Facial Recognition Technology: Face Recognition Errors, No Appeal

Picture a guy walking into a store. A video camera scans his face, a screen says "match," and a security guard asks him to leave. He has never been in that store. That is how facial recognition goes wrong, and it is already happening to real people.

Facial recognition errors already touch millions of people, and face checks will reach your venues, jobs, trips and accounts long before anyone gives you a fast way to say "that's not me."

Facial recognition technology is about to be everywhere, and the error problem is already here

Here is my prediction, and I'll say it plainly. Over the next couple of years, you will be asked to show your face in more places. Concert gates. Office doors. Airport security lanes. Account logins. Age checks on websites. Almost none of those places will tell you how to fight a wrong result from their facial recognition technology.

IEEE Spectrum, the engineering magazine, has reported that facial recognition errors affect millions of people around the world. I could not pull up the full article text while researching this piece, so I'm leaning on the underlying research it points toward. The numbers in that research are hard to shrug off.

One estimate says 131 million Americans use facial recognition daily, across mobile phones, borders, shops and workplaces, according to Antivirus Insider. Another figure puts unregulated recognition systems touching over 117 million American adults, many of whom were enrolled in law enforcement databases without knowing it.

Why facial recognition error rates jump outside the lab

That jump is the whole story. In a lab, the image is sharp, front-facing and evenly lit. In real life, you get a hat, a bad angle, a grainy ceiling camera video, a rainy night, and a face that is harder to identify. The Federation of American Scientists analysis walks through how performance changes when the image gets messy.

A facial recognition system that looks nearly perfect on paper can be wrong about nearly one in ten faces when it leaves the test bench. (Would you buy a car that was "tested on a track"?) Nobody walking up to a gate sees the test results. They just see a green light or a red one.

---

How facial recognition systems scale: why a "small" match error rate hits millions

Let's do the math a normal person does at 11pm. A recognition system that is wrong 1 time in 100 sounds great. Run it on a stadium crowd of 50,000 and that is 500 people who might be wrongly flagged. Run it every day at thousands of doors and the count explodes.

Researchers call this the scale penalty. The more faces a recognition system has to search, the more chances it has to confuse two people who look alike. The Springer paper "When facial recognition does not 'recognise'" digs into this, including who may be on the hook when a wrong match hurts someone.

The more individuals that the system must "scan," the higher the percentage of error.

Academic literature on facial recognition accuracy, summarized in our research

The "same recognition system, different results" problem

Errors also do not fall evenly. The top recognition system tested as of March 2025 produced 358 times as many false positives (cases where the system says "match" and it is wrong) for West African women over 65 as for Eastern European men aged 35 to 50. That figure comes from the Federation of American Scientists summary of federal testing.

Read that again. Same software. Same day. Wildly different odds of being wrongly stopped, depending on who you are. Research on watchlist imbalance shows how those gaps turn into real enforcement outcomes. A good overall score can hide a bad score for you.

What the test shows	What the real world shows
Clean mugshots, about 0.1% error	Photos taken in the wild, about 9.3% error
One overall accuracy number	False positive rates that differ by age, sex and region, up to 358 times
A small database to search	A huge database or crowd, where facial recognition errors pile up
A trained technician reviewing the face match	A gate, a security guard or a login screen that just says no

---

What happens to a person when facial recognition gets it wrong?

When facial recognition gets it wrong, the person usually pays first: they get stopped, searched, banned, locked out or questioned. In one case highlighted in the research, a wrongful arrest tied to a face mismatch led to a USD 10 million claim against Macy's and EssilorLuxottica. The error was not an abstract number. It was a human being in handcuffs. Previously in this series: Eu Digital Identity Wallet Optional Hard To Dodge By 2027.

Retail is where this gets very ordinary. In one deployment described in the research, a retailer's facial recognition system produced thousands of false positives. The company had not tested accuracy well before launch and did not consistently track false positives afterward. The FTC banned Rite Aid from facial surveillance for five years over a poorly governed rollout.

Is face detection technology getting better?

Yes, and I'll give the industry its due. In controlled, cooperative settings, error rates fell by a factor of three from 2020 to 2025. Some evaluations show them roughly halving every two years. That is real progress for facial recognition technology.

But it assumes the best algorithm, good cameras and cooperative people. Stores, stadiums and cheap login tools rarely check all three boxes. And the market is racing ahead, with one estimate growing from $7.88 billion in 2025 to $9.26 billion in 2026. Money flows in faster than the repair process does.

---

Who do you call when a face scan or identity verification is wrong?

This is the part that bothers me most. Picture the flow. A scan says you are not who you say you are. The person at the security desk did not build the system, cannot see why it said no, and has no button to override it. You get a shrug and a phone number that rings nowhere.

If you have ever wondered whether an image or a profile is really who it claims to be, that is the exact question facial recognition technology exists to answer. The catch is that identity verification needs a human check behind it, with a record someone can look at. A bare "computer says no" is not an answer.

So here is the one thing I'd do. Whenever a face check blocks you, ask three plain questions on the spot: What system made this call? Who can review it by hand? Can I get this in writing? Then take a photo of the screen and note the time and place. It feels like overkill until you need it, and the people who get fixed fastest are the ones who brought proof.

The availability trick is worth knowing, too. Stories of wrongful arrests stick in your mind, which makes the risk feel huge or like it only happens to others. The truth is more boring and more worrying. Most errors are quiet: a delayed boarding, a locked account, a hiring portal that will not accept you. Nobody writes headlines about those.

Here is the question I'd leave you with. When a machine is 99% right, who is responsible for the 1% it gets wrong about you? Today, the honest answer is nobody, and facial recognition keeps growing anyway.

Facial recognition: Frequently Asked Questions