Biometrics: Hong Kong Counts 2,800 Arrests, Not Its Errors
Quick answer
How accurate is Hong Kong's facial recognition system?
Nobody outside the police can say, because Hong Kong has not published error figures for SmartView. Police cite arrests and solved cases, but not wrong matches. Research elsewhere shows such systems can misfire often. In Wales, false matches outnumbered correct ones by more than ten to one, so accuracy claims need published numbers.
Biometrics: Hong Kong Counts 2,800 Arrests, Not Its Errors
This episode is based on our article:
Read the full article →Biometrics: Hong Kong Counts 2,800 Arrests, Not Its Errors
Full Episode Transcript
Hong Kong police say their facial recognition network has helped lead to more than two thousand eight hundred arrests. They can't say how many times it pointed at the wrong person. That number isn't published. It may not even be counted.
If you've ever walked past a camera in a train
If you've ever walked past a camera in a train station or a shopping mall, this story is about you. Most of us assume someone, somewhere, checks whether those systems get it right. In Hong Kong, that assumption is being tested right now.
The program is called SmartView. It launched in twenty twenty-four. Police plan to have sixty thousand cameras across the city by twenty twenty-eight. Police say it has helped solve nearly twenty-two hundred criminal cases. Now they're adding real-time alerts, links to shopping mall cameras, and a hundred and eighty days of stored footage. So the question threading through this episode is simple. When the system gets someone wrong, who finds out?
Start with what's been promised. Hong Kong says an officer ranked inspector or above must sign off on the system's use. That's a safeguard, and it's real. But sign-off on using the tool isn't the same as checking a specific match. Nothing in the reporting describes a human double-checking a match before police act on it. And nothing describes a way for the person flagged to push back.
Why does that gap matter
Why does that gap matter? Look at Wales. According to a study of real-time facial recognition used by police there, false matches outnumbered correct ones by more than ten to one. For every person correctly identified, over ten people were wrongly flagged. Imagine being one of those ten, just walking to work.
Now add a second problem. The U.S. National Institute of Standards and Technology, known as N.I.S.T., tested these algorithms in twenty nineteen. It found that error rates for Asian and African American faces were often ten to a hundred times higher than for white faces. That depended on the algorithm. But the pattern showed up again and again. In a city that's overwhelmingly Asian, that research is not a footnote. It's the main question.
That's the professional lesson. An investigator's reputation depends on knowing when the tool is wrong. Without published accuracy numbers, nobody can know that. For everyone else, it means a computer's guess could be the reason someone stops you, and you'd never see the guess.
Then there's the stored footage
Then there's the stored footage. A hundred and eighty days is about six months. That's far longer than most investigations need to look back. Critics call it archive-building. Footage of ordinary people, kept just in case.
Police do have a response. They say Selfie Id Verification 34 Of Rejections Arent The Face will run on specific searches, not constantly on every camera, because of computing limits. Hong Kong's privacy commissioner says organizations must follow the city's privacy laws. If the tool stays limited to missing persons and named suspects, the risk looks very different.
But narrow pilots rarely stay narrow. A tool that starts as one lead among many can quietly become the reason police act. And once a match carries that weight, the missing error count stops being a gap in the paperwork. It becomes the whole problem.
The Bottom Line
Hong Kong is connecting tens of thousands of cameras to software that guesses at faces, and it hasn't said how often the software guesses wrong. Research from Wales and from N.I.S.T. shows those wrong guesses are common, and they don't fall evenly on everyone. Whether you investigate cases or just walk through a mall, the question isn't whether the camera works. It's whether anyone is counting when it doesn't.
Full breakdown's in the show notes.
Ready for forensic-grade facial comparison?
Full forensic reports with detailed similarity scoring. Results in seconds.
Run My First SearchMore Episodes
Selfie ID Verification: 34% of Rejections Aren't the Face
More than one in three selfie verification rejections have nothing to do with your face. According to Shufti Pro's twenty twenty-five failure statistics, thirty-four percent come from a blurry or incomplete picture of you
PodcastAge Verification: Nairobi Court Rejects Typed Birthdays
Typing a fake birthday into a box is all it took for a child in Nairobi to reach online alcohol listings. A judge looked at that and said it isn't good enough. Last month, Kenya's High Court suspended o
PodcastAge Verification: AUD 99M Fines Can't Fix a Face Scan Guess
Researchers found that about two-thirds of kids who got onto social media platforms they were banned from did it with a simple trick. They typed in an older birthday. Or they uploaded a selfie that the platform accepted a
