AI Learning Hub › Learn AI › Self-Paced Courses › Introduction to AI › Beginner › AI: What It Is—and Isn’t
Level 2 of 4 · Beginner · Lesson 1 of 6
AI: What It Is—and Isn’t
What artificial intelligence, machine learning, generative AI, and large language models are, and why AI is a tool.
About 15 minutes
Skip to lessonBy the end of this lesson you should be able to
- Explain how artificial intelligence, machine learning, generative AI, and large language models relate to one another
- Tell whether an automated system is following fixed rules, classifying or predicting, retrieving, or generating
- Say why it helps to approach AI as a tool rather than as an authority
Why this matters
“AI” is used for a spam filter, a library catalog’s search box, a system that suggests the next word as you type, and a chat assistant that drafts a memo. Those systems work in very different ways, fail in different ways, and need different kinds of checking. A question such as “Can I trust the AI?” has no useful answer until you know which kind of system you are dealing with and what it is actually doing.
This lesson gives you the vocabulary to make that distinction. Later lessons build on it: how generative systems produce their output, why they can sound so human, and what you still have to check.
Four terms, from broad to specific
Artificial intelligence is the broad field: building computer systems that perform tasks we associate with human intelligence, such as recognizing images, understanding and producing language, making predictions, or planning. It names an area of work, not one technology.
Machine learning is one approach within that field, and today the dominant one. Instead of a person writing out every rule, developers give a system many examples, and a training process adjusts the system to perform the task on them. The aim is for it to generalize—to work on examples it has not seen—but training does not guarantee that, so developers test it on held-out examples that were kept out of training. A spam filter trained on thousands of messages that people marked as spam or not spam is a classic case: nobody wrote a rule for each kind of scam. (Generalization and testing on separate data are explained in Google’s Machine Learning Crash Course: “Overfitting” (opens in a new tab) and “Dividing the original dataset” (opens in a new tab); optional and technical.)
Generative AI refers to systems that produce new content—text, images, audio, video, or code—that resembles the patterns in the material they were trained on. Stanford HAI’s definition puts it this way: these systems “learn patterns from training data and generate novel outputs that resemble the original data.” (Source: Stanford HAI, “What is Generative AI?” (opens in a new tab))
A large language model, or LLM, is one kind of generative model, trained on very large amounts of text to predict what text should come next. LLMs are the engines inside most chat assistants. Many current models also accept images or audio, but the core idea is the same one you will meet in the next lesson.
For the machine-learning-based generative systems this course is about, the terms nest, each one a kind of the one above it. This is a simplified map, not a complete classification of every AI system or language model:
- Artificial intelligence — the broad field
- Machine learning — systems that learn from examples
- Generative AI — systems that produce new content
- Large language models — generative models of text
- Generative AI — systems that produce new content
- Machine learning — systems that learn from examples
Nesting means “is one kind of,” not “is the only kind of.” Much machine learning is not generative (the spam filter labels mail; it does not write any). Much generative AI is not a language model (many image generators work quite differently). And some AI systems, past and present, work from explicit rules rather than learned examples—while ordinary automation can follow rules without involving AI at all. The words are related, but they are not interchangeable.
Four things an automated system might be doing
A more practical question than “Is this AI?” is “What is this system doing with my request?” Most automated features do one or more of four things:
- Follows fixed rules
- What comes out: an action or answer someone specified in advance.
What that means for checking: check the rule itself, the inputs it acts on, and whether it was implemented as intended. - Classifies or predicts
- What comes out: a label, score, or likelihood. It may come from a model trained on past examples or from explicit rules; the operation alone does not tell you which.
What that means for checking: a label is a prediction, not a finding. Ask how well the system performs on this kind of case, and look at its errors, especially on unusual cases. - Retrieves
- What comes out: existing items that match a request.
What that means for checking: check that each result is the source it claims to be, that the results are relevant, current, and complete enough for your purpose, and that any extract shown matches the source. A record’s existence does not make its contents true. - Generates
- What comes out: new text, images, or other content.
What that means for checking: it can be fluent and still wrong. Check the claims against something other than the output itself.
Many products combine these. A research tool might retrieve five articles and then generate a summary of them: the articles are real, but the summary is new writing and can misstate them.
Worked example: one inbox, four systems
A constructed example; the office and its settings are fictional.
