Skip to main content
CertRampLearning

ISTQB · Free practice questions

ISTQB CT-GenAI Practice Questions

15 original CT-GenAI practice questions across all five chapters of the ISTQB Testing with Generative AI syllabus v1.1, weighted roughly like the real exam, with the most questions on prompt engineering and GenAI risks. Each one comes with the answer, a short explanation and the reason every other option is wrong.

Written for the syllabus v1.1 (minor update of 27 April 2026; learning objectives unchanged from v1.0 of 25 July 2025) · Updated October 27, 2026

These are original practice questions written by CertRamp to the official ISTQB exam outline. They are not real exam questions and do not come from any exam. Try each question first, then open the answer to see the explanation for every option.

Domain 1

Introduction to GenAI for Software Testing

GenAI-1.1.3 Distinguish between foundation, instruction-tuned and reasoning LLMsFoundationDomain: Introduction to GenAI for Software Testing

Question 1. A test team needs a model that works through a complex business rule step by step, checking intermediate conclusions before deriving test conditions. Which type of LLM is designed for this kind of task?

  1. Option A: A foundation LLM
  2. Option B: An instruction-tuned LLM
  3. Option C: A small language model fine-tuned on test reports
  4. Option D: A reasoning LLM
Show answer and explanation

Correct answer: D. A reasoning LLM

Foundation models are trained on broad data, instruction-tuned models are further trained on prompt–response pairs to follow instructions, and reasoning models add multi-step logical problem solving on top. Choose the type that matches the complexity of the test task.

  • A (incorrect)Foundation LLMs are general-purpose models trained on broad data and usually need further adaptation to follow instructions well.
  • B (incorrect)Instruction-tuned LLMs follow instructions better than foundation models, but are not specifically extended for multi-step reasoning.
  • C (incorrect)Fine-tuning adds domain knowledge; it does not by itself make a model suited to multi-step reasoning.
  • D (correct)Correct. Reasoning LLMs extend instruction-tuned models with logical inference and multi-step, chain-of-thought problem solving.

Reference: ISTQB: CT-GenAI certification page (syllabus v1.1) (opens in a new tab)

GenAI-1.1.2 Explain the basics of generative AI and large language modelsFoundationDomain: Introduction to GenAI for Software Testing

Question 2. A tester pastes a very long requirements document into an LLM chat and notices that the answers ignore requirements from the beginning of the document. Which concept best explains this?

  1. Option A: The context window: the model can only consider a limited number of tokens at once
  2. Option B: Temperature: the model is set too high and becomes random
  3. Option C: Fine-tuning: the model has not been trained on requirements
  4. Option D: Multimodality: the model cannot read text documents
Show answer and explanation

Correct answer: A. The context window: the model can only consider a limited number of tokens at once

LLMs split input into tokens and can only consider a limited amount of preceding text: the context window. Larger windows help with long documents but increase computational complexity and processing time, so splitting or summarising input is often necessary.

  • A (correct)Correct. Text beyond the context window, measured in tokens, cannot be taken into account by the model.
  • B (incorrect)Higher temperature increases variability, but it does not make the model lose earlier parts of the input.
  • C (incorrect)Lack of fine-tuning affects domain knowledge, not whether earlier input is considered.
  • D (incorrect)The model clearly processes text; multimodality concerns combining text with images or other data types.

Reference: ISTQB: CT-GenAI certification page (syllabus v1.1) (opens in a new tab)

GenAI-1.2.2 Compare interaction models when using GenAI for software testingIntermediateDomain: Introduction to GenAI for Software Testing

Question 3. A company wants GenAI to generate draft test cases automatically for every new user story in its backlog tool, using the same prompts and output format each time, for hundreds of stories per month. Which interaction model fits best?

  1. Option A: Individual testers using a public AI chatbot
  2. Option B: An LLM-powered testing application integrated through an API
  3. Option C: A vision-language model reviewing screenshots
  4. Option D: Manual prompting by the product owner in a chatbot
Show answer and explanation

Correct answer: B. An LLM-powered testing application integrated through an API

AI chatbots offer direct, conversational interaction suited to routine and exploratory tasks. LLM-powered testing applications embed the model through APIs, which gives more customisation and scalability for well-defined tasks.

  • A (incorrect)Chatbots suit ad hoc and exploratory work, but they do not scale well to repeatable, automated generation for hundreds of stories.
  • B (correct)Correct. Integrating LLM capabilities through APIs into a testing application suits well-defined, automated and scalable tasks with consistent prompts.
  • C (incorrect)Vision-language models add image understanding; this does not address the integration need.
  • D (incorrect)Same limitation as any chatbot use, and it moves test design away from the testers.

