Classifier

A classifier is a program that assigns something to one of several predefined categories. You give it a text, an image, or a series of measurements, and it responds with a label. For an email, the answer is “spam” or “not spam.” For a photo, it might be “dog,” “cat,” or “horse.” The special thing is: nobody writes the rules for this decision by hand. The program is given thousands of already correctly sorted examples and derives on its own what it needs to pay attention to. The term for this is simply “classifier,” and in practice almost everyone uses this English term as well.

Workhorse status in AI practice

Classifiers are among the oldest and most widely used AI building blocks of all. They are inconspicuous but omnipresent. Banks use them to flag suspicious card payments. Hospitals use them to pre-sort X-ray images. Large platforms use them to detect posts that violate their rules.

They are economically interesting because they are much cheaper than a large language model. A language model is an AI system that generates text, and it requires enormous computing power to do so. A specialized classifier for a single yes-or-no question, by contrast, can be tiny. It sometimes runs directly on a phone, without any connection to a data center.

But the downside is important too. A classifier is only as good as the examples it was fed. If those examples were biased, it adopts that bias. This is exactly why classifiers come up so often in debates about algorithmic discrimination.

From examples to a decision boundary

You can picture the task as drawing a line on a sheet of paper. Every email becomes a point whose position is determined by measurable properties: length, frequency of certain words, number of links. Spam points tend to land in one corner, normal emails in another. Training consists of placing the dividing line so that as few points as possible end up on the wrong side.

In reality, this sheet doesn’t have two dimensions but hundreds or thousands. The dividing line is then no longer a line but a complicated surface. This is exactly what methods like neural networks are needed for — computational models made up of many simple, interconnected computing units.

One detail is often misunderstood. A classifier usually doesn’t output a hard decision but a probability. It might say, for instance: 87 percent spam. Where the line for a verdict is drawn is decided by the operator. If it’s set low, more spam is caught, but genuine emails are lost too. This trade-off is called the threshold, and it’s a business decision, not a mathematical one.

Spotting classifiers in everyday life

The spam folder in an email inbox is the best-known example. So are the suggestions for which person appears in a photo, or the automatic language detection when translating. The browser’s warning about a fraudulent website also comes from a classifier.

In news about AI, the term is mostly encountered in connection with safety and moderation. Chatbots place additional classifiers both before and after themselves, which check whether a request or a response is problematic. This checking layer is a separate, small model and not part of the language model itself.

The classifier should be distinguished from regression. A regression estimates a number, such as the price of an apartment. A classifier picks from a fixed list of labels. Mixing the two up means choosing the wrong method — a common beginner’s mistake in machine learning, that is, in computers learning from data.