Tutorial T-04 · Unit 1
AI, Machine Learning, and Deep Learning
Three terms that appear in every headline, often used as if they meant the same thing. They do not. This tutorial clarifies the relationship between them — and identifies the exception that most simple explanations of this hierarchy quietly leave out.
01 Learning Objectives
By the end of this tutorial you will be able to
- Define artificial intelligence, machine learning, and deep learning, and explain how each term relates to the others.
- Identify the exception to the nested hierarchy: systems that are AI but not machine learning (symbolic AI).
- Explain why the three terms are so often conflated, and why that conflation is misleading.
- Classify a given system as AI-only, machine learning, or deep learning, with justification.
- Recognise when a claim about “AI” is better described as a claim about machine learning or deep learning.
02 Motivating Example
Consider three headlines, all from the same week:
“New AI system beats doctors at detecting skin cancer.”
“Researchers use machine learning to predict crop yields.”
“Deep learning model writes news articles indistinguishable from human work.”
A casual reader would reasonably assume these three stories are describing three different kinds of technology. They are not. All three are almost certainly describing neural networks — the same underlying technology, applied to different problems. The headlines vary because the journalists writing them chose different words from an overlapping vocabulary, and used whichever term felt most appropriate for the audience.
This is not a small problem. It means that when someone reads that “AI is transforming healthcare,” they cannot easily tell whether the claim refers to a rule-based system that any hospital IT department could build, a random forest trained on a modest dataset, or a large language model consuming enormous compute. These are vastly different claims in scope, cost, and risk.
The purpose of this tutorial is to make the three terms precise — so that when you encounter them, you know what is actually being claimed.
03 Concept Explanation
The nested view — and its limits
The standard textbook relationship between the three terms is often drawn as nested circles:
- Artificial intelligence is the broadest term. It refers to any technique that makes a computer behave in ways we would describe as intelligent.
- Machine learning is a subset of AI. It refers specifically to systems that learn patterns from data, rather than being programmed with explicit rules.
- Deep learning is a subset of machine learning. It refers specifically to machine learning using multi-layer neural networks.
The nested picture is a useful starting point and is correct in most practical cases. But it hides one important exception that we will get to in a moment.
Artificial intelligence
The term artificial intelligence was coined at the Dartmouth Conference in 1956, which you read about in T-02. It has always been an ambitious umbrella term, defined more by its goal — making computers behave intelligently — than by any particular method.
This breadth is why the term is so difficult to pin down. A symbolic system that plays chess by searching a game tree is AI. A neural network that recognises faces in photographs is AI. A hand-written rule engine that classifies support tickets is, technically, AI. So is a program that plays tic-tac-toe perfectly.
The term does not tell you how the system works. It tells you what the system is trying to achieve.
Machine learning
Machine learning is more specific. It refers to systems whose behaviour is not programmed explicitly, but is instead learned from examples. The distinction was the subject of T-01: classical programming takes rules as input and produces answers; machine learning takes answers as input and produces rules.
Everything that is machine learning is also AI — ML is a subset of the broader field. But not everything that is AI is machine learning. A symbolic system built entirely on hand-written rules is AI in the broad sense, but it does not learn from data, so it is not machine learning.
This is the exception that the tidy nested diagram omits. The full picture has a region of AI — symbolic AI — that sits outside the ML circle. We will return to it below.
Deep learning
Deep learning is a subset of machine learning, and refers specifically to models built from multi-layer neural networks. The word “deep” refers to the number of layers: a network with many layers is “deep,” whereas a network with one or two layers is not.
Deep learning is currently responsible for most of the impressive results that appear in the news: image recognition, speech transcription, language translation, and the large language models that have dominated public discussion of AI since 2022. It is also the most computationally expensive branch of machine learning, and the one that requires the most data.
Not all machine learning is deep learning. Linear regression, decision trees, random forests, and support vector machines — the algorithms you will meet in Modules 2 and 3 — are all machine learning but none of them is deep learning. They remain extremely useful, and for many problems they outperform deep learning while being faster, cheaper, and more interpretable.
