Machine learning is not the same thing as AI. It is the subfield of AI that learns from data instead of following explicit pre-set instructions, and deep learning is in turn a subfield of machine learning. For teaching the machine learning vs AI question, that nesting is the right spine, as long as students also learn that AI includes logic- and knowledge-based methods that do not learn from data at all. Our judgement is that one continuum (AI, then ML, then deep learning, with generative AI and computer vision as the applications built on them) serves introductory students better than two separate tracks.
TechnoLynx can help an instructor team pressure-test that framing against real engineering work. TechnoLynx would start by agreeing how much nuance students need compared with what client conversations demand, then check that each layer of the slide matches how AI, ML, generative AI and computer vision projects are actually built and run.
Machine learning vs AI: a subset, not a synonym
AI is an umbrella term. It has no universal definition and is used for many different but related ideas. That makes a precise slide harder to write, so we anchor it on the official definitions. The OECD defines an AI system as a machine-based system that, for explicit or implicit objectives, infers from its input how to generate outputs such as predictions, content, recommendations or decisions.
Its core is objectives and outputs. It does not say “learns from data”. Learning is one way to get there, not the definition.
Recital 12 of the EU AI Act names the two technique families explicitly: machine learning approaches that learn from data how to achieve objectives, and logic- and knowledge-based approaches that infer from encoded knowledge or a symbolic representation of the task. The OECD likewise lists machine learning and knowledge-based approaches as technique categories, alongside application areas such as computer vision and natural language processing. For a slide, the takeaway is simple: ML sits inside AI as one of its technique families.
Michael Jordan argued in Harvard Data Science Review (2019) that most of what is labelled AI today, particularly in public, is actually machine learning, and that “AI” is often used as an intellectual wildcard. A 2020 framework paper on ML’s role within AI makes the matching point from the other side: the terms should not be used interchangeably, and it should be clear what role ML plays in a given system. Students should leave knowing both facts. Most of what is publicly labelled AI is ML, and the two words still mean different things.
Can AI exist without machine learning?
Yes. Expert systems, knowledge bases, deductive inference engines, symbolic reasoning and search and optimisation methods are all AI under the European Commission’s guidelines, and none of them learns from data: they reason over rules, facts and relationships encoded by human experts. ISO/IEC 22989 defines symbolic AI as manipulating symbols according to explicitly defined rules to produce declarative outputs, and contrasts it with subsymbolic AI, which uses statistical approaches and produces outputs with a given probability of error.
This is where the neat nested picture starts to strain, and students deserve to hear it. The OECD notes that symbolic AI may itself use machine learning: inductive logic programming learns symbolic logical representations from data, and decision-tree learning learns symbolic rules. Production systems blur the line too. A 2021 survey of ML in K-12 education observes that nearly all ML applications are really combinations of rule-based code and ML models.
None of this breaks the hierarchy. It means the circles describe families of technique, not boxes that every real system fits into cleanly.
The standard one-slide figure
One figure already does this: the Venn diagram in Goodfellow, Bengio and Courville’s Deep Learning, which shows deep learning as a kind of representation learning, representation learning as a kind of machine learning, and machine learning as something “used for many but not all approaches to AI”. That last phrase is the important one. The figure itself carries the caveat about non-ML AI, so instructors do not have to bolt it on.
Our version of that figure, with the applications added:
| Layer | What it is | Example to put on the slide | Anchor |
|---|---|---|---|
| AI | Systems that infer outputs (predictions, content, recommendations, decisions) from input for given objectives | Any of the below, plus rule-based expert systems | OECD, EU AI Act |
| Knowledge-based AI (outside ML) | Reasoning over rules and facts encoded by experts | Expert system, search and optimisation | European Commission guidelines |
| Machine learning | Techniques that improve performance through exposure to training data instead of explicit instructions | Decision tree, linear regression | OECD, CS2023 |
| Representation learning | A kind of machine learning; deep learning is one kind of it | Learned feature extractor | Goodfellow et al. |
| Deep learning | ML using layered neural networks that learn features from raw data | Convolutional neural network | European Commission guidelines |
| Applications | Problem areas, not technique layers | Computer vision, NLP | OECD application areas |
The bottom row is the one to draw with most care. On our version of the slide, computer vision and generative AI are not a fifth and sixth circle inside deep learning. They are application areas, drawn as bands running across the nested circles.
Where deep learning, generative AI and computer vision sit
Deep learning is a subset of machine learning that uses layered neural networks for representation learning, can learn features from raw data without manual feature engineering, and typically needs large amounts of training data. It is the technology behind many recent breakthroughs in AI.
