Every time your phone unlocks with a glance, or an airport camera waves you through a gate, you are watching one of the quiet triumphs of modern AI. That same technology also sits inside police databases, retail security systems, and companies that have quietly copied billions of photos from the open web. Understanding how facial recognition works and its risks matters because the two are inseparable: the machinery that makes a face convenient to match is the very same machinery that makes it dangerous to get wrong. The story begins with a 2012 breakthrough called AlexNet, which taught computers how to see — and it is really about the long shadow that breakthrough cast.
To be clear up front: this is not a claim about that 2012 paper itself. Face recognition is a downstream consequence of the deep-learning era AlexNet kicked off. The convolutional neural network — the kind of model that won in 2012 — turned out to be very good at faces, and once it was cheap and accurate enough, it spread from research labs into places where a mistake has real consequences.
🧩 How a Face Becomes a Number
A facial-recognition system does not “recognize” a face the way a friend does. It works in two mechanical steps.
First, it detects and normalizes: the software finds the region of an image that contains a face, then straightens and crops it to a standard size and pose, so a tilted head or an off-center photo doesn’t throw it off.
Second — and this is the heart of it — a convolutional neural network converts that cleaned-up face into a list of numbers called an embedding. You can think of the embedding as a numeric fingerprint: a few hundred numbers that together describe the geometry and texture of a particular face. The network is trained so that two photos of the same person produce embeddings that sit close together, while photos of two different people produce embeddings that sit far apart. The machine never “knows” your name or your story. It produces a point in a high-dimensional space, and everything after that is arithmetic.
That reframing is worth sitting with. A face, to the system, is not a face. It is coordinates. And once you are a set of coordinates, comparing you to someone else is just a matter of measuring distance.
🔍 Finding the Match: One-to-One vs One-to-Many
There are two very different jobs people call “facial recognition,” and the difference matters enormously for the risks.
One-to-one verification asks: are these two faces the same person? Your phone stores the embedding of your enrolled face; when you look at it, it computes a fresh embedding and checks whether the distance between the two is small enough to clear a threshold. Passport e-gates work the same way — comparing your live face to the chip in your document. Because it’s comparing exactly two faces you have chosen to enroll, the stakes are contained.
One-to-many identification asks a much harder question: who is this, out of everyone in a database? Here the system takes one unknown face, turns it into an embedding, and compares it against millions of stored embeddings — mugshots, driver’s-license photos, or scraped web images — returning the closest matches as a ranked list of candidates. This is the mode police use, and it is where the trouble concentrates. The output is not a certainty. It is a list of guesses, sorted by similarity, and the top guess can be wrong.
⚖️ Where the Accuracy Breaks Down
For years, vendors sold facial recognition as if it were essentially solved. Careful, independent testing told a more complicated story — and it’s important to keep two landmark studies straight, because they measured different things.
The first is Gender Shades (2018), by Joy Buolamwini and Timnit Gebru. They audited commercial products from IBM, Microsoft, and Face++ — but on a classification task: guessing a person’s gender from their face, not identifying who they are. The gaps were stark. Error rates for darker-skinned women ran as high as 34.7%, while the maximum error for lighter-skinned men was 0.8%. That is a system that worked almost flawlessly for some people and failed roughly a third of the time for others.
The second is a government study of face recognition specifically: the U.S. National Institute of Standards and Technology (NIST) report NISTIR 8280 (December 2019), which tested 189 algorithms from 99 developers. It found that false-positive rates — wrongly declaring two different people a match — were often 10 to 100 times higher for Asian and African American faces than for white faces in one-to-one matching, depending on the algorithm. In one-to-many search, African American women saw elevated false positives — precisely the scenario, NIST noted, where a false match “could include false accusations.”
Here is the crucial nuance, and it cuts against fatalism: NIST found the disparity was not universal. Some algorithms showed no such gap. That points to training data — unrepresentative datasets — as a major driver, which means this kind of “bias” is a measurable, technical failure with fixable causes, not an inevitable law of the technology. “Bias” here means documented error disparities, not a rhetorical flourish.
🚔 Why a False Match Is Worst in Policing
Put those two facts together — one-to-many search returns a ranked list of guesses, and the error rates are highest for exactly the groups most likely to be over-policed — and you can see why the courtroom, not the smartphone, is where this technology has done real damage.
The failure in every documented case is depressingly consistent, and it is human as much as technical: a face-recognition “lead” was treated as confirmation instead of a starting point. The software is designed to hand back a candidate to investigate. When that candidate is instead treated as the answer, the ordinary police work that would catch an obvious mismatch gets skipped.
As reported by the American Civil Liberties Union (ACLU) and news outlets, Robert Williams was arrested by Detroit police in January 2020 after a facial-recognition search wrongly matched his driver’s-license photo to store surveillance footage of a shoplifter — the first publicly reported wrongful arrest tied to the technology. Detroit settled his ACLU lawsuit in June 2024 for $300,000 and agreed to significant policy limits, including no longer arresting people on the basis of a facial-recognition result alone. In a separate case, Porcha Woodruff — eight months pregnant at the time — was arrested in Detroit in 2023 on robbery and carjacking charges after a false match; the charges were later dismissed. The ACLU has documented more than a dozen wrongful arrests nationwide linked to the technology.
These are real people, and the point is not that a computer is malicious. It is that “the computer matched him” is not evidence. A match is a lead that a human still has to verify — and the reform Detroit agreed to, that a face match can never be the sole basis for an arrest, is exactly the guardrail whose absence caused the harm.
