible categories.

For each agency and period, capture the tested position, exam scope, unit of count, definitions, treatment of incomplete cases, and relationship to suitability or hiring. Link the original source beside the row.

Keep scientific accuracy claims in a separate table from administrative outcome rates. A program’s share of adverse technical outcomes is not the probability that its calls match truth. The discussion of common polygraph problems addresses the individual experience; it should not be used as a denominator for an agency statistic.

Likewise, employment consequences deserve their own analysis. The article on failing a polygraph while keeping a job explains why applicant screening and incumbent employment questions should not be merged.

Frequently Asked Questions

What is the CIA polygraph failure rate?

Treat any CIA failure rate as unsupported unless it comes with a primary source, a period, a population, and an outcome definition. GAO’s published agency comparison covers the FBI, DEA, Secret Service, and CBP, not the CIA.

What is the FBI polygraph failure rate?

GAO reported a 66 percent average pass rate for FBI applicants from fiscal years 2018 through 2023. It also reported 56 percent for DEA applicants, 36 percent for Secret Service applicants, and annual CBP rates ranging from 28 to 40 percent from fiscal year 2018 through the first half of fiscal year 2024. These are applicant pass rates for the agencies GAO compared, not failure rates for an agency’s whole workforce, and GAO says applicant mix, timing, and agency protocols may explain differences.

Does an inconclusive result count as a failure?

Only the source’s rules can answer that for a particular dataset. Some tables separate inconclusive outcomes; others exclude or regroup them, so the table definition controls.

Can pass rates show which agency has the fairest exam?

No. A pass rate alone cannot isolate exam quality or fairness from applicant mix, test scope, policy, selection, withdrawals, retesting, and outcome definitions. Comparable evidence would be needed for that conclusion.

What should I save when I find a public rate?

Save the original table and its footnotes. Record the source date, the unit of count, the population, and the definitions; mark a missing field unknown before making a comparison.