Skip to content

P1-1.3 Relationship Among AI, Machine Learning, Deep Learning, and Generative AI

Section ID: P1-1.3 Version: v2026.07.20

Section 1.1 organized the scope of the word AI, and Section 1.2 looked at the kinds of problems AI deals with. This section organizes the relationship among AI, machine learning, deep learning, generative AI, and LLM, all of which will keep returning later.

What matters here is not memorizing perfect inclusion relations. What matters is fixing a baseline so that terms from different conceptual levels are not mixed as if they meant the same thing.

In Part 1, the baseline relationship among AI, machine learning, deep learning, generative AI, and LLM is fixed in this section. Even when the terms reappear later, only the minimum connection needed for the current question is carried forward there. When the relationship among the terms themselves needs to be reorganized again, return to this section and to the Concept Glossary.

Separating the Levels of AI, Machine Learning, and LLMs

This section organizes the following questions.

  • What is the safest way to read the relationship among AI, machine learning, deep learning, generative AI, and LLMs?
  • What part can be explained through inclusion relation, and what part becomes mixed differently inside real services?
  • Why does confusion arise when the terms are narrowed into something like AI = LLM?

This section first closes the point that AI, machine learning, deep learning, generative AI, and LLM are terms from different levels. The internal structure of each level continues in Part 4, Part 5, and Part 6. Here the focus is first on distinguishing the place of each term.

Distinguishing Inclusion Relations from Service Mixtures

  • Distinguish the basic relationship among AI, machine learning, deep learning, generative AI, and LLMs.
  • Avoid understanding all of AI too narrowly from recent service experience alone.
  • Understand that even if a broad inclusion relation is roughly useful, real services still mix several technologies together.

Concepts to Connect First

This section is the baseline for the term relationship that Part 1 will keep reusing. The concepts below should be fixed only by position first, and checked again through their glossary entries when fuller definitions are needed.

Concept Meaning to fix first here Why it is needed now
AI the broadest field and system category to see where all other terms sit
machine learning a learning approach that improves performance from data to separate rule-based and learning-based approaches
deep learning an approach that strongly uses neural networks and representation learning to fix a major expansion path inside machine learning
generative AI an output category that produces new content to keep learning method and output character from collapsing into one
LLM a family of large language models to avoid reading generative AI and AI as a whole too narrowly as the same thing

Main Learning Points

What should remain first from this section is that the five terms are not all the same kind of name. The three standards below form the overall map.

Standard Why it matters Level of understanding needed here
AI is the widest category, and machine learning and deep learning are approaches inside it It reduces the mistake of comparing different kinds of terms as if they were peers. Distinguish the broad flow AI > machine learning > deep learning.
Generative AI is a term about what is produced It prevents deep learning, LLMs, and service experience from being collapsed into one block. Distinguish generative AI as an output category that produces text or images.
LLMs are representative examples of generative AI, but not the whole of AI It prevents recent usage experience from shrinking the whole field too much. Distinguish clearly that AI is not the same thing as LLM.

AI is the broadest field and system category. Machine learning is a learning approach that improves performance through data. Deep learning is a method inside that approach that strongly uses neural networks and representation learning. Generative AI is an output category that produces new content. LLM is a representative language-model family inside that space. In other words, read them as AI as outer category, machine learning and deep learning as learning approaches, generative AI as output category, and LLM as a language-model family.

Detailed Learning

Start With the Big Picture

The widest term is AI. AI is the broad field and system category that aims to perform some functions associated with human intelligence through computer systems, machines, and algorithms. Inside it sit different approaches such as rule-based methods, search, planning, probabilistic reasoning, and machine learning.

Machine learning is the approach that improves performance through data or experience. Instead of having people write every rule directly, it lets a model learn decision criteria from input-output examples, rewards, or structure inside data.

Deep learning is an approach inside machine learning that learns representations through neural networks, especially deep networks with multiple layers. Earlier approaches often relied heavily on people designing features directly. Deep learning strongly expands the direction in which useful representations are learned together from raw data.

Generative AI refers to AI models and services that create new content such as text, images, audio, video, or code. Recent generative AI is usually built on deep learning and large-scale models, but deep learning and generative AI are not the same term. Deep learning is closer to a learning method and model structure. Generative AI is closer to the character of the output.

The reason it is described as an output character becomes clearer by comparison. If a system reads a customer inquiry and chooses one among categories such as refund inquiry, delivery inquiry, and account inquiry, that is closer to classification. If the same inquiry is read and the system produces a new answer sentence such as refunds are usually completed within three business days, that is closer to generation. So the more direct boundary for generative AI is not that the input is text, but whether the output is a predefined category or a newly created piece of content.

