What Is Artificial Intelligence?
Before we can learn Machine Learning, we need an honest answer to a deceptively hard question: what do we actually mean by "artificial intelligence" — and why does the field keep redefining it?
You're a software engineer, so your instinct is to build. But "AI" is one of the most overloaded terms in computing — it has meant chess programs, medical rule-bases, spam filters, and chatbots, often in the same decade. If you don't have a mental map of the territory, every later concept ("is a decision tree AI or ML?") becomes confusing. This lesson gives you that map. By the end you'll be able to place any technique — including everything else in this course — in its correct box.
What this lesson delivers
- A working definition of AI, and why no single definition satisfies everyone.
- The four historical ways people have tried to define "thinking machines".
- A compressed history: why AI boomed, why it crashed (twice), and why it works now.
- The major approaches to building intelligent systems — and which one Machine Learning belongs to.
- A precise placement of ML inside AI, so the rest of the course has a home.
1 · What is it? Defining an intelligence we can't define
Here is the uncomfortable truth at the center of the field: we cannot define intelligence precisely, so we cannot define artificial intelligence precisely either. Instead, AI has always been defined by what it is trying to do. A useful starting definition:
Artificial Intelligence is the branch of computer science concerned with building systems that perform tasks which, when done by humans, are said to require intelligence — such as perceiving, reasoning, learning, planning, understanding language, and making decisions under uncertainty.
Notice the slipperiness built into that sentence. It defines AI relative to human ability, which creates a famous moving target known as the AI effect: as soon as a machine can do something reliably, we stop calling it intelligent. Optical character recognition, route planning, and chess were all "AI" until they worked — then they became "just algorithms." This is why the boundary of AI feels like it keeps shifting: it does.
A cleaner way to organize the definitions comes from Russell & Norvig, who point out that historical definitions fall along two axes:
| Concerned with thought | Concerned with behavior | |
|---|---|---|
| Measured against humans | Thinking humanly — model the actual human mind (cognitive science) | Acting humanly — the Turing Test: be indistinguishable from a person |
| Measured against an ideal | Thinking rationally — the "laws of thought", formal logic | Acting rationally — the rational agent: do the right thing to achieve goals |
Modern AI overwhelmingly adopts the bottom-right box: the rational agent. We don't require a system to think like a human or to be conscious. We require it to act so as to achieve its goals as well as possible, given what it knows and perceives. This is the definition that makes AI an engineering discipline rather than a philosophy seminar — and it's the one this course quietly assumes.
Forget "can it think?". Ask: "Given some perception of the world, does this system choose actions that move it toward a goal, and can we measure how well?" That reframing — agent, perception, action, goal, performance measure — is the load-bearing idea of the entire field.
2 & 3 · Why was it needed? What problem does it solve?
As a software engineer you already know how to make a computer do exactly what you specify. So why is a whole field needed? Because a vast category of valuable problems cannot be solved by writing explicit rules, for one of three reasons:
- We can't articulate the rules. You recognize a friend's face or a handwritten "7" instantly, but you cannot write down the procedure you used. The knowledge is real but tacit.
- The rules are too many or too fluid. Consider spam filtering, medical diagnosis, or fraud detection. You could try to enumerate the rules, but there are thousands, they interact, and adversaries change them faster than you can patch.
- The problem requires searching an enormous space. A chess position has more continuations than atoms in the observable universe. No lookup table can hold the answer; the system must reason or search its way to a good move.
AI exists to attack exactly these problems: tasks where the classical software recipe — "a human works out the algorithm, then codes it" — breaks down because the human can't work out the algorithm, or the algorithm is intractable to specify.
If we can't write the rules, maybe we can write a program that discovers the rules from examples. Instead of "programmer → rules → program," we do "data + desired outcomes → learning procedure → rules." That inversion is the single idea that turns AI into Machine Learning. Everything in Parts II–VIII of this course is a way of carrying it out.
4 · How the field developed (the parts that still matter)
I'll spare you the exhaustive timeline and keep only the beats that explain why the field looks the way it does today. The pattern is important: AI has gone through cycles of soaring optimism and crashes ("AI winters"), and the survivor of those crashes is precisely the paradigm you're about to study.
The founding (1943–1956)
McCulloch & Pitts modeled a neuron as a logic gate (1943). Turing asked "Can machines think?" and proposed his imitation game (1950). The field was formally named at the Dartmouth workshop of 1956, whose proposal confidently claimed a summer's work could make real progress on language, abstraction, and self-improvement. That optimism set the tone — and the trap.
The symbolic era & the first winter (1956–1980)
Early AI bet that intelligence is symbol manipulation: represent knowledge as logical statements and search or deduce over them. It produced genuine milestones (theorem provers, the first chatbots, planning systems) but hit two walls: problems exploded combinatorially, and encoding real-world knowledge by hand proved staggeringly hard. Funding collapsed — the first AI winter.
Expert systems & the second winter (1980–1990s)
Industry revived AI with expert systems: hand-crafted rule bases capturing a specialist's knowledge (e.g. MYCIN for infections). They worked in narrow domains and made money — but they were brittle, couldn't learn, and were nightmarish to maintain as rules multiplied. When they failed to scale, the second winter hit.
Hand-coding intelligence — whether as logic or as expert rules — does not scale to the messy, uncertain real world. Both crashes came from the same root cause. The field's escape was to stop hand-coding the knowledge and start learning it from data.
The statistical turn & the modern boom (1990s–now)
From the 1990s, AI reoriented around probability, statistics, and learning from data — the Machine Learning paradigm. Three forces made it explode:
- Data — the internet and digitization produced oceans of labeled examples.