The exception: symbolic AI
The neat nested diagram suggests that AI contains ML, and ML contains DL. This is true for the vast majority of modern systems, but it omits an important category that still exists: symbolic AI, which you read about in T-03.
Symbolic systems — rule engines, expert systems, theorem provers, compilers, classical planners — are artificial intelligence, but they are not machine learning. They do not learn from data; they operate on rules that humans wrote. So they sit inside the AI circle but outside the ML circle.
This category is not a historical curiosity. Every time you use a compiler, a database query optimiser, or a formal verification tool, you are using a symbolic AI system. These tools are everywhere, and they are likely to remain in use for the foreseeable future.
Why the terms get conflated
If the distinctions are clear, why are the three terms so often used interchangeably? There are three main reasons.
- In modern practice, they mostly do overlap. Almost every AI system built today is a machine learning system, and most of the most visible machine learning systems are deep learning systems. For the average company building an AI product, the three terms do effectively refer to the same thing.
- “AI” is a better marketing word. A press release that says “we use AI to detect fraud” sounds more impressive than one that says “we use logistic regression to detect fraud,” even though the second might be more accurate. The term “AI” carries connotations of cutting-edge technology that “machine learning” and “deep learning” do not.
- The terms evolved historically. In the 1970s, “AI” meant symbolic systems. In the 2000s, it increasingly meant statistical learning. Today it encompasses both. When people use the term, they are often drawing on different historical usages without realising it.
Why the distinction still matters
If the terms are effectively interchangeable in practice, does precision matter? It does, for three reasons.
First, for evaluating claims. A headline that says “AI can diagnose disease” could be describing a rule-based expert system from 1985 or a foundation model from 2025. These are radically different technologies with radically different capabilities. Knowing which is being claimed changes how you should interpret the story.
Second, for choosing tools. If you have a problem to solve, “we should use AI” is not a plan. “We should use a decision tree, because we need to explain our reasoning to regulators” is a plan. The distinction between ML and DL in particular matters enormously for cost, latency, and interpretability.
Third, for thinking clearly. The word “AI” has accumulated so much cultural baggage that it is nearly impossible to reason about it precisely. “Machine learning” and “deep learning” are more specific and more tractable. When you find yourself making a claim about AI, it is worth asking: do I mean machine learning? Do I mean deep learning? Or do I mean something broader?
04 Visual Illustration
The diagram below shows the relationship between the three terms. Read it carefully: the nesting is not perfect. Symbolic AI sits inside AI but outside machine learning, and the diagram makes this exception visible rather than hiding it.
One more nuance
In practice, the boundaries between these regions are less crisp than the diagram suggests. Modern systems often combine symbolic and statistical components (as you saw in T-03). A chess engine that uses a learned evaluation function alongside symbolic tree search is simultaneously both. The diagram is a useful map, not a rigid taxonomy.
05 Worked Example
Let us take a single problem — recognising whether a photograph contains a cat — and consider how it might be solved at each of the three levels. The purpose is to see what the distinctions we have just discussed actually look like in practice.
Detecting cats in photographs: three levels
Level 1: AI without machine learning (symbolic).
- A programmer studies photographs of cats and writes rules describing visual features: triangular ears, whiskers, four legs, a tail, fur texture.
- For each new photograph, the system extracts those features using hand-written image processing routines (edge detection, shape matching, etc.).
- The system applies rules: if the image contains two triangular ear shapes AND a cluster of thin lines resembling whiskers AND a pattern of fur texture, classify as cat.
- The programmer iterates, adding rules for edge cases (cats lying down, cats seen from behind, cats with their tails out of frame).
This approach is brittle. Cats in unusual poses, partial occlusion, and unusual lighting all defeat hand-written rules. It was attempted seriously in the 1970s and 1980s and never worked well.
Level 2: Machine learning (not deep learning).
- A programmer decides which features might be relevant: image statistics such as colour histograms, edge density, texture measures, and symmetry. These are called hand-engineered features.