An LLM is a language-model family trained on large-scale text data. It is one of the representative technologies behind today’s generative AI, but it is not the whole of generative AI. Image-generation models, speech-generation models, and video-generation models are also forms of generative AI without being language models in the same sense.

flowchart TD
  AI[AI]
  Rule[Rule-based AI]
  Search[Search and planning]
  Prob[Probabilistic reasoning]
  ML[Machine learning]
  DL[Deep learning]
  Gen[Generative AI]
  LLM[LLM]
  OtherGen[Image, audio, and video generation models]

  AI --> Rule
  AI --> Search
  AI --> Prob
  AI --> ML
  ML --> DL
  AI --> Gen
  DL -.-> Gen
  Gen --> LLM
  Gen --> OtherGen

This diagram is a learning map. In real research and real services, the boundaries are more complex. For example, when an LLM is attached to a search service, the result may look like generative AI from the outside, but the full system can also include a search engine, a database, permission controls, rule-based filters, and recommendation models.

The key is not memorizing the arrows as a rigid genealogy. The key is reading the place of each term: AI as the widest outer category, machine learning and deep learning on the learning side, generative AI on the output side, and LLMs as a representative language-model family inside generative AI.

Do Not Treat the Terms as Belonging to the Same Level

Confusion happens when terms from different conceptual levels are compared side by side.

Term Main conceptual level Short explanation Point of caution
AI field and system category the broad field of building machines that perform functions associated with intelligence should not be narrowed to one recent model
machine learning learning approach an approach for building models that improve through data or experience not all AI is machine learning
deep learning machine-learning method an approach using deep neural networks and representation learning not all machine learning is deep learning
generative AI output character and service category models and services that produce new text, images, audio, or code generation does not mean factual correctness
LLM model family large language models centered on language data not all generative AI is LLM-based

For that reason, a question such as “Which is better, AI or machine learning?” is unstable from the start. Machine learning is one approach inside AI. Likewise, “Which is newer, deep learning or generative AI?” mixes levels. Deep learning belongs more to model-learning method, while generative AI belongs more to the character of the produced output and to service use.

Inclusion Relation Is Only a Starting Point

At the beginning, one rough guide can be written like this:

AI > machine learning > deep learning > generative AI > LLM

But that line is a very strong simplification. Generative AI is closer to an output and service category, so it should not be treated only as a simple lower stage beneath deep learning. A more accurate explanation has to keep several exceptions in mind.

  • Rule-based AI can still fall inside AI without being machine learning.
  • Machine learning includes many methods that are not deep learning.
  • Deep learning is used not only for generation, but also for classification, recognition, prediction, and control.
  • Generative AI can create text, images, audio, video, and code.
  • LLMs are a representative model family inside generative AI, but they do not mean the whole service.

So inclusion relation is useful as a guide, but when an actual system is explained, it is still necessary to separate what problem is being handled, what model or rules are being used, and what kind of output is being produced.

Separating Again Through Examples

The following table is a practice in distinction.

Case Is it AI? Is it machine learning? Is it deep learning? Is it generative AI? Is it an LLM?
a system that judges loan eligibility by human-written rules yes no no no no
a model that predicts fraud risk from past transaction data yes yes depends usually no no
a neural model that locates defect areas in an image yes yes yes usually no no
a chatbot that produces a summary from a sentence input yes yes usually yes yes usually yes
a model that produces an image from a text prompt yes yes usually yes yes usually not an LLM in the narrow sense

Expressions such as yes, usually, and depends are used because the internal implementation still has to be checked. Services that look similar from the outside can still differ internally.

What beginners most often confuse is one service name and the several conceptual layers inside it. For example, even one chatbot that answers questions over documents can be described in several ways at once.

Description attached to the same service Why that description can still be right What to watch out for
AI service because it is a broad system that receives questions, produces answers, and affects human judgment this only gives the widest category, not the internal method
machine-learning or deep-learning-based service because the core model improving answer quality can rely on data learning and neural representation learning this does not mean every component is deep learning
generative AI service because the final output is a newly generated answer sentence this does not mean the output has already been fact-checked
service containing an LLM because the core language model processes questions and answers retrieval, permissions, logs, and safety filters are not the LLM itself

The key here is not which description is wrong but which level is being described. The same system can be an AI service from the outside, deep-learning-based from the model perspective, generative AI from the output perspective, and an LLM-containing system from the language-model perspective.