- Compute — cheap parallel hardware (GPUs) made training on that data feasible.
- Algorithms — mature statistical methods, and later deep neural networks (the 2012 ImageNet moment).
The through-line: the approach that survived the winters and now dominates is learning from data. That's not a fashion — it's the paradigm that solved the scaling problem that killed its predecessors. And that paradigm is exactly where this course lives.
5 · The major approaches to building intelligent systems
Across that history, humanity has really only tried a handful of fundamentally different strategies for making machines act intelligently. Knowing these four lets you classify any AI system you meet.
Real systems mix these. A self-driving car uses learning (C) to perceive pedestrians, search/planning (A) to choose a trajectory, and hand-written rules (B) for legal constraints like "never cross a solid line." Intelligence, in practice, is a composite.
The classic, precise definition of the learning approach:
This isn't decoration — it's the template we will fill in for every algorithm in the course. For spam filtering: T = label email as spam/not, E = a corpus of labeled emails, P = fraction classified correctly. Keep the <T, P, E> triple in your head; we'll return to it constantly.
6 · Where Machine Learning fits — the map you'll use all course
Now the payoff. Here is the containment relationship, from widest to narrowest:
Two adjacent terms people confuse with these:
- Data Science is not a subset of AI — it's an overlapping discipline about extracting insight from data (statistics + programming + domain knowledge). It uses ML as one of its tools.
- Generative AI / LLMs (ChatGPT and friends) are applications of Deep Learning — box 4. They sit two levels below where we're stopping, which is exactly why this course won't cover them beyond context.
Linear Regression, Logistic Regression, k-NN, Decision Trees, SVMs, Random Forests, k-Means, PCA — every algorithm in this course lives in box 2 (Machine Learning) and specifically in approach C (statistical learning). None of them are Deep Learning. This is "classical" or "traditional" ML: powerful, interpretable, the right tool for most real-world tabular problems, and the essential foundation before any neural network makes sense.
A first taste of ML's internal structure
Within Machine Learning, problems split by what kind of experience the system learns from. You'll meet this properly in Lesson 4, but here's the preview so the roadmap reads clearly:
| Type | The data has… | Goal | Course examples |
|---|---|---|---|
| Supervised | inputs and correct answers (labels) | predict the answer for new inputs | Regression, Classification (Parts III–IV) |
| Unsupervised | inputs only, no answers | find hidden structure | k-Means, PCA (Part VII) |
| Reinforcement | rewards from acting in an environment | learn a strategy by trial and error | context only — out of scope |
7 · How this is "implemented in practice" (at the field level)
There's no code in this lesson, but it's worth grounding the abstractions. When an organization "does AI" today, the concrete reality is almost always an ML pipeline: collect and clean data → represent it as numbers (features) → choose a model family → train the model by optimizing a numerical objective → evaluate honestly on held-out data → deploy and monitor. Every stage in that sentence is a lesson later in this course. The philosophy of Section 1 becomes, in practice, a very concrete engineering loop — one that will feel natural to you as a software engineer.
8 · Limitations — what "AI" is not
- No general intelligence today. Every system in real use is narrow — brilliant at one task, helpless outside it. "Artificial General Intelligence" remains a research aspiration, not a product.
- Learning ≠ understanding. An ML model captures statistical patterns in data. It has no concept of meaning, causation, or truth unless we engineer that in. It will happily learn our biases and spurious correlations.
- Data quality is destiny. Because the rules come from data, garbage data yields a confident, garbage model. Much of real ML work is data work, not algorithm work.
- Interpretability varies. A decision tree explains itself; a deep network largely doesn't. Choosing a model is partly a choice about how much you need to trust and audit it — a recurring theme in this course.
How this connects to the next lesson
You now have the map: four approaches, and ML as the one that learns from data. In Lesson 2 — How Intelligent Systems Are Built, we zoom into approaches A and B (search and symbolic logic) just enough to feel why they hit walls — the very walls that made the statistical, learning-based approach (C) win. Understanding what ML replaced makes ML itself far less mysterious. After that, Lesson 3 sharpens the AI/ML/DL boundary, and Lesson 4 opens the machine we'll spend the rest of the course building.
Exercises & reflection
No coding yet — these are calibration questions. If you can answer them in a sentence or two, you're ready to move on.
For each system, name which of the four approaches (Search, Symbolic/Logic, Statistical Learning, Bio-inspired) it primarily uses — some use more than one:
- A GPS app finding the shortest route.
- A spam filter that improves as you mark messages.
- A tax-preparation program that applies the legal rulebook.
- A program that plays chess by looking many moves ahead.
Show discussion
(1) Search/planning. (2) Statistical learning — it's ML, specifically supervised classification. (3) Symbolic/logic — a hand-coded rule base, an expert system. (4) Primarily search/planning; modern engines add a learned evaluation (statistical learning), so it's a hybrid — exactly the "intelligence is a composite" point.
Pick a task you'd like a machine to do (e.g. "predict apartment prices"). Write out its <T = task, P = performance measure, E = experience>. If you can't name the experience E as concrete data, the problem isn't yet an ML problem — reframe it until you can.
In one sentence, explain the "AI effect" to a non-technical friend, and give an everyday technology that was once considered AI but no longer is.
AI is the engineering of systems that act to achieve goals on tasks we associate with intelligence. Because many such tasks resist explicit rules, the field's winning strategy became learning the rules from data — Machine Learning. ML is a strict subset of AI; Deep Learning is a strict subset of ML's neural-network family. This course lives entirely at the ML level (classical, statistical learning), which is both the historically robust core of the field and the necessary foundation for anything deeper.