IB COMPUTER SCIENCE | 2027 CASE STUDY | SL & HL

01.2 What is generative AI?

Explain generative AI, distinguish generation from classification and retrieval, and connect models, prompts and outputs to Visionary Studios.

Understand the technology

What changes when the output is new content?

For May and November 2027. This introductory lesson is shared SL and HL content.

01 | A COMPUTER THAT CREATES AN IMAGE

A designer types: “An illustrated explorer standing beside a mountain lake at sunrise.” A moment later, an image appears. The designer did not specify the colour of every pixel or draw every line. A trained system produced an image in response to the description.

Generative artificial intelligence is AI that produces content using patterns learned from data. Its outputs can include images, text, sound, video or code. In this case study, we focus on images and how they could support the work of Visionary Studios.

The explorer is our fictional teaching example, not a character named in the official booklet. A convincing output would still need checking against the designer’s brief.

02 | WHAT DOES ARTIFICIAL INTELLIGENCE MEAN HERE?

Artificial intelligence is the broad area of developing computer systems to perform tasks associated with intelligence, such as recognising patterns, making predictions or generating content. It does not mean that every AI system can perform every intellectual task.

Machine learning is an approach in which a system learns patterns from examples rather than having every decision explicitly programmed. Many modern generative systems use neural networks, which contain adjustable numerical parameters.

A useful relationship is: AI is the broad field; machine learning is one approach within it; generative AI describes systems that produce content. Not all AI is generative, and generative AI is not limited to images.

03 | RECOGNISING AND GENERATING ARE DIFFERENT TASKS

Imagine giving a photograph to a system and asking, “Does this contain a mountain?” A classifier processes the image and returns a category or score. It is recognising something about an existing input.

Now ask a system to create an image of a mountain that matches a description. It must produce image content rather than simply return a label. That is a generative task.

Classification

Input: an image.

Output: a category such as “mountain”.

Generation

Input: a description and other generation inputs.

Output: a newly produced image.

Both tasks may use learned patterns. A larger application can combine them, for example generating candidates and then checking them with another model.

04 | WHAT IS A MODEL?

A model is a learned computational system used to make predictions or produce outputs. In a neural network, training adjusts many numerical values called parameters . These influence how inputs are transformed into outputs.

Think of the model as the learned machinery, rather than the whole website or app you click on. An image-generation application may also contain a user interface, safety checks, storage and editing tools.

Different applications can use the same underlying model, and an application can offer several models. Visionary Studios therefore needs to evaluate the complete workflow as well as the model’s outputs.

05 | LEARNING FROM EXAMPLES

A training dataset is a collection of examples used during learning. Depending on the system, it may contain images with captions or other information. Training can develop patterns involving shapes, textures, objects, styles and relationships between visual content and descriptions.

The training algorithm changes parameters according to a learning objective. We will examine training more closely in 01.3; for now, separate this learning stage from a designer later requesting an image.

The examples and learning process influence what the model can produce. A model trained on a narrow range of visual material may struggle with requirements outside that range. More data alone does not guarantee better, fairer or legally suitable results.

06 | WHAT DOES “LEARNED PATTERNS” ACTUALLY MEAN?

Suppose training images frequently show reflections below objects in water. A model can learn statistical relationships between those visual features. When generating a lake scene, those relationships can influence the output.

This is not the same as a person standing beside a lake, understanding the physics and checking whether each reflection is correct. A generated image may look plausible while containing an impossible reflection.

Students should avoid explaining the process as either “the computer understands everything” or “it just pastes stored photographs together”. Both descriptions hide the learned numerical process and its limitations.

07 | GENERATION USES A TRAINED MODEL

After training, a model can be used to produce outputs. This use is called inference ; for image creation we often call it generation . A designer supplies the inputs supported by the application, and the system computes an output using the trained model.

A normal image request does not necessarily retrain the model. Revising a prompt and generating again usually changes the generation inputs, rather than changing the learned parameters.

Some services have separate data collection or training policies. Those policies should be investigated rather than assuming that every request immediately teaches the model or that no request can ever be retained.

08 | A PROMPT GUIDES THE RESULT

A prompt is an input used to guide generation. For text-to-image generation, it is a textual description. It might specify the subject, setting, composition, lighting and style.

Compare “an explorer” with “a full-body illustration of an explorer in a green jacket beside a mountain lake, with space on the left for an advertising headline”. The second gives more guidance about the desired result.

A prompt is guidance, not a guarantee. The system may omit a detail, misunderstand a relationship or create an unsuitable arrangement. A longer prompt can also contain conflicting instructions, so useful prompting involves clarity and review rather than length alone.

