Accuracy and Bias

Independent testing demonstrates that top-performing biometric algorithms achieve exceptional accuracy with minimal demographic variation, and IBIA supports continued testing, transparency, and improvement to maintain public trust.

Background

One of the most prominent debates surrounding biometric technologies—particularly facial recognition—concerns the accuracy of the underlying algorithms and whether they perform equitably across different demographic groups.

Critics have argued that facial recognition is inherently “biased” against certain populations, particularly people of color, and that this justifies bans or moratoria on the technology. However, independent testing by the National Institute of Standards and Technology (NIST)—the globally recognized authority on biometric algorithm evaluation—tells a more nuanced and encouraging story.

While NIST testing has shown wide variation in algorithm performance overall, the best-performing algorithms achieve accuracy rates of 99 percent or higher with virtually undetectable demographic differences across racial groups and sexes. These top-performing algorithms are more than 20 times more accurate than even skilled professional human examiners.

Importantly, NIST itself uses the term “demographic differences”—not “bias”—to describe performance variations, confirming that such variation is technical and scientific in nature.

IBIA Position

IBIA is committed to promoting the development, deployment, and use of biometric technologies that are accurate, fair, and equitable across all demographic groups.

IBIA’s key positions on this issue include:

Government Agencies Should Deploy Top-Performing Algorithms that Demonstrate Minimal Demographic Differences

NIST testing of hundreds of algorithms from laboratories and vendors worldwide has found that the most accurate identification algorithms display virtually undetectable differences among demographic groups, with false non-match rates of less than three per thousand. The most accurate verification algorithms similarly exhibit both low false positives and low false negatives across demographics. These are the algorithms that government agencies and responsible commercial users should be procuring and deploying.

Performance Variation Is Not "Bias"

Differences in algorithm performance most likely result from natural variations among people in facial bone structures, skin tones, and image capture conditions—not from prejudice introduced by developers. When such variations are identified through testing, researchers fine-tune algorithms to address them, as was successfully done when fingerprint matching challenges were discovered with certain demographic groups. The use of the term "bias" to describe technical performance differentials is semantically misleading and does not reflect the scientific reality.

Automated Recognition Outperforms Human Recognition

The evidence is clear that automated facial recognition augmenting human decision-makers is significantly more accurate than human recognition alone. Measured accuracy of human visual passport inspection, for example, has been found to be 80 percent or less. Human recognition also suffers from well-documented "own-race bias," whereas top-performing algorithms do not. Banning or restricting automated biometric tools would leave society reliant on these far less accurate and more biased human alternatives.

Independent Third-Party Testing Is Essential

IBIA strongly supports independent, third-party evaluations such as NIST's Face Recognition Vendor Test (FRVT) and DHS Science and Technology Directorate Biometric Technology Rallies as the foundation for informed procurement and deployment decisions. These evaluations measure both overall accuracy and accuracy across demographic groups under realistic operational conditions, and they drive continuous improvement across the industry.

Deploy Only Sufficiently Accurate Technologies for the Use Case

IBIA advocates that developers and users procure and deploy only biometric technologies that are sufficiently accurate for the given application. In law enforcement and national security settings, only the most accurate technologies—as validated by established third-party evaluations—should be used. Any biases in an AI-based biometric system should be calibrated out, or at least characterized and stated, so that any derived evidence can be fairly weighed.

Continued Research and Improvement Must Be Supported

The state of facial recognition and other biometric technologies continues to improve rapidly, but further advances require continued use, refinement, and access to diverse training data representative of operational environments. Bans and moratoria would halt this progress and cede U.S. leadership in the development of critical identification technologies to other nations.