ISTQB® CT-GenAI chapter 1: introduction to generative AI for software testing

In short

Chapter 1 of the ISTQB® CT-GenAI syllabus ("Testing with Generative AI", v1.1) sets up the vocabulary for everything that follows. It answers three questions: where generative AI sits within the wider AI field, how a large language model (LLM) actually produces an answer, and in which ways a tester can put it to work. The syllabus allocates 100 minutes of training to it, the second-smallest share after chapter 5.

This is a chapter about understanding, not applying: there are no K3 questions. But its concepts come back everywhere. If context windows or non-deterministic behavior are still fuzzy, the questions in chapters 2 and 3 get much harder. An "easy" chapter on paper, then, and one worth nailing early.

What the exam expects

Chapter 1 accounts for 7 of the 40 questions, worth 7 points (1 point each): 1 K1 question (remember) and 6 K2 questions (understand).

Objective	Level	Questions	What you need to be able to do
GenAI-1.1.1	K1	1	Place symbolic AI, classical machine learning, deep learning and generative AI relative to one another
GenAI-1.1.2	K2	2	Explain how an LLM works: tokenization, embeddings, the transformer model, context window, non-determinism, LLM vs SLM
GenAI-1.1.3	K2	1	Tell foundation, instruction-tuned and reasoning LLMs apart
GenAI-1.1.4	K2	1	Understand what a multimodal LLM (text plus image) brings to a test task
GenAI-1.2.1	K2	1	Give examples of LLM capabilities that help with testing
GenAI-1.2.2	K2	1	Compare AI chatbots with LLM-powered testing applications

GenAI-1.1.2 carries two questions on its own, which makes it the best return on study time in this chapter.

Key concepts

The AI landscape. Symbolic AI reasons with explicit rules. Classical machine learning learns from data, but someone has to pick the useful input variables (the features). Deep learning, built on neural networks, discovers those features by itself in large datasets. Generative AI uses deep learning to create new content. Why it matters for testing: the models come pre-trained, so you can apply them to test tasks straight away, without a training phase on your side.

What happens inside an LLM. Text is split into tokens, each token is turned into a vector (an embedding) that encodes its meaning and its relationships, and the transformer model then predicts the next token, over and over. The output is statistically plausible, which is not the same as correct. Two practical consequences: the context window, measured in tokens, caps how much the model can consider at once (a long execution log can blow past it), and non-deterministic behavior means the same prompt can return two different answers.

Three kinds of LLM. A foundation LLM is general-purpose. An instruction-tuned LLM is derived from it and fine-tuned on prompt/response pairs so that it follows instructions reliably. A reasoning LLM goes further, with multi-step inference that suits high-cognitive-load tasks. In testing, the last two are the ones you actually use, depending on how complex the task is. Small language models (SLMs) are compact versions that are lighter to run.

Multimodal models. A multimodal LLM handles several data types, for example a screenshot together with a user story. Vision-language models are a subset of them. A typical test use: comparing a screen mock-up with the written business rules to spot mismatches.

What an LLM can do for a tester. Analyze and improve requirements, support test case creation, propose expected results (test oracles), generate test data, support test automation, analyze test results, and draft testware such as test plans and reports.

Two interaction models. The AI chatbot: a conversation, instant answers, ideal for exploring or clarifying. The LLM-powered testing application: the model is called through an API from a tool, for well-defined, repeatable tasks at scale. Either way, prompt quality is what makes the difference.

Common pitfalls

• Treating a token as a word. A token can be a character, part of a word or a whole word. Context limits and costs are counted in tokens, not words.
• Mistaking fluency for accuracy. The model produces the most probable continuation. A confident tone says nothing about whether the content is right (see our article on LLM hallucinations in software testing).
• Always reaching for the most powerful model. A reasoning LLM adds nothing when you just need a table reformatted, and a larger context window increases processing time.
• Seeing chatbots and applications as rivals. They are two complementary modes, not two maturity levels. One team can use both at the same time.

How to study it

Plan for one to two hours. Start by reading section 1.1 of the syllabus and restating each concept in your own words. Then do the most useful exercise in the chapter: run the same text (an anonymized user story from your project) through a token counter and look at how it gets split. Finally, ask a chatbot the same question twice and compare the answers. You have just watched non-determinism in action.

To check where you stand before going further, take the free level check (10 questions, 5 minutes). The CT-GenAI certification guide walks through all five chapters, and the exam guide covers the format and scoring. If you want a structured path with exercises, the full course covers every objective in depth.

Two related reads: how CT-GenAI differs from CT-AI, to be sure you are preparing for the right certification, and what changed in syllabus v1.1.

Frequently asked questions

Do I need to know how to code for chapter 1?

No. Chapter 1 asks you to understand mechanisms, not implement them. You need to be able to explain what a token, an embedding or a context window is, and why that matters for a test task, but the chapter's objectives do not ask you to write code or calculate anything.

Are the questions on types of AI technical?

They stay at the level of definitions. GenAI-1.1.1 is a K1 objective: you need to recognize what characterizes each family (explicit rules, features chosen by humans, neural networks, content generation). The trap is usually the similar wording between classical machine learning and deep learning, so remember who selects the features.

Why does chapter 1 matter if it is only worth 7 points?

Because its concepts are reused everywhere. Non-determinism explains the mitigation techniques in chapter 3, context windows and embeddings return with RAG in chapter 4, and the two interaction models shape the tooling choices in chapter 5. Getting this chapter right makes the others easier to understand.