Week 2: AI, Machine Learning, and Generative AI

Welcome to Week 2 of the AI Literacy Course.

In Week 1, we learned to notice AI-supported systems in everyday life. In this lesson, we’ll build a clearer vocabulary for understanding different kinds of AI.

These distinctions matter because “AI” is often used as a broad marketing term. Not every automated system is AI, and not every AI system is a chatbot. Understanding what kind of system we’re using helps us decide what evidence, verification and human review are needed.

Learning goals

By the end of this lesson, learners should be able to:

• Explain how automation, AI, machine learning and generative AI are related.

• Define what a model is in simple language.

• Distinguish training from inference.

• Explain the relationship between prediction, probability and classification.

• Identify important differences between generative and predictive AI systems.

• Compare general-purpose and specialised AI systems.

• Explain why context and possible harm affect the need for human review.

1. How the Concepts Relate

The terms automation, AI, machine learning and generative AI are related, but they don’t mean the same thing.

• Automation follows a designed process without requiring a person to repeat every step.

• Artificial intelligence is a broad term for systems that recognise patterns, make predictions, generate content or support decisions.

• Machine learning is one approach within AI. It uses examples to train a model.

• Generative AI produces new content such as text, images, audio, video or code.

A large language model is one type of generative AI model. We’ll examine how language models produce answers in Week 3.

2. Machine Learning and Models

In traditional programming, people give a computer detailed rules to follow. In machine learning, a system uses examples to adjust a model.

A model is a mathematical system trained to find patterns and produce outputs from new input.

For example:

• A spam filter learns patterns associated with unwanted messages.

• An image system learns patterns from labelled pictures.

• A speech-recognition system learns patterns from sound recordings.

• A fraud-detection system learns patterns associated with suspicious payments.

The examples matter. If the data contains gaps, errors or uneven representation, the model may perform less well for some people or situations.

A closer look: labelled and unlabelled examples

Some machine-learning systems train with labelled examples. An image might be labelled “cat” or “dog,” or an email might be labelled “spam” or “not spam.”

Other systems look for patterns or groups without receiving predetermined labels.

In both cases, people still need to decide what the patterns mean and whether they are useful, fair and appropriate for the intended purpose.

3. Training and Inference

Training and inference are two different stages.

Training is the process in which a model adjusts its internal parameters by analysing examples.

Inference happens when the trained model receives new input and produces an output.

For example, an image model may be trained using many labelled images. During inference, it receives a new image and estimates which category it belongs to.

A model may perform well during development but less well in real conditions. Changes in language, population, lighting, background noise or user behaviour may affect the result.

Testing should therefore reflect the real people and situations in which the system will be used.

4. Prediction, Probability, and Classification

Many AI systems predict probabilities or assign categories.

Prediction means estimating a possible outcome. A navigation app may predict travel time, while a bank system may estimate the probability that a payment is fraudulent.

Classification means assigning something to a category. An email may be classified as “spam” or “not spam.”

Prediction and classification are often connected. A spam filter might estimate an 85% probability that an email is spam. If this passes a chosen threshold, the system classifies the message as spam.

A probability isn’t a fact, and thresholds create trade-offs:

• A strict spam filter may block legitimate messages.

• A loose filter may allow harmful messages through.

• A strict fraud detector may inconvenience innocent customers.

• A loose detector may miss real fraud.

People decide where thresholds are placed and which kinds of error are considered acceptable.

5. Generative AI

Generative AI produces new content from patterns learned during training.

It can generate:

• Text and summaries

• Images and illustrations

• Speech and music

• Video and animation

• Computer code

Generative AI can support learning, creativity and work. However, plausible output isn’t automatically accurate, original, fair or suitable for publication.

A generated answer may sound confident and still contain incorrect information. Generated content may also raise questions about personal data, copyright, consent and representation.

The person using the system must review the output and take responsibility for its final use.

6. General-Purpose and Specialised Systems

A general-purpose AI system can support many different tasks. A general chatbot, for example, may draft text, summarise information, explain concepts and generate ideas.

A specialised AI system is designed for a narrower purpose. Examples include systems that analyse medical images, detect fraud or support industrial maintenance.

Broad capability doesn’t mean universal expertise. A chatbot that’s useful for brainstorming may not be suitable for medical diagnosis, legal decisions or formal assessment.

The right tool depends on:

• The purpose of the task

• The quality of the available evidence

• The people who may be affected

• The possible consequences of an error

• The level of qualified human review available

7. Human Choices Remain

AI systems don’t remove human responsibility. People and organisations make important choices about:

• Which problem the system should address

• What data it uses

• What the model should optimise

• How examples are labelled

• Where thresholds are placed

• How results are presented

• When and where the system may be used

Calling a result “automatic” can hide these human and organisational decisions.