- Thousands of photographs are collected, each labelled “cat” or “not cat”.
- Each photograph is converted into a vector of its feature values. A learning algorithm (such as a random forest or a support vector machine) is trained to map from feature vectors to labels.
- The model learns which combinations of features best predict the presence of a cat.
This works far better than rules, and was the standard approach in the 2000s. The human designer still decides what features to extract, which places a ceiling on how good the system can be: if the relevant visual cues are not among the chosen features, the model cannot use them.
Level 3: Deep learning.
- Thousands or millions of labelled photographs are collected.
- Each photograph is represented as a grid of pixel values — for a 224×224 colour image, that is over 150,000 numbers.
- A convolutional neural network is trained on the raw pixels. The network's early layers learn to detect simple visual patterns (edges, curves, colour gradients). Its later layers learn to detect more complex structures (fur texture, ear shapes, eyes) built from the earlier layers.
- No human specifies what features to look for. The network discovers them from the data.
This is dramatically more accurate than the previous two approaches, and it is what modern image recognition systems use. The cost is that it requires far more data, far more compute, and the resulting model is far less interpretable — it is difficult to say exactly why a particular network classified a particular image as a cat.
The progression illustrates the distinction between the three terms. All three approaches are attempts at artificial intelligence. Only the second and third are machine learning. Only the third is deep learning. And while the third is the most accurate, the second remains a perfectly reasonable choice in many practical settings — especially when data is limited, compute is constrained, or interpretability matters.
06 Interactive Activity
Which term is most appropriate?
For each system below, decide whether it is best described as AI but not ML (symbolic), ML but not DL (classical machine learning), or DL (deep learning). If a system combines multiple approaches, pick the term that best captures its dominant method and note the hybrid nature.
| # | System | Your answer |
|---|---|---|
| 1 | Your mobile phone's predictive text keyboard. | |
| 2 | A tax preparation program that checks whether your return complies with current legislation. | |
| 3 | A system that automatically generates captions for photographs. | |
| 4 | A spam filter built on a Naive Bayes classifier. | |
| 5 | The GPS routing on a navigation app that computes the shortest path between two locations. | |
| 6 | ChatGPT or a similar conversational assistant. | |
| 7 | A bank's system for scoring loan applications using logistic regression. | |
| 8 | A compiler that checks whether a program is syntactically valid. | |
| 9 | A model that predicts tomorrow's temperature using a random forest trained on historical weather data. | |
| 10 | A chess engine combining a neural network with tree search. |
Show suggested answers
1. Predictive text — DL (typically). Modern predictive keyboards use small transformer-based language models. Earlier versions used n-gram statistics, which would be ML but not DL.
2. Tax compliance — Symbolic. The rules are defined by legislation and must be exactly reproducible. No learning involved.
3. Image captioning — DL. Combines a convolutional network (for the image) with a language model (for the caption). Both are deep learning.
4. Naive Bayes spam filter — ML but not DL. A probabilistic classifier from classical machine learning. Still widely used, and still effective.
5. GPS routing — Symbolic. Shortest-path algorithms (Dijkstra's, A*) are classical computer science, not machine learning. The system uses predefined graph structure and rules.
6. Conversational assistant — DL. Modern large language models are deep neural networks. Some systems also include symbolic components (tool use, retrieval, etc.), making them hybrids.
7. Loan scoring — ML but not DL. Logistic regression is a classical ML model. Banks often choose it precisely because it is interpretable and auditable — a regulatory requirement that deep learning models struggle to satisfy.
8. Compiler — Symbolic. Language grammars are formal systems. Parsing is a classical computer science problem.
9. Weather prediction — ML but not DL. Random forests are ensembles of decision trees, not neural networks. Note that some modern weather forecasting uses deep learning — but the system described here is classical ML.
10. Chess engine — Hybrid, spanning symbolic and DL. Modern engines like Leela Chess Zero combine a neural network (DL) with Monte Carlo tree search (symbolic). Neither alone achieves grandmaster level play in the modern game.
