Supervised learning

Supervised learning needs a teacher, and the teaching happens through labels. You assemble a dataset where every example already carries its correct answer: ten thousand emails each marked spam or not spam, a million login events each marked legitimate or compromised. The algorithm adjusts itself until its predictions match those labels as closely as it can, and the finished model applies what it learned to inputs it has never seen. Most of the machine learning in security tooling works this way.

The labels are not a preliminary step. They are the product. A model can only ever learn the definition of truth encoded in its training data, which means whoever produced those labels has quietly written the policy the model enforces. If your analysts marked ambiguous logins as compromised whenever they were unsure, the model inherits that caution and floods you with false positives. If they marked them legitimate to close tickets faster, it inherits the opposite. Nobody wrote a policy. The labels became one. The same dependency runs through every supervised method built on them, whether that is a decision tree, a linear regression, or a convolutional neural network.

That makes labelling both a quality ceiling and an attack surface. No amount of model sophistication recovers accuracy the labels never contained, so a mediocre labelled dataset caps what any algorithm can achieve on it. It is also the cheapest place for an adversary to interfere. Corrupting a subset of labels is far easier than reaching model weights, and the result looks like a model that simply performs badly rather than one that has been attacked (see data poisoning). The governance question is therefore not only where training data came from, but who decided what it meant.