A face unlocks a phone, clears a border and, increasingly, confirms an identity to a service that has never met you. What biometrics genuinely fix, the property that makes them different from every other credential, and where the distinction between matching on your device and matching on somebody else's server decides everything.
What they genuinely fix
Two problems at once, and they are the two that decades of password advice never solved.
The memory problem disappears: there is nothing to recall. The reuse problem disappears with it, since there is no string to carry between sites.
And crucially, the secure path became the fast path. Every previous improvement asked people to do something slower in exchange for safety, and people declined. This one asked them to do something quicker. That is why it succeeded where instruction failed, and it is worth understanding as a design lesson rather than only as a risk.
The property that makes them different
A password is a secret you know, and if it leaks you replace it.
A face is a characteristic you display continuously in public and cannot replace.
The consequence is precise: the security of a biometric does not rest on secrecy, because there is none. It rests entirely on the difficulty of presenting a convincing copy. Your face is not confidential and was never going to be; the system is betting on the cost of forging it, and that is a different kind of bet with different failure modes.
On your device, or on their server
This is the distinction that decides the whole question, and almost nothing in the interface reveals which one you are in.
On the device. A template is generated and stays in hardware designed to hold it. The comparison happens locally. The service learns one bit: pass or fail. Nothing about your face leaves.
On a server. A template of your face is transmitted and stored. It can be breached. It can be compared against other collections. It can be retained long after you stop using the service, and what happens to it if the company is sold is a question about a contract rather than about cryptography.
From where the user stands these are the same gesture — look at the camera, get in. Beneath them they are not remotely the same arrangement.
Liveness, and the arms race
Because the input is not secret, the system's real job is telling a person from a representation of a person: a photograph, a replayed video, a mask, a generated face.
That is a contest rather than a solved problem, and it moves. Defences raise the cost of a convincing presentation; generation lowers it. Each improvement in synthesis is an improvement in the attacker's economics, and each countermeasure adds friction that users notice.
No prediction is needed about who wins. The shape is enough: this is a cost race, not a proof, and it will be re-run with every generation of tooling.
Irreversibility
A breached template cannot be reissued. There is no equivalent of choosing a new one, and the exposure does not expire.
And the second half, which is easy to miss: your face is being captured by systems you never enrolled with — cameras in shops, photographs taken by other people, images posted by friends. Enrolment is not what put your face into circulation. It only attached your face to an account.
So the risk is not created by choosing to use face unlock. It is created by the world, and enrolment decides who holds a usable reference copy.
Where it is reasonable, and where it is not
Reasonable: unlocking a device in your hand, matched locally, falling back to something you know. The convenience is real, the exposure is contained, and the alternative is a four-digit code somebody can watch you type.
Questionable: as the only factor protecting something consequential. As a remote check against a server-held template. And anywhere the fallback is weaker than the biometric, which is common — an attacker facing a good face check and a weak recovery flow does not attack the face check.
Close
Two questions, and they settle most of it. Where is the comparison performed, and what is stored after it succeeds?
A service that cannot answer plainly has usually answered by not answering.