Storing a template, not the raw biometric
A biometric template is a mathematical representation extracted from a raw scan — a fingerprint, a face — stored for later comparison rather than the raw image or scan itself. A well-designed template is derived so it isn't reversible back into the original raw biometric, which matters precisely because a biometric characteristic can't be changed the way a compromised password can: there's no way to issue someone a new fingerprint if the raw data itself is what leaked.
This is why insisting on a genuinely irreversible template, rather than storing the raw biometric data directly, is worth treating as a hard requirement rather than an implementation detail — a leaked template is a real, serious problem; a leaked raw fingerprint is a permanent one, for exactly the population whose data was exposed.
Matching alone isn't the whole authentication
Liveness detection verifies that a biometric sample actually comes from a live person present at the sensor, rather than a photo, a recording, or an increasingly convincing synthetic reproduction. A matching algorithm can be excellent at comparing a presented sample against a stored template and still be fully vulnerable to a convincing fake, if nothing separately confirms the sample came from a genuine, present person rather than a captured reproduction of one — the two checks are answering different questions, and both are necessary.