Reference: ISTQB: CT-GenAI certification page (syllabus v1.1) (opens in a new tab)

Domain 2

Prompt Engineering for Effective Software Testing

GenAI-2.1.1 Give examples of the structure of prompts used in generative AI for software testingFoundationDomain: Prompt Engineering for Effective Software Testing

Question 4. A prompt reads: "You are an experienced test analyst. Our app is a hotel booking system. Derive test conditions from the user story below. Consider only the cancellation rules. Return a numbered list." The user story follows. Which prompt component has been placed in the sentence "Consider only the cancellation rules"?

  1. Option A: Role
  2. Option B: Output format
  3. Option C: Constraints
  4. Option D: Context
Show answer and explanation

Correct answer: C. Constraints

A structured prompt for test tasks combines role, context, instruction, input data, constraints and output format. Recognising each component helps you spot what is missing when an LLM's output disappoints.

  • A (incorrect)The role is "You are an experienced test analyst".
  • B (incorrect)The output format is "Return a numbered list".
  • C (correct)Correct. Constraints set restrictions or special considerations the LLM must follow: here, only the cancellation rules.
  • D (incorrect)The context is the background information "Our app is a hotel booking system".

Reference: ISTQB: CT-GenAI certification page (syllabus v1.1) (opens in a new tab)

GenAI-2.2.2 Apply generative AI to test design and test implementation tasksIntermediateDomain: Prompt Engineering for Effective Software Testing

Question 5. A team wants an LLM to write test cases in their company's specific Gherkin style, with fixed naming rules and step phrasing. Plain instructions have produced inconsistent results. Which prompting technique is most likely to help?

  1. Option A: Zero-shot prompting with a longer instruction
  2. Option B: Few-shot prompting with several examples of correctly written scenarios
  3. Option C: Raising the temperature
  4. Option D: Removing the role from the prompt
Show answer and explanation

Correct answer: B. Few-shot prompting with several examples of correctly written scenarios

Few-shot prompting includes several input–output examples in the prompt, so the model can follow a specific format or style. It is especially useful for generating test cases or scripts that must match existing conventions.

  • A (incorrect)Zero-shot prompts contain no examples, and the team has already seen that instructions alone give inconsistent style.
  • B (correct)Correct. Examples in the prompt show the model the exact style to imitate, which is the strength of few-shot prompting.
  • C (incorrect)A higher temperature increases variability, which makes style consistency worse.
  • D (incorrect)Removing a prompt component does not teach the model the required style.

Reference: ISTQB: CT-GenAI certification page (syllabus v1.1) (opens in a new tab)

GenAI-2.2.5 Select and apply appropriate prompting techniques for a given context and test taskAdvancedDomain: Prompt Engineering for Effective Software Testing

Question 6. A tester wants to use an LLM to analyse a complex epic: first identify ambiguities, then propose acceptance criteria, then derive prioritised test conditions. Each intermediate result must be checked by the tester before it is used. Which technique fits best?

  1. Option A: Meta prompting
  2. Option B: One single prompt that asks for all three results at once
  3. Option C: One-shot prompting with one example epic
  4. Option D: Prompt chaining
Show answer and explanation

Correct answer: D. Prompt chaining

Prompt chaining suits complex test tasks that break down naturally into stages, because the tester can verify each intermediate result. This human verification between steps also reduces the impact of hallucinations and reasoning errors.

  • A (incorrect)Meta prompting lets the LLM generate or improve prompts; it does not structure a task into verified intermediate steps.
  • B (incorrect)A single large prompt gives no opportunity to verify intermediate results and increases the risk of compounding errors.
  • C (incorrect)One example may help with format, but it does not provide step-by-step verification.
  • D (correct)Correct. Prompt chaining splits the task into steps, and each output is checked or refined before it becomes the input to the next prompt.

Reference: ISTQB: CT-GenAI certification page (syllabus v1.1) (opens in a new tab)

GenAI-2.3.1 Understand the metrics for evaluating the results of Generative AI on test tasksIntermediateDomain: Prompt Engineering for Effective Software Testing

Question 7. An LLM generates 50 automated test scripts. The team measures that 42 of them run successfully in the CI pipeline without manual changes. Which evaluation metric are they measuring?

  1. Option A: Execution success rate
  2. Option B: Recall
  3. Option C: Diversity
  4. Option D: Time efficiency
Show answer and explanation

Correct answer: A. Execution success rate

GenAI results for test tasks can be evaluated with metrics such as accuracy, precision, recall, relevance and contextual fit, diversity, execution success rate and time efficiency. Because LLM output is non-deterministic, these metrics should be based on enough runs to be meaningful.

