AI vs. Machine Learning vs. Deep Learning: One Spam Filter, Three Approaches

Three teams are building spam filters. One writes rules. One trains a statistical model on word counts. One trains a neural network on sequences of tokens.

All three want the same outcome: fewer unwanted messages in the inbox. Calling the third approach “more AI” does not tell us which filter is best. The terms describe overlapping families of methods, not a ranking of product quality.

The rule-based filter

The first team starts with explicit conditions: block known malicious domains, flag certain combinations of phrases, allow trusted senders. People supply the decision logic.

Rule-based systems have a long history within artificial intelligence. AI is the broad field concerned with building systems that perform tasks associated with intelligence, such as reasoning, perception, language, or planning. There is no requirement that every such system learn from examples.

Rules can be excellent when the requirement is precise and relatively stable. They are inspectable: someone can point to the condition that triggered a decision. They become expensive to maintain when each new case produces another exception.

There is also a naming limit here. An ordinary conditional statement is not automatically worth calling AI. The label depends on the system and its purpose; attaching it to a feature is not a technical explanation.

The filter trained on examples

The second team converts each email into features, such as word counts, and trains a classifier using labeled messages. A model such as logistic regression learns how those features relate to the target label.

This is machine learning: the decision function is fitted from data rather than completely specified by hand. People still choose the examples, representation, algorithm, and evaluation procedure.

The model can generalize to combinations it has not seen exactly before. It can also learn shortcuts. If all the training spam happens to be in one language, language may become a proxy for spam. The algorithm does not know that this is an unacceptable interpretation unless the development and evaluation process exposes it.

Learning from examples changes where the rules come from. It does not establish that those rules are sensible.

The neural filter

The third team uses a neural network with multiple layers. Rather than relying only on manually chosen word-count features, it can learn intermediate representations from the input during training.

This is deep learning, a family within machine learning. The layers can support richer representations, but the system still needs an objective, training data, and evidence that its predictions work in the intended setting.

The network may capture distinctions the simpler model misses. It may also cost more to train and operate, be harder to inspect, and provide little practical improvement on a small or straightforward task. “Deep” describes the method; it is not a certificate of accuracy.

Compare the failures, not the vocabulary

Give all three filters the same held-out messages. Count legitimate mail incorrectly hidden, spam missed, processing time, and maintenance effort. Check whether the test includes the languages and message types people actually receive.

A layered product might use all three approaches: explicit rules for known attacks, a statistical classifier for ordinary filtering, and a neural model for difficult cases. The categories are compatible because they describe different levels of the system.

One final distinction matters in every version: making a prediction does not necessarily update the system. A trained model can remain fixed while processing millions of emails. Whether feedback leads to retraining is a separate engineering decision.

If someone says a product “uses AI,” ask which part makes a decision, how that decision was constructed, and how its failures are measured. Those answers tell you considerably more than the label.