A student-services office’s shared inbox has four automated features:
- Forwarding. Any message with “transcript” in the subject line goes to the records queue. That is a fixed rule someone typed in. It does exactly what it says, so its gaps are consistent: it will miss a records question titled “Question about my grades” every time.
- Junk filtering. Suspected phishing moves to a junk folder. In this office, that is a classifier trained on examples of phishing and ordinary mail. It can misfile a real message, and nobody in the office knows how often, so someone skims the junk folder.
- Search. Searching for “room booking” lists past messages. That is retrieval. Every result is a message that exists, though what a message says may still be wrong, and a message that says “reservation” instead may not appear.
- “Draft a reply.” A button writes a response to a question about deadlines. That is generation. It writes something that did not exist before, and it can state a deadline that appears nowhere in the office’s actual policies.
Notice two things. Each feature needs a different kind of check. And the forwarding rule did not need AI at all: a simple rule is cheaper, more predictable, and easier to audit. Not every automated process needs AI, and not every automated feature is generative AI.
A tool, not an authority
Every one of these systems was built by people, for a purpose. None is guaranteed to behave as its builders intended: a rule can be written wrongly, a trained system can fail in ways nobody foresaw, and a generative model can produce unreliable output. That includes the most capable generative models. Capability does not give a system authority, judgment, or responsibility for the result. You decide whether to use it, what to give it, and whether its output is good enough to rely on.
That is the stance the rest of this level builds on. Lesson 3 looks at why that stance can be hard to hold when a system writes like a person, and Lesson 6 at what reviewing its work involves. If you want a map of the products themselves, the Getting Started level covers the kinds of AI tools and features you are likely to meet.
Try it: what is each system doing?
Each of these systems is fictional and described in enough detail to decide. For each one, choose: follows fixed rules, classifies or predicts, retrieves, or generates. Some involve more than one.
- A course-registration page refuses to enroll a student unless a prerequisite course appears on the student’s record.
- A library chat service automatically tags incoming questions as “probably about citation format” so they reach the right librarian. The tagger was trained on past questions that librarians had labeled by topic.
- A catalog search returns the three books whose records match “administrative law” published between 2020 and 2024.
- An assistant asked for three practice questions on promissory estoppel writes three new questions.
- A research tool finds five articles in a database and then writes a paragraph summarizing them.
- A calendar app notices that a team usually meets on Tuesdays and suggests Tuesday for the next meeting.
Answers and discussion
- Fixed rule. Someone wrote the condition “prerequisite on record.” If the rule is wrong, it is wrong the same way every time.
- Classifies. The tag is a prediction from a model trained on labeled examples. How far to trust it depends on how well it performs on questions like these, which the description does not tell you; look at its mistakes, especially on unusual questions.
- Retrieves. The results are existing records. The risks are what the search missed and whether the records themselves are accurate, not something the search invented.
- Generates. The questions are new text. They may be excellent, or they may misstate the doctrine; someone who knows the subject has to check them.
- Retrieves, then generates. This combines two different operations. The articles exist, but the summary is generated and can misstate what they say, so “it cited real sources” is not the same as “the summary is accurate.”
- Predicts, with genuine ambiguity. The suggestion comes from a pattern in past data. The app may phrase it in a sentence, but a sentence filled in from a template is not really generated content. Reasonable people could also call this a simple rule (“suggest the most common day”); the product description alone may not tell you which it is.
Human review starts here. Before you decide how much to trust an output, ask what kind of system produced it. A retrieved record can be traced to its source, though what the record says may still need checking. A generated sentence has to be checked against something else.
Optional: where the term came from
The phrase “artificial intelligence” comes from a 1955 proposal by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon for a summer research project at Dartmouth College in 1956. Its opening conjecture was that “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.” (The 1955 proposal, hosted on John McCarthy’s Stanford site (PDF) (opens in a new tab))
Much of today’s generative AI rests on a much later development: the Transformer, a model design introduced in a 2017 paper. Google’s machine learning course describes it as the most successful and widely used architecture for building large language models. (Vaswani and colleagues, “Attention Is All You Need” (2017) (opens in a new tab), a technical research paper; Google, “LLMs: What’s a large language model?” (opens in a new tab))
Sources and further reading
- Stanford HAI: What is Generative AI? (opens in a new tab) — a short definition from Stanford’s Institute for Human-Centered AI, with links to related terms such as large language models.
- Google Machine Learning Crash Course: Introduction to Large Language Models (opens in a new tab) — optional and more technical than this lesson; it assumes earlier modules on machine learning.