  • A (correct)Correct. This metric is the share of generated tests or scripts that execute successfully: here 84%.
  • B (incorrect)Recall is about finding all relevant instances, such as covering every equivalence partition, not about scripts running.
  • C (incorrect)Diversity measures how varied the generated inputs and scenarios are.
  • D (incorrect)Time efficiency compares the time needed with GenAI against manual creation.

Reference: ISTQB: CT-GenAI certification page (syllabus v1.1) (opens in a new tab)

Domain 3

Managing Risks of GenAI in Software Testing Process

GenAI-3.1.2 Identify hallucinations, reasoning errors and biases in LLM outputIntermediateDomain: Managing Risks of GenAI in Software Testing Process

Question 8. Asked to derive test cases from a user story about password reset, an LLM adds a test for "reset via SMS code". Neither the user story nor any other input mentions SMS, and the product has no SMS feature. What kind of problem is this?

  1. Option A: A reasoning error
  2. Option B: A bias
  3. Option C: A hallucination
  4. Option D: Non-deterministic behaviour
Show answer and explanation

Correct answer: C. A hallucination

Hallucinations are outputs that look plausible but are incorrect or irrelevant to the task. Testers must check generated test artefacts against the test basis, because hallucinated test cases waste effort and can create false expectations.

  • A (incorrect)A reasoning error is a flaw in logical steps, such as a wrong calculation or conclusion; here content has been invented.
  • B (incorrect)Bias means systematically favouring certain information or assumptions; a single invented feature is better described otherwise.
  • C (correct)Correct. The output contains content that is factually incorrect or not grounded in the input: a hallucination.
  • D (incorrect)Non-determinism means the output varies between runs; it does not describe invented content.

Reference: ISTQB: CT-GenAI certification page (syllabus v1.1) (opens in a new tab)

GenAI-3.1.4 Recall mitigation techniques for non-deterministic behavior of LLMsFoundationDomain: Managing Risks of GenAI in Software Testing Process

Question 9. A team uses an LLM in its pipeline to classify failed test results, but the same input sometimes gets a different classification. Which measure reduces this variability?

  1. Option A: Increasing the temperature
  2. Option B: Lowering the temperature and setting a fixed random seed where the model supports it
  3. Option C: Switching to a larger context window
  4. Option D: Adding more examples with different classifications
Show answer and explanation

Correct answer: B. Lowering the temperature and setting a fixed random seed where the model supports it

LLM output is non-deterministic by nature. Lowering the temperature and fixing seeds reduce variation, but because reproducibility cannot be fully guaranteed, automated verification of outputs is still recommended.

  • A (incorrect)A higher temperature makes the output more random, not less.
  • B (correct)Correct. A lower temperature reduces randomness, and a fixed seed improves reproducibility, although full reproducibility is not guaranteed.
  • C (incorrect)A larger context window lets the model consider more text; it does not make the output deterministic.
  • D (incorrect)More varied examples may even increase variation; it does not address non-determinism directly.

Reference: ISTQB: CT-GenAI certification page (syllabus v1.1) (opens in a new tab)

GenAI-3.2.3 Summarize mitigation strategies to protect data privacy and enhance security in Generative AI for software testingIntermediateDomain: Managing Risks of GenAI in Software Testing Process

Question 10. Testers want to use a third-party LLM service to analyse failed transactions that contain real customer names and account numbers. Which measure most directly reduces the data privacy risk?

  1. Option A: Asking the LLM in the prompt not to store the data
  2. Option B: Using a reasoning LLM instead of an instruction-tuned LLM
  3. Option C: Increasing the number of human reviewers of the output
  4. Option D: Anonymising or pseudonymising the data and sending only what the task needs
Show answer and explanation

Correct answer: D. Anonymising or pseudonymising the data and sending only what the task needs

Key privacy risks of GenAI in testing include unintended data exposure and loss of control over how data is used, with compliance consequences such as under the GDPR. Mitigations include data minimisation, anonymisation or pseudonymisation, secure storage and transmission, and regular security audits.

  • A (incorrect)A prompt instruction does not control how the provider processes or retains data.
  • B (incorrect)The model type does not change what personal data leaves the organisation.
  • C (incorrect)Output review helps with quality, but the personal data has already been sent to the service.
  • D (correct)Correct. Data minimisation and anonymisation or pseudonymisation limit what personal data is exposed in the first place.