Generative models are a subset of machine learning designed to create new, realistic data from learned patterns. NIST’s 2025 adversarial machine learning taxonomy divides AI systems into two broad classes, predictive AI and generative AI, and notes that most industrial applications are still predictive. That split is worth a sentence on the slide, because most industrial AI is still predictive.
Generative AI is broader than large language models. LLMs belong in the generative band as one prominent family, not as its definition.
Computer vision is a field of AI that derives meaningful information from images, video and other visual inputs. Modern computer vision applications use AI and ML for tasks such as object identification, character recognition and classification, and they now include generative applications that answer questions about images or create new ones. So computer vision crosses the generative band as well, which is why we draw the application row across the circles instead of nesting it inside one.
Placing ChatGPT on the teaching slide
Both: ChatGPT is a machine learning model, and machine learning is part of AI. OpenAI describes ChatGPT as fine-tuned from a model in the GPT-3.5 series, with reinforcement learning from human feedback used in its training. That makes it a machine learning system, hence AI, and a generative application in the predictive/generative split. The European Commission’s guidelines use GPT-based systems as their example of machine learning models that generate content. On the slide, ChatGPT is a good example to place in the machine learning circle and the generative band at the same time.
Should AI and ML be taught as one continuum or two tracks?
Teach one continuum at introductory level. This is our judgement, and the curriculum frameworks below are its reasons.
The AI4K12 guidelines, developed since 2018 as national guidelines for K-12 AI teaching, organise everything under Five Big Ideas: perception, representation and reasoning, learning, natural interaction and societal impact. Learning is Big Idea 3, and machine learning is defined there as a learning algorithm that constructs a reasoner by adjusting the internal representations of a reasoning model such as a decision tree or neural network. That definition only makes sense once students know what a reasoner and a representation are, which is the argument for one continuum: ML is easiest to understand as a new way of building something students have already met. The same guidelines want students to see learning as a simple, mechanical process with no self-awareness or magic involved.
At undergraduate level, the CS2023 curriculum increased ML’s recommended core hours because ML underlies most recent AI advances. Its authors are candid that there is no universal agreement on what an AI curriculum should contain. Both unified AI literacy curricula and separate curricula exist, so a separate-track design is not wrong in itself.
Split the tracks only when a unified treatment is unworkable: for example, when timing forces ML to be taught before any AI context, or when students’ maths background makes the ML half far heavier than the rest. The K-12 survey notes that rule-based programming rests on Boolean logic and discrete mathematics that can be taught early, while ML depends on statistics and probability. That gap is a sequencing problem first, and only a reason to separate tracks when the sequence cannot absorb it.
A sequence that keeps the continuum honest
- Start with the definition shared by the OECD, the EU AI Act and ISO/IEC: a system that infers outputs from input for given objectives.
- Introduce the two technique families, knowledge-based and learning-based, and let students build a small rule-based system first, since its logic is accessible early.
- Present ML as a learning algorithm that constructs a reasoner from training data.
- Narrow to representation learning and deep learning using the Venn figure.
- Add computer vision and generative AI as applications cutting across the layers, with generative AI framed wider than LLMs.
- Close on where the picture strains: symbolic AI that learns, and real systems that mix rules with models.
A framing that still holds up in front of clients
The simple framing survives client conversations if it keeps the hierarchy and drops nothing structural. The failure mode to guard against is a slide that silently equates AI with ML, or generative AI with chatbots.
This matters beyond vocabulary. The European Commission’s guidelines say whether a piece of software counts as an AI system should be decided from its specific architecture and functionality, and they exclude systems based solely on rules defined by people to execute operations automatically.
The Commission’s guidelines list content as a separate output category because of its prevalence in generative AI systems. Revisit the slide as terminology shifts, against the official definitions rather than the news cycle.
The continuum also has a practical end that a conceptual course does not cover. Once a model leaves the notebook it has to be versioned, deployed, monitored and retrained, the production discipline of MLOps.
What an instructor team still has to settle is how much of the strain to show, and when. Show none and the gaps described above stay hidden from advanced students. Show it too early and beginners lose the spine before they have it.
FAQ
How much maths do you need for machine learning?
ML rests on more advanced maths than rule-based AI. A 2021 survey of K-12 ML teaching notes that ML rests on statistics and probability, and CS2023 recommends that AI students have calculus-based probability and statistics, optimisation and experimental design.
Is AI or machine learning harder to learn?
It depends on which part of AI. The basics of rule-based AI rest on Boolean logic and discrete mathematics and can be taught early, while ML needs more advanced mathematical concepts.
