Researchers Say Facial Recognition Accuracy Works in Labs but Fails on Streets

Facial recognition technology has long been promoted as a near-perfect tool, with accuracy rates in lab tests often exceeding 99 percent. But new research suggests those numbers are misleading. According to academics from the University of Oxford and the University of Pennsylvania, real-world use of facial recognition technology (FRT) is far less reliable and, in some cases, dangerously flawed.

Why Accuracy in the Lab Doesn’t Translate Outside

The US National Institute of Standards and Technology (NIST) runs the Facial Recognition Technology Evaluation (FRTE), a benchmark used worldwide to justify the adoption of facial recognition in law enforcement and security. Police forces in the UK and agencies across the US have leaned on these high-scoring evaluations to defend their growing reliance on FRT.

But Oxford researchers Teo Canmetin, Juliette Zaccour, and Luc Rocher argue that these benchmarks don’t reflect reality. They highlighted three critical problems:

1. Controlled conditions – Real-world images are often blurry, poorly lit, or partially obscured, unlike the clean datasets used in testing.

2. Small sample sizes – Limited datasets mean there’s a greater risk of errors, especially in diverse populations.

3. Demographic bias – The benchmark datasets fail to capture the full spectrum of real-world faces, leading to disproportionate errors against certain groups.

These shortcomings aren’t just theoretical. In 2020, a Detroit man was wrongfully arrested after being misidentified by facial recognition. A University of Essex study found London’s live facial recognition system could only correctly identify 8 out of 42 people it scanned, an alarming miss rate for technology deployed in policing.

Evidence from Recent Studies

Backing up the Oxford team, a May 2025 study from the University of Pennsylvania found that image quality plays a huge role in accuracy. Their research showed that blur, poor resolution, and pose variation significantly degraded performance and that these errors were not evenly distributed. Marginalized racial and gender groups were more likely to be misidentified.

False positive and false negative rates increase with image degradation, disproportionately affecting individuals from marginalized race and gender groups.

-UPenn researchers wrote.

At the same time, they acknowledged that facial recognition still performs better than some traditional forensic tools such as fingerprinting or firearm comparison.

Growing Concerns About Law Enforcement Use

Critics say accuracy is only part of the problem. A 2023 US Government Accountability Office report revealed that many US agencies had been using FRT without proper training or civil rights safeguards. Advocacy groups argue that lack of oversight has turned the technology into a liability for both law enforcement and citizens. In a number of countries, the Transportation Security Administration are using FRT on travelers without proper consent. Passengers often aren’t told they can opt out.

Facial Recognition Accuracy: What This Means for the Future

The debate over facial recognition is reaching a boiling point. On one hand, law enforcement and private companies see FRT as an indispensable security tool, particularly with global investments in AI-driven surveillance. On the other hand, human rights advocates warn that flawed technology, paired with weak oversight, risks amplifying racial injustice and eroding civil liberties.

As Oxford researcher Juliette Zaccour noted,

Numbers like 99.95 percent sound impressive, but they tell us little about how this technology works in practice. Out of the lab, the stakes are higher, and the risks are real.

Until benchmarks catch up with real-world conditions, experts say, the glowing lab results behind facial recognition may remain more marketing than reality.

ALSO READ: ShinyHunters Hackers Steal Customer Info From Database – Google Confirms