Reference: ISTQB: CT-GenAI certification page (syllabus v1.1) (opens in a new tab)

GenAI-3.3.1 Explain the impact of task characteristics and model usage on the energy consumption of Generative AI in software testingFoundationDomain: Managing Risks of GenAI in Software Testing Process

Question 11. A team's nightly pipeline calls a large LLM to regenerate all 5,000 test cases from scratch, even though only a few requirements change each day. Which change would most reduce the energy consumption of this GenAI use?

  1. Option A: Only send the changed requirements to the LLM and avoid unnecessary model interactions
  2. Option B: Run the same pipeline at night when energy is cheaper
  3. Option C: Lower the temperature
  4. Option D: Ask the LLM to produce longer, more detailed test cases
Show answer and explanation

Correct answer: A. Only send the changed requirements to the LLM and avoid unnecessary model interactions

The energy use of GenAI depends on task complexity, the model used and how often it is called, and it adds up across many users and runs. Choosing an appropriately sized model and avoiding unnecessary interactions are the most direct levers.

  • A (correct)Correct. Energy consumption rises with usage volume and task complexity, so limiting unnecessary interactions reduces it directly.
  • B (incorrect)The time of day changes the cost, not the amount of energy the model interactions consume.
  • C (incorrect)Temperature controls output randomness, not the computational effort of the requests.
  • D (incorrect)Longer outputs need more tokens to be generated, which increases rather than reduces consumption.

Reference: ISTQB: CT-GenAI certification page (syllabus v1.1) (opens in a new tab)

Domain 4

LLM-Powered Solutions for Software Testing

GenAI-4.1.2 Summarize Retrieval-Augmented GenerationIntermediateDomain: LLM-Powered Solutions for Software Testing

Question 12. A team wants an LLM to answer questions about its own, frequently updated test documentation without retraining the model each time the documents change. Which approach fits best?

  1. Option A: Retrieval-augmented generation (RAG): retrieve relevant document chunks from a vector database and add them to the prompt
  2. Option B: Fine-tuning the LLM after every documentation change
  3. Option C: Raising the temperature to make the model more creative
  4. Option D: Switching to a model with a smaller context window
Show answer and explanation

Correct answer: A. Retrieval-augmented generation (RAG): retrieve relevant document chunks from a vector database and add them to the prompt

RAG splits documents into chunks, stores them as embeddings in a vector database, retrieves the most relevant chunks for each query and passes them to the LLM as context. This keeps answers grounded in up-to-date project knowledge.

  • A (correct)Correct. RAG grounds the response in retrieved, current content at run time, without changing the model itself.
  • B (incorrect)Fine-tuning changes the model by further training; repeating it after every change is costly and slow.
  • C (incorrect)Creativity does not give the model access to the team's documents.
  • D (incorrect)A smaller context window makes it harder, not easier, to provide relevant documentation.

Reference: ISTQB: CT-GenAI certification page (syllabus v1.1) (opens in a new tab)

GenAI-4.1.3 Explain the role and application of LLM-powered agents in automating test processesIntermediateDomain: LLM-Powered Solutions for Software Testing

Question 13. Which statement about LLM-powered agents in test automation is correct?

  1. Option A: Agents eliminate hallucinations because they call external tools
  2. Option B: Agents can only answer questions and cannot perform actions
  3. Option C: Agents can process instructions, retrieve context and invoke predefined tools, and several agents can be coordinated through orchestration
  4. Option D: Agents should always run fully autonomously, even for critical test tasks
Show answer and explanation

Correct answer: C. Agents can process instructions, retrieve context and invoke predefined tools, and several agents can be coordinated through orchestration

LLM-powered agents combine an LLM with context retrieval and tool use to carry out multi-step test tasks. Because they inherit LLM risks, critical work needs verification, for example through semi-autonomous operation.

  • A (incorrect)Agents are built on LLMs and carry the same risks of hallucinations, reasoning errors and bias.
  • B (incorrect)Agents can invoke predefined tools to perform actions, which distinguishes them from a plain chat interaction.
  • C (correct)Correct. This describes semi-autonomous or autonomous agents and multi-agent systems.
  • D (incorrect)For critical tasks, semi-autonomous agents with human or automated verification are a recommended mitigation.

Reference: ISTQB: CT-GenAI certification page (syllabus v1.1) (opens in a new tab)

Domain 5

Deploying and Integrating GenAI in Test Organizations

GenAI-5.1.1 Recall the risks of shadow AIFoundationDomain: Deploying and Integrating GenAI in Test Organizations

Question 14. Before any official GenAI strategy exists, several testers paste requirements and test data into personal accounts of public AI chatbots. What is the main risk of this "shadow AI" use?