The pattern: systems with formal, stable rules tend toward symbolic; systems solving classical computational problems tend toward symbolic; systems with perceptual or linguistic tasks tend toward DL; systems needing interpretability or working with limited data tend toward classical ML.
07 Self-Check Questions
Try each question before revealing the answer. If you get one wrong, re-read the section it refers to before continuing.
Question 1
Which of the following statements about the relationship between AI, ML, and DL is correct?
Show answer
B. The nesting runs from broadest to narrowest: AI contains ML contains DL. The important exception is that symbolic AI falls within AI but outside ML, so “all AI is ML” is false.
Question 2
A compiler that checks whether a program's syntax is valid is best described as:
Show answer
C. A compiler is a symbolic system — it applies hand-written rules (a grammar) to determine validity. It is AI in the broad sense but not machine learning, because it does not learn from data. Option D is a common and understandable objection, but in the technical literature, symbolic systems that perform tasks associated with intelligence are counted as AI.
Question 3
A news headline reports that “AI can now predict heart disease from patient records.” Without reading the article, what is the most reasonable assumption about the underlying technology?
Show answer
B. Patient records are tabular data, and tabular prediction tasks are typically handled by classical ML models (logistic regression, gradient boosting, random forests). Deep learning tends to dominate on perceptual tasks — images, audio, text — not on structured records. That said, the headline uses “AI” precisely because it is vague; only the article will tell you what was actually used.
Question 4
Why is it useful to distinguish between machine learning and deep learning, given that they are both forms of AI?
Show answer
B. The distinction matters practically. Deep learning typically requires more data, more compute, and sacrifices interpretability — in exchange for higher accuracy on perceptual tasks. For many problems, classical ML is the better engineering choice. Option A is false: deep learning is not universally more accurate.
Question 5
A marketer says: “Our product uses AI to recommend movies to users.” What is the most likely technology behind this claim, and why does the choice of wording matter?
Show answer
B. Recommendation systems are dominated by machine learning — often a mix of matrix factorisation and neural approaches. The marketer's use of “AI” is a deliberate choice: the term carries more prestige than “machine learning” while being less specific. Option D is a common but incorrect belief; as this tutorial has shown, the terms are not synonymous.
08 Practice Exercises
-
Classify the system. For each system, state whether it is
best described as AI but not ML (symbolic), ML but not DL,
or DL.
- A compiler that checks the syntax of a program.
- A spam filter using logistic regression.
- A modern voice assistant.
- The routing algorithm in a GPS app.
- A chess engine combining a neural network with tree search.
- A credit scoring model using gradient boosted trees.
- Nest the categories. Place each term in the correct relationship: AI, ML, DL, Symbolic AI. Which is the broadest? Which is the narrowest? Which one does not nest cleanly?
-
Which term? For each claim, which of the three terms would
be most accurate? Justify your answer in one sentence.
- “Our product uses advanced intelligence to optimise your supply chain.”
- “The model was trained on two million labelled photographs.”
- “The algorithm applies the rules of grammar to parse each sentence.”
- Identify the exception. Give three examples of systems that are AI but not ML. For each, explain in one sentence why they fall outside the ML circle.
09 Problem Sets
Analytical problems
-
A company's marketing department announces that their new product “uses AI to provide personalised recommendations.” The engineering team, when asked, says the product uses a decision tree trained on purchase history.
(a) Is the marketing claim accurate? Is it misleading? (b) If the same product were built using a neural network instead, would the marketing claim be more or less accurate? (c) Does the choice of algorithm affect whether the product can reasonably be described as “AI”?
-
The nested diagram presented in this tutorial has a non-nested region (symbolic AI sits inside AI but outside ML). Construct an alternative diagram that would make this exception visually cleaner — one in which every category nests neatly. What would you have to change about the categories themselves to achieve this?
