ISTQB CT-GenAI Sample Questions: 10 Free Practice Questions with Explained Answers
Test yourself on prompting, the risks of generative AI and LLM-powered solutions before you sit the CT-GenAI exam. Each answer comes with an explanation. For the syllabus, prerequisites and exam format, see the ISTQB CT-GenAI certification page.
Sample questions written by AutomationDataCamp, not official ISTQB questions. They are original, they are not taken from the ISTQB syllabus or from any official exam, and they do not reproduce the official exam format or difficulty. Use them to check your understanding, then work from the official syllabus published on istqb.org.
Read each question and choose one answer. Then open the explanation to compare. Try to answer in about a minute each, as in a timed exam.
Question 1 of 10 · Concepts of generative AI for testing
A tester pastes a very long specification into an LLM chat and asks for test cases. The test cases cover the first chapters well but ignore the last ones entirely. What is the most likely cause?
A. The model is not trained on software testing standards.
B. The input is larger than the model's context window, so part of it was not taken into account.
C. LLMs are unable to read documents longer than a few pages by design.
D. The model deliberately skips the sections it considers low risk.
Show the answer and explanation
Correct answer: B. LLMs process a limited amount of text at once (the context window, measured in tokens). Content beyond that limit is truncated or not used effectively. Good practice is to split the specification into parts, summarise it, or retrieve only the relevant passages for each request. The other options describe behaviours that LLMs do not have.
Question 2 of 10 · Concepts of generative AI for testing
An LLM-generated test case refers to a requirement identifier, “REQ-482”, that does not exist in the project's requirements. How is this behaviour best described, and what is the right reaction?
A. Prompt injection: block the user who wrote the prompt.
B. Overfitting: retrain the model on the project data.
C. A hallucination: the output looks plausible but is not grounded in the real source, so it must be checked against the requirements.
D. Data drift: the production data has changed since the model was trained.
Show the answer and explanation
Correct answer: C. A hallucination is plausible-looking content that is not backed by facts or by the supplied source. Because LLMs generate the most probable text rather than retrieve verified facts, every reference, figure and requirement in an output must be verified before use. The other terms describe different phenomena.
Question 3 of 10 · Prompt engineering for testing
A team needs test cases written in its own template (ID, preconditions, steps, expected result). The LLM keeps inventing its own layout. Which change to the prompt is the most direct fix?
A. Add two or three complete examples of test cases in the required template (few-shot prompting).
B. Write the whole prompt in capital letters to stress the importance of the template.
C. Ask the model to “be creative” so that it explores more layouts.
D. Remove the requirement text so that the model concentrates on the format.
Show the answer and explanation
Correct answer: A. Giving the model a few worked examples of the expected output (few-shot prompting) is a reliable way to impose a format, and it complements an explicit description of the template. Capital letters and “be creative” do not constrain the format, and removing the requirement would remove the information the test cases must be based on.
Question 4 of 10 · Prompt engineering for testing
Which of the following prompts is the most likely to produce a useful set of test ideas for a login feature?
A. “Test login.”
B. “Give me every possible test for everything around authentication, as many as you can.”
C. “Write test cases like the best testers in the world do.”
D. “You are a senior test analyst. Context: a web banking app with email and password login, lockout after 5 failed attempts. Task: list boundary-value and negative test ideas. Format: a table with ID, input, expected result. Do not invent features that are not described.”
Show the answer and explanation
Correct answer: D. A well-structured prompt states a role, the context, a precise task, the expected output format and constraints. The prompt that names a senior test analyst role does all of these and also asks the model not to invent features. The others are vague, unbounded or give no usable context.
Question 5 of 10 · Managing the risks of generative AI
A team wants to use a public, cloud-hosted LLM to generate test data by pasting real customer records exported from production. What is the main risk?
A. The generated data will be too similar to the original and therefore useless.
B. Personal and confidential data would be sent to a third party, creating privacy, security and compliance exposure.
C. The model will refuse to process numbers.
D. The test environment will run slower because of the added data volume.
Show the answer and explanation
Correct answer: B. Sending production personal data to an external service can breach data protection obligations and company policy. Safer options are synthetic data, properly anonymised data, or an LLM deployment approved by the organisation. Check the provider's terms on data retention and on the use of prompts for training.