  1. Option A: The test cases will be generated too slowly
  2. Option B: Confidential data may leave the organisation without control, creating security and compliance risks
  3. Option C: Public chatbots cannot process requirements written in natural language
  4. Option D: The testers will need a Foundation Level certificate first
Show answer and explanation

Correct answer: B. Confidential data may leave the organisation without control, creating security and compliance risks

Shadow AI is staff using personal or unapproved GenAI tools outside the organisation's policies. It risks data exposure, compliance breaches and intellectual-property disputes, which is why a GenAI strategy defines approved tools and rules for sensitive data.

  • A (incorrect)Speed is not the concern with unapproved tool use.
  • B (correct)Correct. Unapproved GenAI use bypasses the organisation's data protection, security and compliance controls.
  • C (incorrect)Public chatbots can process natural language; the issue is control, not capability.
  • D (incorrect)Certification is unrelated to the risk of uncontrolled tool use.

Reference: ISTQB: CT-GenAI certification page (syllabus v1.1) (opens in a new tab)

GenAI-5.2.2 Recall strategies for cultivating AI skills within test teams to support the adoption of Generative AI in test activitiesIntermediateDomain: Deploying and Integrating GenAI in Test Organizations

Question 15. A test organisation has selected an approved LLM-powered tool. Several experienced testers are sceptical and keep working the old way. Which action best supports the change?

  1. Option A: Make the tool mandatory immediately and measure each tester's usage
  2. Option B: Wait until the sceptical testers leave the team
  3. Option C: Replace manual test design entirely so the old way is no longer possible
  4. Option D: Build skills through hands-on practice with the tool on real test tasks, guided learning, and a community of practice that shares prompts and lessons learned
Show answer and explanation

Correct answer: D. Build skills through hands-on practice with the tool on real test tasks, guided learning, and a community of practice that shares prompts and lessons learned

Adopting GenAI changes testers' tasks and responsibilities; without structured change management, tools are underused, misused or resisted. Capability grows from basic prompting to reusable, test-specific prompt patterns, while testers stay accountable for reviewing GenAI output.

  • A (incorrect)Enforcement without building skills and trust tends to increase resistance rather than adoption.
  • B (incorrect)Ignoring resistance does not manage the change and loses experienced staff.
  • C (incorrect)Removing human test design ignores the need for human verification of GenAI output.
  • D (correct)Correct. The syllabus recommends building GenAI capability through hands-on practice, guided learning, communities of practice and shared prompt libraries, not through theory or mandates alone.

Reference: ISTQB: CT-GenAI certification page (syllabus v1.1) (opens in a new tab)

Want more practice?

15 questions are a sample. The exam is 40 items.

The full CertRamp ISTQB CT-GenAI practice exams have 6 tests that grow with you, from a diagnostic start to an exam-level Test 4 and a challenge set beyond the exam, with an explanation for every option.

Who writes the exams

Written by a practitioner, not a question mill

Joshua Ravnjak, founder of CertRamp Learning

Joshua Ravnjak

Senior Consultant, digital transformation and AI

Founder, CertRamp Learning

921 learners and 61 reviews on Udemy (as of November 3, 2026)

I have more than 12 years of professional experience and work as a senior consultant in an IT management consultancy, focusing on digital transformation, AI and requirements engineering in the insurance industry. I have earned many certifications myself, so I know what exam preparation feels like. That is why every CertRamp practice exam follows the same progression: an easy start to see where you stand, then real exam level, then beyond it. You walk into the exam with a buffer and a good feeling.

Selected certifications

  • IREB® CPRE Foundation Level
  • ITIL® 4 Foundation
  • Professional Scrum Master I (PSM I)
  • Professional Scrum Product Owner I (PSPO I)
  • Certified SAFe® 6 Scrum Master
  • Certified AI-Empowered SAFe® 6 Product Owner/Product Manager

Plus further certifications from Microsoft and Anthropic (Claude).

ITIL® is a registered trademark of PeopleCert. IREB® is a registered trademark of IREB e.V. SAFe® is a registered trademark of Scaled Agile, Inc. PSM and PSPO are trademarks of Scrum.org.

Also in print

Cover: CPRE Foundation Level Prüfungsvorbereitung 2026 by Joshua Ravnjak

CPRE Foundation Level Prüfungsvorbereitung 2026

6 Übungsprüfungen mit 270 Fragen und ausführlichen Erklärungen

Paperback, German, October 2026

My CPRE Foundation Level exam-prep book (in German) follows the same progressive approach.

View on Amazon (opens in a new tab)