AI Is More Than Machine Learning

AI is not a synonym for a chatbot

A lot of today's internet curriculum accidentally teaches a chain that looks like this: AI equals a neural network, a neural network equals a large language model, and a large language model equals the product you chat with.

That is a category error.

AI is the category of machine-based systems. Machine learning is one way some of those systems get from inputs to outputs. Search, rules, planning, optimization, and probabilistic reasoning are others. A neural network is one kind of learned model. A large language model (LLM) is one kind of neural network trained on text. A chat product wraps a model in retrieval, tools, filters, and an interface.

Say this out loud:

AI is not machine learning. Machine learning is not a neural network. A neural network is not an LLM. An LLM is not an AI product.

Those five things overlap. They are not the same object.

The category, then the methods

NIST describes AI as a machine-based system that can, for human-defined objectives, make predictions, recommendations, or decisions that influence real or virtual environments. [nist-ai-glossary] The OECD's updated definition is a machine-based system that infers from inputs how to generate outputs such as predictions, content, recommendations, or decisions. [oecd-ai-2024]

"Infers" is broader than "was trained on a dataset." A system can infer by following a rule, searching options, optimizing a score, updating a belief, learning from examples, or mixing those methods.

So when someone says "this uses AI," the useful beginner questions are the same ones as in What is AI?:

What goes in, what comes out, and which method turns one into the other?

The same everyday problem, solved seven ways

Take a problem everyone already understands: getting from home to a coffee shop with a maps app.

1. Lookup / rule

Always take Main Street. An explicit instruction. No examples required. If Main is closed, the rule is stale.

2. Search

Try possible streets until one reaches the shop. Search explores options. It does not, by itself, learn from past trips.

3. Optimization

Pick a route that minimizes time or fuel under constraints such as one-way streets or no left turns. Optimization needs an objective (what "best" means) and constraints. It fails when the objective is wrong, not only when the map is wrong.

4. Probabilistic inference

Update a belief. Main floods when it rains. Choose a route given that uncertainty. This is reasoning with incomplete information, not fitting a predictor from a pile of past trips.

5. Learned prediction

From past trips, predict duration at this hour. That is machine learning. It needs data that actually represent the world you will drive in. If the data are old, biased, or from a different city, the prediction can look precise and still be wrong.

6. Planning

A sequence of actions: leave, park, walk the last block. Planning considers later steps, not one isolated prediction. A good duration forecast is not a plan if it ignores "the lot is full on Saturdays."

7. Hybrid

A real maps app mixes these. Traffic prediction may be learned. The route is searched and optimized. Closures can be rules. Arrival time is uncertain. When the app is late, you cannot assume "the AI" failed as one blob. The learned model, the search, or a stale rule might each be the part that broke.

A chat product is often a hybrid too

The language model generates text. Retrieval is search over documents. Tools are actions (look something up, run code, call an API). Safety filters can be rules. The product is not "just the model."

That is why a fluent demo is not an evaluation, and why a true sentence about the model can be false about the product. For how to check a claim about a named system, see How to Check an AI Claim. Fluency is a style of output, not evidence.

What you can say after this lesson

• AI is the category. Machine learning is one family of methods inside it.
• Search, rules, planning, optimization, and probabilistic reasoning sit beside learning.
• Neural networks and LLMs are specific learned systems, not the definition of AI.
• Products people use are often hybrids. Name the part that did the work before you name the failure.
• Related lessons: Machine Learning Fundamentals, Hardware, Software, and Models, RAG: Retrieval-Augmented Generation, AI Agents & Tool Use.