-
Consider the following claim: “Deep learning is just machine learning with more parameters.” Evaluate this claim. In what ways is it accurate, and in what ways does it miss something important about what distinguishes deep learning from other ML approaches?
-
The tutorial distinguishes four regions in the AI/ML/DL diagram: symbolic AI, ML-but-not-DL, DL, and (implicitly) the empty region outside AI. Are there real systems that occupy each region? Give an example for each of the four, and explain what makes each one land in its specific region.
Technical problems
-
The following table describes six systems. Fill in the last column with the most specific correct label (Symbolic AI, ML, or DL) and the middle column with the capability area from T-06 (Perception, Prediction, Language, Ranking, Detection, Generation).
System Capability area Paradigm A grammar checker that flags sentences violating syntactic rules A model that predicts customer lifetime value from purchase history A transformer model that translates between languages A statistical model that predicts which search results a user will click A convolutional network that classifies skin lesions A rule engine that approves or rejects insurance claims -
Estimate the parameter count of three different models you have encountered in this course's examples:
- A linear regression with three features.
- A random forest with 100 trees, each of depth 10.
- A small convolutional neural network for image classification.
For each, explain whether the parameter count puts it in the DL category or not. Is there a specific parameter threshold that separates ML from DL? If not, what actually distinguishes them?
-
A news headline reports that a new system achieves “superhuman performance” on a benchmark. Write the three most important questions you would ask to determine whether the “AI” claim in the headline refers to deep learning, classical ML, or a symbolic system. Then explain how each answer would change your interpretation of the result.
Theoretical questions
- Is deep learning a fundamentally different kind of technique from other machine learning methods, or is it a special case that happens to be particularly powerful? What would it mean for the boundary between the two to be a matter of degree rather than kind?
- The tutorial argues that the terms AI, ML, and DL are conflated partly because “AI” is a better marketing word. Is this an innocent linguistic convenience, or does it have real consequences for how the public understands and evaluates AI systems?
- Symbolic AI sits inside the AI circle but outside the ML circle. Is this the right way to draw the boundary? Could symbolic AI equally be considered as its own separate field, related to AI but not a subset of it?
- If a model is so complex that no human can explain why it made a particular prediction, is it still meaningful to say the model is doing “machine learning” in Mitchell's sense? Does it matter that we cannot verify that the model has actually learned?
10 Discussion and Projects
Discussion questions
- Should journalists and companies be required to be more precise about which term they use? Or is the vagueness inherent to how language works, and something we should simply learn to interpret?
- The tutorial notes that symbolic AI is still widely used in compilers, query optimisers, and verification tools. Why is it that these applications are rarely described as “AI” in the press, even though they technically are?
- If we invented a new term tomorrow that precisely distinguished ML from DL from symbolic AI, would the public's understanding of AI improve? Or would the new terms just acquire the same vagueness as the old ones?
- Do you think the distinction between ML and DL is stable, or will it blur as classical algorithms gain neural capabilities and deep learning models incorporate symbolic components?
Case study: the “AI” hiring decisions at a major tech company
Case study
A large technology company announces that it will use “AI” to screen engineering applicants at the top of its hiring funnel. The announcement is reported widely, with headlines describing an “AI hiring system” that will review millions of applications per year.
The actual system, when described in technical detail by the company's engineering blog, is a rules-based filter: candidates are rejected if their resume does not contain a minimum number of technical keywords, if their years of experience fall outside a range, or if their educational institution is not on a predefined list.
There is no machine learning model anywhere in the system. It is a symbolic filter with hand-written rules. The rules were, however, informed by analysis of historical hiring data — statisticians examined which features correlated with success at the company, and the rules were designed to match those correlations.
Analysis questions:
- Is it accurate to describe this system as “AI”? Which term from this tutorial is most appropriate?
- Does the fact that the rules were informed by statistical analysis of historical data change your answer? Why or why not?
- What practical consequences might follow from the way the system was described in the press?
- If you were the engineering team, would you have pushed back on the marketing description? What would you have said?