Responsibility may be shared among developers, organisations, professionals and users, but it doesn’t disappear because a technical system is involved.

8. Classroom Activity: Build a Concept Map

Create a concept map connecting these terms:

• Automation

• Artificial intelligence

• Machine learning

• Model

• Training

• Inference

• Prediction

• Classification

• Generative AI

• Human review

Use arrows and short explanations to show how the terms relate.

Then compare two systems, such as a spam filter and a general-purpose chatbot.

For each system, answer:

1. What does it do?

2. Is it automation, machine learning or generative AI?

3. What data might it use?

4. What happens during inference?

5. What output does it produce?

6. What could go wrong?

7. Who could be affected?

8. What kind of human review is needed?

9. Reflection Questions

Write short answers to the following questions:

1. Why is machine learning considered part of AI?

2. What is the difference between training and inference?

3. How can training data affect a model’s performance?

4. Why is a probability not the same as a fact?

5. How are prediction and classification connected?

6. What distinguishes generative AI from a spam filter?

7. Why might a model perform differently in real conditions?

8. Why doesn’t automation remove human responsibility?

10. Key Vocabulary

Automation:
A system carrying out a designed process without a person repeating every step.

Artificial intelligence:
A broad term for systems that recognise patterns, make predictions, generate content or support decisions.

Machine learning:
An approach within AI in which a model learns patterns from examples.

Model:
A mathematical system trained to find patterns and produce outputs from new input.

Training:
The process in which a model adjusts its internal parameters using examples.

Inference:
Using a trained model to produce an output from new input.

Classification:
Assigning something to a category.

Prediction:
Estimating a probability or possible outcome.

Threshold:
A chosen point at which a score or probability leads to a particular classification or action.

Generative AI:
AI that produces new content from patterns learned during training.

General-purpose AI:
AI that can support many different tasks.

Specialised AI:
AI designed for a narrower purpose or field.

11. Summary

Automation, artificial intelligence, machine learning and generative AI are connected, but they aren’t interchangeable terms.

Machine-learning models learn patterns from examples during training and apply those patterns to new input during inference. Some systems predict or classify, while generative AI produces new content.

Every system has a purpose, data, model, output and context of use. People also make choices about objectives, labels, thresholds and acceptable risks.

The central lesson is:

Understanding what kind of system we’re using helps us choose appropriate evidence, verification and human review.

Lesson 2 Interactive Quiz: AI, Machine Learning, and Generative AI

Choose one answer for each question. Then click Check my answers. This self-checking quiz does not collect names or scores.

1. Which statement best describes the relationship between AI, machine learning, and generative AI?

A. AI and machine learning are unrelated technologies B. Machine learning is an approach within AI, while generative AI produces new content C. Every automated system is a form of generative AI D. Machine learning only refers to systems that generate text

2. What is a model in machine learning?

A. A complete list of instructions written for every possible situation B. The device or screen used to access an AI service C. A mathematical system trained to find patterns and produce outputs D. A collection of answers that has already been checked by a person

3. What happens during training?

A. A model adjusts its internal parameters by analysing examples B. A trained model receives new input and produces a final output C. A user decides whether to accept or reject an AI-generated answer D. A system converts every probability into a verified fact

4. Which example describes inference?

A. Developers collect and label thousands of training images B. A trained image model classifies a new image C. An organisation decides what problem a system should address D. A teacher explains what artificial intelligence means

5. How can prediction and classification be connected?

A. Classification always happens without probabilities or data B. Prediction creates content, while classification checks its spelling C. They are two names for exactly the same process D. A probability may be compared with a threshold before a category is assigned

6. Why might a model perform well during development but less well in real conditions?

A. A model stops using patterns after training B. Testing results always describe every future situation C. Real users and environments may differ from the development data D. Models can only be used once after they are trained

7. What mainly distinguishes generative AI?

A. It produces new content from patterns learned during training B. It only assigns existing information to fixed categories C. It produces content that is automatically accurate and original D. It works without training data, input or human choices

8. Which comparison between general-purpose and specialised AI is most accurate?

A. A general-purpose system is automatically reliable in every professional field B. A specialised system has a narrower purpose, while a general-purpose system supports many tasks C. A specialised system never requires human review D. General-purpose and specialised systems always use identical data and testing

9. Why is the threshold chosen for a classification system important?

A. It guarantees that every classification will be correct B. It removes the need to examine possible consequences C. It affects the balance between different kinds of error D. It determines how much training data the system has already used

10. Which question best supports responsible selection of an AI system?

A. Can this system produce the longest possible answer? B. Is this the most popular AI tool at the moment? C. Can the system complete the task without any human involvement? D. Does the system fit the task, evidence, risks and available human review?

Reflection: After checking your answers, explain one difference between training and inference, and describe one reason why an AI system might require stronger human review.

}