09 | GENERATION CAN USE OTHER GUIDANCE

Text is not the only possible input. The booklet discusses conditional generation using class labels, sketches and segmentation maps. A sketch can guide a layout; a segmentation map identifies image regions by category, such as sky, road or person.

Conditional generation produces content guided by specified information. Text-to-image generation is also conditional because the text guides the output. Image-to-image translation uses an input image to guide a transformed result.

Unconditional generation does not use specified content guidance such as a prompt or class label for each output. It still uses a trained model and generation inputs such as a random starting value. “Unconditional” does not mean “no computation or no inputs of any kind”.

10 | WHY CAN THE SAME PROMPT GIVE DIFFERENT IMAGES?

Many image generators use randomness as part of the generation process. The same description can therefore lead to different candidate images. This variation can help a designer explore alternative compositions.

A seed is a value used to initialise a pseudorandom process. Recording a seed can help reproduce a result when the model, software and other settings also remain compatible and unchanged. A seed alone does not guarantee identical output across different services or versions.

For Visionary Studios, variety is useful during concept exploration, but uncontrolled changes can be a problem when a recurring character must remain recognisable across a campaign.

11 | GENERATION IS DIFFERENT FROM IMAGE SEARCH

An image search system retrieves existing images or references to them. A generative model computes an image using learned parameters and generation inputs. It is not normally explained as searching its training collection for the single closest photograph.

However, this distinction does not guarantee originality or freedom from copying concerns. Models can memorise training material and sometimes produce close reproductions. Some applications also combine generation with retrieval or reference images.

Visionary Studios should therefore investigate training data and inspect outputs. “AI-generated” is not proof that an image is unique, factually correct or unrestricted for commercial use.

12 | A USEFUL RESULT STILL NEEDS A DESIGNER

Consider a proposed workflow: a designer prepares a brief, generates candidates, reviews them, changes the guidance, edits a promising result and obtains approval. The model contributes to the process; it does not automatically complete every stage.

A designer might reject an image because the product is inaccurate, the character changes appearance or the layout leaves no room for text. An output can be visually impressive and still fail the job.

The case study says the company is evaluating AI to streamline production. It does not state that designers will be removed. Explain potential changes to the workflow as possibilities that need evidence.

13 | POSSIBILITIES AND LIMITATIONS

Generative AI can help produce many visual directions, transform descriptions into candidate images and support creative experimentation. These capabilities may reduce some forms of manual work or help a team explore options it would otherwise not try.

Limitations include imperfect instruction following, inconsistent characters, biased representations and computational demands. An output is not evidence that the depicted event happened or that the objects would work in the real world.

Assess suitability for the actual use. An unusual distortion might inspire concept art but be unacceptable in an image intended to show a product accurately.

14 | GENERATION TYPE AND MODEL ARCHITECTURE

“Text-to-image” describes how content is guided and what is produced. “Diffusion” describes a modelling approach. These labels answer different questions: a diffusion model can be part of a text-to-image system.

Diffusion models are central to the shared SL and HL case study content. GANs and hybrid models are additional HL sections. You do not need their mechanisms to understand this introduction; later lessons build them carefully.

Avoid treating text-to-image, conditional generation and diffusion as three mutually exclusive options. A system may fit all three descriptions.

15 | APPLY IT TO VISIONARY STUDIOS

For our fictional explorer campaign, generative AI could help the studio produce candidate settings from a written brief. A designer could compare compositions before selecting and refining a direction.

The studio would still need evidence that the system maintains the character and style across images. It should also record rejected candidates and editing time, because rapid generation does not necessarily mean rapid production of acceptable work.

Explain: Generative AI could help Visionary Studios explore more campaign concepts because it can produce candidate images from descriptions. However, inconsistent outputs may require repeated generation and editing, therefore the company should measure time to an approved result.

16 | TRY IT ON PAPER

Write a brief for a fictional illustrated campaign image. State the subject, setting, composition and style. Predict two ways the output could fail your brief and explain how a designer would detect those failures.

Then describe how your task differs from classifying an existing image and from searching for a photograph. You do not need access to a paid generator or an account to complete this activity.

Reveal a possible approach

“An illustrated explorer in a green jacket beside a mountain lake at sunrise, positioned on the right with clear space on the left for a headline.” A reviewer could check the jacket and composition. If the explorer fills the left side, the image fails the layout requirement even if it looks attractive.

Classification would assign a label to an existing input. Search would retrieve an existing image. Generation computes candidate image content using a trained model and the supplied guidance.

Read this lesson alongside the official 2027 case study. The campaign example is independent teaching material.

CHECK YOUR UNDERSTANDING | CLASSIFICATION, GENERATION OR RETRIEVAL?

Identify the task described. Focus on whether the result is a label, produced content or an existing image.