Why Generative AI and LLM Must Be Separated

Today, many users first encounter AI through chatbots. Because of that, AI, generative AI, LLM, and product names such as ChatGPT can feel like one block.

But for learning, they still need to be separated. An LLM is a model family centered on language. Generative AI is the broader category that creates new content. The NIST generative-AI profile describes generative AI as a category of AI models that generate synthetic content derived from patterns and properties in the input data, with examples including images, video, audio, and text. It also treats LLMs as one representative common activity within that larger generative-AI space.

From that viewpoint, LLMs are an important case of generative AI, but they cannot explain all of generative AI by themselves. The reverse is also true: building a generative-AI service does not end with one LLM. Real services still need prompt handling, retrieval, permissions, logs, safety filters, user interfaces, and cost control.

The Standard This Book Will Use

From here on, the book uses the following standards.

  • AI is used for the widest field and system category.
  • machine learning is used when a model improves performance from data or experience.
  • deep learning is used when deep neural networks and representation learning are central.
  • generative AI is used when models and services that create new content are being explained.
  • LLM is used for large language models trained on language-centered data.

These standards are refined more closely in later chapters. For now, what matters is fixing the place of the terms. Part 3 first reorganizes data modeling, and the detailed algorithms and learning methods are then recovered again in Part 4, Part 5, and Part 6.

The terms are not finished here all at once. Instead, the section first connects where each one will be read fully again.

Term fixed now Where it is revisited more fully later Why it matters at the current stage
machine learning Part 4 because that is where problem, data, and evaluation are first organized in full structure
deep learning Part 5 because representation learning, backpropagation, CNNs, attention, and Transformers are recovered there as actual computational structure
generative AI Part 6 because generation structure, RAG, tool use, and agents are connected there all the way to service perspective
LLM Part 6 because tokens, next-token prediction, GPT-family models, and instruction tuning are fully explained there as one stream

A Short Practice in Separating Levels

The cases below are meant as a practice in dividing whether something is AI, machine learning, deep learning, generative AI, or LLM.

Case Distinction question to ask first Point most likely to be confused
a system that helps insurance review with a hand-written rule table Did it learn from data, or were the rules written directly by people? It can feel like AI but still not be machine learning.
a prediction model that detects abnormal transactions from transaction records Were patterns learned from data, and is a neural network actually necessary? Machine learning and deep learning are easily collapsed into one.
a service that makes an image from a text prompt Does it create new content, and is the central output text or image? It is generative AI but may not be an LLM.
a chatbot that answers questions over documents Is the central model a language model, and are retrieval and rules attached together with it? It is easy to treat the LLM and the whole service as if they were the same thing.

The point of this exercise is not memorizing inclusion relations. The point is to separate whether the thing in front of you is being named as a field, a learning approach, a model family, or an output category. The same service can still be described at several levels at once: as a generative-AI service, as a system containing an LLM, and as an application where retrieval and rules are attached together.

Cases and Examples

Case 1. Why an Image-Generation Service Is Not Fully Explained as an LLM

Suppose a user sees a service that creates a poster image from a text prompt. Because many recent AI services have looked like chatbots, it can feel tempting to call such a system simply an LLM. But even if the service belongs to generative AI, its central model may not be a language model in the narrow sense. So generative AI and LLM often overlap, but they are not the same term, and the output type still matters. This case shows that generative AI is broader than LLMs.

Case 2. Why a Rule-Based Judgment System May Not Be Machine Learning

Suppose an internal approval system works only through human-written conditions. The user may feel it is AI because it supports business judgment, but if there is no structure that learns patterns from data, it is hard to call it machine learning. In other words, it may belong to AI in the broad sense while still sitting at a different level from machine learning, deep learning, or generative AI. This case shows why the large map AI > machine learning > deep learning still needs to be read together with the actual implementation style.

Checklist

  • You can explain that AI is a broader term than machine learning.
  • You can explain the difference between machine learning and deep learning through the viewpoints of data learning and neural representation learning.
  • You can explain that deep learning and generative AI are not the same term.
  • You can explain the relationship between generative AI and LLMs as representative case, but not the whole.
  • You can explain that a real AI service can be a combination of several components, not one model alone.
  • You can explain that AI, machine learning, deep learning, generative AI, and LLM are not terms from the same level.
  • You can explain that generative AI is a broader category than LLMs, and that LLMs are a representative language-model family inside it.
  • You can explain that real services do not end with one model and should be read together with components such as retrieval, permissions, logs, and safety filters.

Sources and References