The conversation about facial recognition bias has been shaped largely by studies from 2018 and 2019 that revealed significant accuracy disparities across demographic groups. Those findings were important and drove meaningful change across the industry. But the technology has evolved substantially since then, and the accuracy picture in 2026 is meaningfully different from what those early studies described.

The National Institute of Standards and Technology's Face Recognition Vendor Test remains the most authoritative benchmark. The most recent evaluations show that the top-performing algorithms have achieved near-parity in accuracy across demographic groups for cooperative, well-lit verification scenarios — the type used in identity verification and border control. The accuracy gaps that were measured in percentage points a few years ago are now measured in fractions of a percent for the leading systems.

This improvement is not accidental. It is the result of specific, deliberate investments. Training datasets have been diversified to include representative samples across ethnicities, ages, and genders. Algorithm architectures have been redesigned to reduce the features that correlated with demographic-specific error patterns. And evaluation methodologies have been refined to measure not just overall accuracy but accuracy equity across groups — making bias a first-class metric rather than an afterthought.

The nuance that often gets lost in the conversation is the distinction between different use cases. Cooperative verification — where a person willingly presents their face to a camera for identity confirmation — performs very differently from surveillance identification, where a system attempts to identify individuals from crowd footage captured at a distance, in varying lighting, and often at unfavourable angles. The bias concerns that apply to surveillance applications do not apply in the same way to cooperative verification, and conflating the two undermines both the legitimate concerns about surveillance and the legitimate uses of biometric verification.

In the identity verification context specifically, the accuracy improvements have practical consequences. Lower error rates mean fewer legitimate users are incorrectly rejected — a friction and fairness issue that disproportionately affected darker-skinned individuals in earlier systems. They mean fewer fraudulent attempts are incorrectly accepted, improving security across the board. And they mean that biometric verification can be deployed more confidently in diverse populations without the risk of systematic exclusion.

The remaining challenges are real but specific. Performance in extreme lighting conditions, with significant facial occlusion, or with low-quality camera hardware still shows greater variability. Age-related changes to facial features create matching challenges over long time periods. And the presentation attack landscape — deepfakes, 3D masks, and other spoofing techniques — requires ongoing investment in liveness detection and deepfake detection to maintain security.

The responsible path forward is to continue improving accuracy and equity through better training data, better evaluation standards, and transparent reporting of performance metrics across demographic groups. It is also to deploy facial recognition in appropriate contexts — cooperative verification with user consent — rather than treating all applications as equivalent.

For organisations implementing biometric matching, choosing a provider that publishes performance data across demographic groups and participates in independent benchmarks like NIST FRVT is the most reliable way to ensure accuracy and fairness. deepidv is committed to equitable biometric performance and ongoing evaluation across all demographic groups.