ISTQB CT-GenAI: Certified Tester Testing with Generative AI (CT-GenAI)

Last checked November 10, 2026 against the official ISTQB exam page.

1. Question 1Introduction to GenAI for Software Testing

   A test lead uses a generative model to draft test cases from user stories. In one batch, the model cites a requirement ID that does not exist anywhere in the backlog. Which behaviour best explains this?

   1. AOverfitting, where the model memorised its training set and repeats it exactly
   2. BHallucination, where the model invents plausible content without grounding
   3. CData drift, where live inputs have changed since the model was trained
   4. DModel poisoning, where an attacker deliberately altered the model weights
   Show answer and explanationHide answer

   Correct answer: B

   Hallucination means the model produces fluent but unsupported output, such as a requirement ID it has invented. Overfitting would show up as memorised training text, data drift is a change in live inputs over time, and poisoning needs deliberate tampering, for which the scenario gives no evidence.

2. Question 2Prompt Engineering for Effective Software Testing

   A tester asks an LLM to write tests for the login page and receives vague, generic cases. Which change to the prompt is most likely to improve the output?

   1. AAdd the feature scope, the output format and the constraints the tests must meet
   2. BRaise the temperature setting so the model explores more unusual test ideas
   3. CPaste the entire regression suite into the prompt to give the model more context
   4. DSend the same request again in a new session that has no earlier messages
   Show answer and explanationHide answer

   Correct answer: A

   Stating scope, output format and constraints gives the model what it needs to produce targeted, consistent tests. A higher temperature adds randomness rather than specificity, a large pasted suite can dilute focus or exceed the context window, and a fresh session with the same vague request will usually return the same vague result.

3. Question 3Managing Risks of GenAI in Software Testing

   A GenAI tool that generates test personas always assumes customers live in one country and use one date format. Which risk does this best illustrate?

   1. AEnergy consumption, where the tool uses more compute than the team expected
   2. BData privacy, where personal details from the training set appear in outputs
   3. CRegulatory non-compliance, where the tool does not meet a required standard
   4. DBias, where the training data skews outputs toward a narrow set of assumptions
   Show answer and explanationHide answer

   Correct answer: D

   The output consistently reflects one narrow set of assumptions, which is the signature of bias inherited from training data or prompts. Energy use, privacy leakage and regulatory gaps are different risks, and the scenario does not describe them.

4. Question 4LLM-Powered Solutions for Software Testing

   A team wants an assistant that answers questions about its internal test standards, which change every month. Which architecture suits this without retraining the model each time?

   1. ARelying on the model's built-in knowledge and asking it to recall the standards
   2. BFine-tuning the model once a year on a copy of the test standards
   3. CTraining a new foundation model from scratch on the company's source code
   4. DRetrieval-augmented generation that supplies current standards as context
   Show answer and explanationHide answer

   Correct answer: D

   Retrieval-augmented generation pulls the latest standards into the prompt at query time, so changes take effect without retraining. Annual fine-tuning would lag behind monthly changes, training from scratch is impractical, and built-in knowledge cannot contain the company's newest standards.

5. Question 5Deploying and Integrating GenAI in Test Organizations

   A test organisation has little GenAI experience and wants a sensible first step toward adoption. Which step fits best?

   1. ARun a small pilot on one bounded testing activity and measure the results
   2. BRoll the tool out to every project and every tester in the same month
   3. CReplace the existing test automation framework before evaluating the tool
   4. DSet a company-wide target for the number of AI-generated test cases per week
   Show answer and explanationHide answer

   Correct answer: A

   A bounded pilot with measured results produces evidence for deciding on wider adoption. A full rollout, replacing a framework before any evaluation, or a volume target all commit resources before anyone knows whether the tool helps.