Sanyanggae LogoSanyanggae

AI6 min read

How does AI pick out advertising messages?

One feature labels a message as likely advertising; another writes a polite reply draft to the same message. Both are called AI, but are they doing the same job?

An illustration placing message cards being divided into two sorting trays beside a separate reply card and pencil
An AI-generated explanatory illustration comparing message classification and reply generation. It is not an actual service screen, neural-network design or accuracy experiment result.

Is looking for the word “free” enough?

Imagine a fictional inbox. A person could write a rule saying: classify a message as advertising if it contains the word “free.” Simple enough, but it might wrongly filter a personal message saying “I'll help you for free,” while missing an advertisement that never uses that word. When people choose the conditions directly, new expressions and contexts may require them to revise the rules.

AI is a broader field than the name of one such feature. It also covers finding objects in images, planning routes, and solving problems using knowledge and rules. Not every AI learns from data, and the mere presence of a conditional statement does not make every program AI. What matters is the task and how it is done, rather than the name.

When examples change a classifier

Another approach is to train a model with example messages labeled as advertising or not. Machine learning trains models on data to produce predictions or other outputs. People still design which material to use and what objective to pursue. Having examples does not guarantee getting every unfamiliar expression right. This inbox is a fictional scene for comparing principles, not an experiment measuring a real service's accuracy.

Training adjusts a model's internal values to fit the data and objective. Feeding in a new message afterward to obtain a classification is inference: using the trained model. A new input does not mean those internal values are learned anew every time. Using conversation content as context must also be distinguished from adjusting the model's values again. The time spent building from examples and the time spent using the resulting model are connected, but they are different stages.

  1. Gather examples

    Decide what should be learned and from which material.

  2. Build the model

    Adjust its internal values during training.

  3. Try it on a new message

    Obtain a classification through inference and compare it with the actual message.

A conceptual diagram distinguishing building a learning-based classifier from using it. Not every AI system works this way.

Picking out advertisements and writing replies

A model that decides whether a message is advertising and one that creates a new reply have different output purposes. Generative AI refers to systems that create content such as sentences, images and sound; modern generative AI usually uses models trained through machine learning. Machine learning is one approach within AI. A model that only classifies messages and does not write text can still be machine learning.

These distinctions are not mutually exclusive boxes. One service can sort messages with a classifier, draft replies with a generative model, and apply human-defined rules to the results. A conversational interface does not prove that only one model is inside. Even when classification, generation and rules appear on one screen, we can examine their jobs separately.

Sounding good and being right

A fluent, confident reply or explanation can still be factually wrong. Just as a classifier can make the wrong choice, a generative model can produce plausible errors. Check whether the result is correct using material and evaluation suited to its purpose. Verify calculations yourself, and check whether cited titles and links really exist and support the claim. Fluency itself is not evidence of accuracy.

Check reply drafts for invented facts and unnecessary personal information too. Before entering personal information or nonpublic material, review the service's handling policies. Combining several approaches does not itself guarantee accurate results or safe handling of data. Even with a convenient feature, looking at what went in, what came out and how to check it makes AI easier to understand concretely.

Following just one message reveals different jobs within the broad name AI. Next time you see an “AI feature” button, first ask whether it selects something or creates something.