🕸️ Your Face, Copied From the Open Web
The accuracy debate assumes a system was fed a database someone was allowed to build. One company blew past that assumption. According to court filings and reporting, Clearview AI assembled a database of more than 60 billion facial images scraped from social media, news sites, and other public corners of the internet, then sold search access to law enforcement and private clients. If you have ever been tagged in a public photo, you may already be in a lineup you never consented to.
Clearview was sued under Illinois’ Biometric Information Privacy Act (BIPA), one of the strictest such laws in the United States. The case produced an unusual settlement, reached in 2024 and approved in March 2025, that gave the plaintiff class a roughly 23% equity stake in the company — valued at about $51.75 million — rather than a conventional cash payout.
The Clearview story is really an argument about why faces deserve special caution. A face is biometric data: uniquely yours, and — unlike a password — impossible to change if it leaks. You can reset a stolen password in a minute. You cannot reset your face. That permanence is the whole reason lawmakers have started singling out biometrics for special rules.
🏛️ What Regulators Are Doing
The legal system is, slowly, catching up. In May 2019, San Francisco became the first U.S. city to ban the use of face recognition by police and city agencies, on an 8–1 council vote; Oakland, Berkeley, and Somerville, Massachusetts, soon followed with their own restrictions.
Across the Atlantic, the EU AI Act takes a broader swing. It prohibits the untargeted scraping of the internet or CCTV footage to build facial-recognition databases — a rule aimed squarely at the Clearview model — and sharply restricts real-time remote biometric identification in public spaces by law enforcement, with those prohibitions taking effect in February 2025. Many other biometric uses are classified as “high-risk,” carrying heavy compliance obligations rather than an outright ban.
None of this makes facial recognition go away, and much of it remains contested and uneven from place to place. But the direction is clear: after a decade of quiet deployment, the technology is finally being treated as something that needs rules. The same convolutional networks that let a machine see also let it watch — and a companion piece in this Secure pillar, on adversarial examples, covers the other shadow the vision revolution cast: how the very models that recognize a face can be fooled by a change too small for a human to see.
❓ Frequently Asked Questions
How does facial recognition actually identify someone? A convolutional neural network converts a detected face into an embedding — a list of a few hundred numbers that acts as a numeric fingerprint. To identify you, the system compares your embedding against stored ones and returns the closest matches by distance. Same person, close numbers; different people, distant numbers. The machine measures similarity; it doesn’t “know” you.
Is facial recognition accurate? It depends heavily on the algorithm and the task. The best systems NIST tested are highly accurate, but NIST’s 2019 study also found that many algorithms produced false positives 10 to 100 times more often for Asian and African American faces than for white faces. Accuracy is real, but it is uneven — and it’s worst exactly where the consequences are most serious.
Has facial recognition led to wrongful arrests? Yes. As reported by the ACLU and news outlets, Robert Williams (Detroit, 2020) and Porcha Woodruff (Detroit, 2023) were both arrested after false matches; Williams’s case settled for $300,000 and Woodruff’s charges were dismissed. The ACLU has documented more than a dozen such cases. In each, the underlying failure was treating an unverified match as proof.
What’s the difference between the Gender Shades and NIST findings? They measured different tasks. Gender Shades (Buolamwini and Gebru, 2018) audited gender classification — guessing gender from a face — and found up to 34.7% error for darker-skinned women versus 0.8% for lighter-skinned men. NIST’s NISTIR 8280 (2019) tested face recognition — matching identity — and found large demographic gaps in false-positive rates. Both revealed disparities, but they aren’t the same study or the same task.
Can I stop my face from being used? Only partly. Tightening photo-privacy settings and limiting public posts helps, but companies like Clearview scraped images that were already public, and you can’t un-share what’s been copied. The more durable protections are legal — laws like Illinois’ BIPA and the EU AI Act — because a face, unlike a password, can never be reset.
📚 Further Reading
- How AI Learned to See — the AlexNet story and how a convolutional network turns pixels into labels, the breakthrough that made face matching cheap and accurate.
- What Is an Embedding? — the plain-English explanation of how a face (or anything) becomes a point in high-dimensional space, which is exactly what a facial-recognition match measures.
- Can You Fool an AI’s Eyes? Adversarial Examples Explained — the companion Secure piece on the other shadow of machine sight: how the same vision models can be tricked by an invisible change.
- The primary sources: Buolamwini & Gebru, Gender Shades (2018); NIST, Face Recognition Vendor Test Part 3: Demographic Effects (NISTIR 8280, 2019); and the ACLU’s case record in Williams v. City of Detroit.
📚 Recommended Resources
Disclosure: some links below are Amazon affiliate links. If you buy through them, I earn a small commission — at no extra cost to you — which helps keep this site running.
Artificial Intelligence: A Guide for Thinking Humans
Melanie Mitchell · Book · ~$18 on Amazon
The best plain-English book for understanding why facial recognition succeeds and fails the way it does. Mitchell is exceptionally clear on how vision models turn images into statistics rather than understanding — the exact gap that explains both the convenience and the wrongful matches in this article. Start here if you want the ideas without the equations.
Hands-On Machine Learning with Scikit-Learn, Keras & TensorFlow
Aurélien Géron · Book · ~$70 on Amazon
The standard practical bridge from “curious” to “building.” If reading about embeddings and nearest-match search made you want to see how a face actually becomes a vector — and how a distance threshold decides a match — this is the field manual for the convolutional networks behind it, clear and genuinely hands-on.
Genius Makers: The Mavericks Who Brought AI to Google, Facebook, and the World
Cade Metz · Book · ~$18 on Amazon
The human story behind the vision revolution that made face matching cheap and ubiquitous — the researchers, the rush, and how the technology (and its blind spots) escaped the lab and landed in police databases and phones. A gripping narrative that gives the ethics debate its backstory.