A facial-recognition result is a comparison made under particular conditions. To judge a claim about its accuracy, establish what the system is trying to do, whose photographs it compares, where its decision threshold sits, and what happens after a possible match.
This guide is for readers examining a proposed deployment, a vendor demonstration or a public statement. It explains the measurements to request and the questions that turn an impressive percentage into evidence you can assess.
Start with the task: verification or identification
Verification, or one-to-one comparison, checks whether two images belong to the same person: for example, a fresh photograph against an enrolled identity. Identification, or one-to-many search, compares a photograph with a gallery and returns possible candidates. A watch-list search also needs to handle people who have no matching identity in that gallery.
NIST evaluates these tasks separately in its FRTE verification results and FRTE identification results. Use the evaluation matching the actual proposal. A result from paired passport-style photographs does not establish how the complete system will perform on distant, moving faces in a crowded entrance.
On a narrow screen, focus the table region and use the arrow keys or swipe to read all columns.
| Measurement | What it counts | Question to ask |
|---|---|---|
| False match rate, 1:1 | Different-person comparisons accepted as matching. | At which threshold, on which image population? |
| False non-match rate, 1:1 | Same-person comparisons rejected as different. | What is the error rate at that same threshold? |
| False positive identification rate, 1:N | Searches with no enrolled mate that nevertheless return one or more candidates above threshold. | How large is the gallery, and how many searches will occur? |
| False negative identification rate, 1:N | Searches with an enrolled mate that fail to return that mate above threshold. | Which gallery, image conditions and acceptance rule were tested? |
NIST's identification page defines its rates and threshold settings alongside the results. Record those definitions when citing a score; later tables may use a different dataset or operating point. For a gallery search, use a measured search-level error rate. Multiplying a pairwise rate by gallery size can be misleading when the search implementation and comparisons differ.
A threshold is a choice with two consequences
The comparison produces a score. A threshold determines which scores count as accepted matches or reportable candidates. Requiring stronger similarity generally reduces false matches while increasing missed matches. A useful evaluation shows both errors at the chosen threshold, along with the cost of each kind of mistake in the proposed workflow.
Write down the algorithm name and version, threshold, gallery size, capture method, test date and sample size. Ask how poor-quality images, people outside the database and repeated attempts were handled. A percentage calculated only after difficult images were discarded describes that filtered set.
Image quality and population also matter. NIST's 2019 demographic study found substantial differences across algorithms and demographic groups. That dated finding is a reason to examine the current system's subgroup results and capture conditions. It is not a current score for every product bearing the same technology label.
Worked example: a small error rate can create a large review queue
The practical lesson is to estimate the work created by the whole flow. In this fictional setting, a reviewer faces roughly eleven false alerts for each correct one. The number of genuine targets, the search volume and the false-alert rate all affect that burden. A candidate photograph is a lead requiring an appropriate verification process, not a conclusion about identity or conduct.
Ask what happens to people and their data
In October 2001, the ACLU challenged a proposed facial-recognition installation at Rhode Island's T. F. Green Airport. Its letter questioned the watch-list database, errors, retention and expansion into other uses. Read it as a dated advocacy document. Its operational questions remain useful to investigate with current evidence; its period accuracy figures should stay attached to their period.
- Purpose and authority: Who approves this specific use, under which applicable rules, and how are changes in purpose reviewed?
- Gallery: Who qualifies for inclusion? Who corrects outdated, duplicate or wrongly attributed records?
- Retention and access: What happens to original images, derived templates and unsuccessful searches? Request the retention schedule and access log policy.
- Human decisions: Who reviews an alert, what independent evidence is required, and how can an affected person challenge a mistake?
- Local evidence: What test represents the intended cameras, lighting and population? What outcome would trigger suspension or redesign?
These are evaluation questions; the applicable legal requirements depend on the place and use. Record unanswered questions explicitly rather than filling them with assumptions. Recheck the evidence when the algorithm, cameras, gallery, threshold or operating purpose changes.
Download the facial-recognition review worksheet to record claims, definitions, evidence and follow-up responsibilities without collecting personal biometric data.