Question 6 of 10 · Managing the risks of generative AI
The same prompt is submitted twice to an LLM and returns two different lists of test cases. Which statement is correct?
A. This shows a defect in the LLM and the tool must be replaced.
B. The second answer is always the more reliable one because the model has “learned” from the first.
C. Variation only occurs when the prompt is written in a language other than English.
D. LLM output is probabilistic, so some variation is expected; outputs should therefore be reviewed rather than assumed to be repeatable.
Show the answer and explanation
Correct answer: D. LLMs generate text by sampling from probabilities, so identical prompts can give different results. Settings such as a lower temperature reduce the variation but do not guarantee identical output. Testers should plan for this: review the results, and do not use an LLM output as a stable, repeatable oracle without checks.
Question 7 of 10 · Managing the risks of generative AI
A tester asks an LLM to produce 20 user personas for test data. Almost all of them share the same cultural background and age group. Which risk does this illustrate and what is a sensible response?
A. Bias inherited from training data: vary and constrain the prompt, and review the output for representativeness.
B. Latency: use a smaller model.
C. Prompt leakage: hide the system prompt.
D. Model collapse: stop using generative AI altogether.
Show the answer and explanation
Correct answer: A. Models reflect the patterns and imbalances of their training data. A tester can reduce the effect by specifying the diversity needed (ages, locales, abilities), by checking the output against the user population, and by reviewing it with a human.
Question 8 of 10 · Managing the risks of generative AI
A team uses an LLM to draft test cases from requirements. Which practice best fits responsible use?
A. Accept the output as it is, since the model was trained on thousands of projects.
B. Review only the first test case; if it is correct, the rest will be too.
C. Have a qualified tester review the generated test cases against the requirements, and track simple measures (defects found, review corrections, time saved) to judge the benefit.
D. Replace the review by asking the same LLM whether its own answer is correct.
Show the answer and explanation
Correct answer: C. Generative AI assists the tester; it does not replace accountability. Human review against the source material is essential, and objective measures show whether the tool is really helping. Asking the same model to confirm its own output does not give independent verification.
Question 9 of 10 · LLM-powered solutions and infrastructure
An internal chatbot answers questions about the company's test standards but often quotes an outdated version. The standards change monthly. Which approach lets the chatbot use the current documents without retraining the model each time?
A. Choose a larger model with more parameters.
B. Retrieval-augmented generation (RAG): retrieve the relevant, current passages and supply them to the model with the question.
C. Lower the temperature to zero.
D. Fine-tune the model again after every change of the standards.
Show the answer and explanation
Correct answer: B. RAG keeps knowledge in an external, updatable source and adds the relevant passages to the prompt, so answers reflect the current documents and can cite them. Fine-tuning after each change is slow and costly, a larger model does not know your documents, and the temperature only affects variation.
Question 10 of 10 · LLM-powered solutions and infrastructure
Three weeks after an LLM-based test assistant went live, its answers become noticeably worse. The provider had updated the underlying model version. Which practice would have detected this earliest?
A. Running a fixed set of reference prompts with expected qualities on a regular schedule and after every model change, and monitoring the results (an LLMOps practice).
B. Waiting for users to complain.
C. Rewriting the system prompt every week.
D. Disabling updates permanently on the provider side.
Show the answer and explanation
Correct answer: A. LLM-based solutions need operational monitoring, like any production system: a stable evaluation set, scheduled and change-triggered runs, and tracked results make regressions visible quickly. The other options are reactive, arbitrary, or often not under the team's control.
What to do next
The CT-GenAI requires the CTFL first: practise with our free CTFL mock exam.
Read the official CT-GenAI syllabus on istqb.org and check the learning objectives behind each question you missed.
For structured preparation with timed mock exams and the “ready to sit” indicator, see our certification preparation method. CT-GenAI preparation: €700 incl. VAT, paid once. Enrol now.