Module 1 — What artificial intelligence really is
"Artificial intelligence" is probably the most misused term in all of computing. Depending on who is speaking, it refers to a chatbot, a résumé-screening algorithm, a science-fiction humanoid robot, or a plain statistical formula rebranded for a funding pitch. This module sets an operational definition — the one you can work with — and draws the boundaries of the field.
A definition you can actually use
Philosophical definitions ("simulating human intelligence") are unusable in practice: nobody agrees on what human intelligence is. The working definition we will use throughout this course is more modest and more solid:
An AI system is a program whose behavior was not entirely written by hand, but learned from data or derived from a search through a space of solutions, to perform a task that would require judgment if a human did it.
Three elements matter in this definition.
The behavior is not entirely hand-written. Payroll software applies rules written by humans; it is not AI, however complex it may be. A modern spam filter learned by itself, from millions of examples, what an unwanted email looks like; that is AI.
There is learning or search. The field's two great traditions. The first learns regularities from data (that is machine learning). The second explores a space of possibilities to find a good solution (that is how classic chess programs worked).
The task would require judgment. Classifying an image, understanding a sentence, estimating a risk: tasks with no exact procedure, only clues to weigh.
The AI, machine learning, deep learning hierarchy
These three terms nest inside each other; they are not interchangeable.
Artificial intelligence is the whole field: any program that performs an "intelligent" task, by whatever means. It includes techniques that learn nothing at all, such as rule-based expert systems and search algorithms.
Machine learning is the subset that learns from data. Instead of writing the rules, you provide examples and the program extracts the regularities. It is by far the dominant way of doing AI today — to the point that the two terms are often conflated.
Deep learning is the subset of machine learning that uses deep neural networks — stacks of computational layers. It is what produced the breakthroughs since 2012: vision, speech, translation, and the large language models.
When someone says "we put AI in the product", ask: does the system learn from data, and which data? The answer immediately separates a genuinely learned system, a set of rules with a new label, and a call to an external API. All three can be sound choices — but they are not the same project, the same cost, or the same risk.
What belongs to the field — and what does not
The real scope of AI in business is wider than commonly believed in some directions, and narrower in others.
In scope: demand forecasting, fraud detection, product recommendation, automated document reading, speech transcription, customer segmentation, predictive maintenance, text and image generation. Bounded tasks, with data.
Out of scope, whatever the marketing says: a dashboard (that is visualization), an automation of fixed rules (that is RPA or a script), a well-indexed database (that is data engineering). These tools are useful; calling them "AI" muddles investment decisions.
The honest gray zone: many products combine both. An enterprise search engine may be plain keyword matching, or may embed a semantic embedding model. The right question is never "is this AI?" but "what is learned, from which data, and what does it change for the user?".
The essential vocabulary for the rest of the course
These terms will come back in every module. Let us fix them now.
| Term | Working definition |
|---|---|
| Model | The learned program: a function that turns an input into a prediction |
| Training | The process that adjusts the model on example data |
| Inference | Using the trained model on a new case |
| Feature | A piece of input information given to the model |
| Label | The correct answer attached to an example, when it exists |
| Parameters | The model's internal numbers, adjusted during training |
| Generalization | The model's ability to stay accurate on cases it has never seen |
Two confusions to eliminate right away.
Model ≠ algorithm. The algorithm is the learning recipe (say, "random forest"); the model is the concrete result of applying it to your data. The same algorithm produces different models on different data.
Training ≠ inference. Training is expensive and happens once (or periodically); inference is fast and happens at every use. Hardware needs, costs and security stakes differ between the two phases. Many bad budget estimates come from conflating them.
A model does not "understand", does not "believe", does not "want" anything. It computes an output from an input, following regularities extracted from its training data. This is not pedantry: the projects that fail most brutally are those where the team credited the system with judgment it does not have — and only found out in production.
Key takeaways
- The useful definition: behavior learned from data (or found by search), for a task that would require judgment.
- AI ⊃ machine learning ⊃ deep learning: three nested scopes, not three synonyms.
- Facing any "AI" product, ask what is learned and from which data.
- Model ≠ algorithm, training ≠ inference: four words, two distinctions that shape every budget and every architecture.
In the next module, we will see where this approach comes from: why the field abandoned hand-written rules in favor of learning from data.