Critical-thinking activity: the AI hierarchy quiz
Activity
Working individually, find five different news articles about AI from the past week. For each article:
- Identify the term the article uses (“AI”, “machine learning”, “deep learning”, or a mix).
- Determine, from the article's description, what the underlying technology most likely is.
- Note whether the term the article uses matches the technology it describes.
- Rate how misleading the article's terminology is on a scale of 1–5.
Bring your findings to a group discussion. Which articles were most accurate? Which were most misleading? Is there a pattern in which kinds of publications tend to use which terms? The goal is to develop an instinct for reading AI claims critically.
Homework assignment
Assignment
Write a 1,000-word essay titled “Does it matter what we call it?” addressing the following question:
Should the terms AI, machine learning, and deep learning be used more precisely in public discourse, or is the overlap between them a natural feature of how technical vocabulary evolves?
Your essay should:
- Take a clear position.
- Acknowledge the strongest argument for the opposite position.
- Use at least two concrete examples of AI claims from news or marketing.
- Reference the conceptual distinctions from this tutorial.
The goal is not to reach a definitive answer but to reason carefully about a question that does not have one.
Mini-project: audit an AI product
Mini-project
Choose a real product or service that claims to use “AI”. It can be a consumer app, an enterprise tool, or a research project. Your task is to determine, using publicly available information, what the underlying technology most likely is.
Your deliverable is a two-page report containing:
- The product's own description of what it does.
- The claims it makes about its technology.
- Your assessment of the most likely underlying technology (Symbolic AI, ML, DL, or hybrid), with a justification.
- Any evidence you found that supports your assessment.
- Any gaps in the public information, and what you would need to know to be more confident.
Where public information is scarce, you may make a reasoned inference based on what the product does and what technologies are typically used for that task. The purpose is to practise the reasoning that the tutorial has been building: taking an AI claim and figuring out what it actually means.
11 Summary
Key takeaways from this tutorial
- Artificial intelligence is the broadest term — any technique that makes computers behave intelligently. It says what the system aims to achieve, not how it works.
- Machine learning is a subset of AI: systems that learn patterns from data instead of being programmed with rules.
- Deep learning is a subset of ML: machine learning using multi-layer neural networks.
- The nesting has an exception. Symbolic AI — expert systems, rule engines, compilers, theorem provers — falls inside AI but outside ML. Most simple diagrams omit this region.
- The three terms get conflated because in modern practice they mostly overlap, because “AI” is a better marketing word, and because the terms have shifted meaning historically.
- The distinction matters for evaluating claims, choosing tools, and reasoning precisely. When you encounter a claim about “AI,” ask which of the three terms would be more accurate.
12 Further Reading
The following sources expand on the distinctions covered in this tutorial.
-
Russell, S., & Norvig, P. (2021). Artificial
Intelligence: A Modern Approach (4th ed.). Pearson. Chapter 1.
The introductory chapter includes a careful discussion of what “AI” means, distinguishing four different historical definitions and their implications. -
Chollet, F. (2017). Deep Learning with Python.
Manning. Chapter 1.
Chollet's first chapter is a short, readable account of what deep learning is and how it relates to machine learning and AI. It uses several historical examples similar to those in this tutorial. -
Mitchell, M. (2019). Artificial Intelligence: A Guide
for Thinking Humans. Farrar, Straus and Giroux.
A book-length treatment aimed at a general audience. Mitchell (a cognitive scientist, not the Tom Mitchell of T-01) provides a careful account of what AI systems can and cannot do, with the distinctions from this tutorial running throughout. -
LeCun, Y., Bengio, Y., & Hinton, G. (2015).
“Deep Learning.” Nature, 521(7553), 436–444.
If you want to understand what deep learning actually is, this is a good technical introduction. Written for a general scientific audience, it can be read profitably once you have finished Module 5 of this course.
Next up
T-05 shifts from terminology to mechanics. Now that the three levels are clear, the next tutorial examines how a learning algorithm actually works: what features, labels, parameters, and loss are, and how they fit together in every machine learning system.