# Free CT-AI practice questions

> 10 free CT-AI practice questions with the correct answer and an explanation.

Page: https://devopscodelab.com/exams/istqb-ct-ai-v2-0/free-practice-questions/
Practice exams: https://devopscodelab.com/exams/istqb-ct-ai-v2-0/

### Question 1 of 10

A test lead is explaining to stakeholders how technical standards relate to AI regulation. Which statement is consistent with the CT-AI syllabus?

- A. Technical standards such as ISO/IEC TR 29119-11 translate high-level aims into practical guidance, for example on testing.
- B. Technical standards are issued by the EU as annexes to the EU AI Act and carry the same financial penalties as the Act itself.
- C. Technical standards are soft law instruments that replace the need for national regulation of AI-based systems.
- D. Technical standards apply only to safety-related systems and are not relevant to demonstrating regulatory compliance.

**Answer:** A

**Explanation:** The syllabus states that technical standards from bodies such as ISO and IEEE translate high-level aspirations into practical implementations, bridging policy and practice; ISO/IEC TR 29119-11 gives guidance on testing AI-based systems, which helps demonstrate regulatory compliance, and the ISO/IEC 42119 series is being developed. Soft law instruments are things like the OECD AI Principles, and they do not replace regulation. Standards are not limited to safety-related systems. They are developed by standards organizations, not issued as EU AI Act annexes with penalties.

### Question 2 of 10

A company is developing an AI-based system that will be used to screen job applicants in the European Union. How does the EU AI Act MOST directly affect the testing of such a system?

- A. It applies the same light-touch requirements to every AI-based system, so testing obligations do not depend on the use case or on its level of risk.
- B. It replaces technical standards such as ISO/IEC TR 29119-11, so testers no longer need any other guidance on testing AI-based systems.
- C. It categorizes systems by risk, and high-risk systems face strict requirements such as rigorous testing, data governance and human oversight.
- D. It is a soft law instrument like the OECD AI Principles, so it offers guidance on testing but carries no financial penalties.

**Answer:** C

**Explanation:** The EU AI Act takes a risk-based approach, categorizing AI-based systems from minimal to unacceptable risk; high-risk systems, such as those affecting fundamental rights, face stringent requirements including rigorous testing, data governance and human oversight, backed by penalties based on global turnover. It does not apply identical light-touch rules to all systems. It is binding regulation, unlike soft law instruments such as the OECD AI Principles. Technical standards such as ISO/IEC TR 29119-11 complement regulation by translating it into practical guidance.

### Question 3 of 10

A quality manager familiar with ISO/IEC 25010 asks how ISO/IEC 25059 relates to it. Which statement is CORRECT?

- A. ISO/IEC 25059 extends the ISO/IEC 25010 quality model with new and modified characteristics for AI-based systems.
- B. ISO/IEC 25059 covers only quality in use, while ISO/IEC 25010 remains the sole source for product quality of AI.
- C. ISO/IEC 25059 reintroduces usability as a characteristic, reversing the change made in the 2023 version of ISO/IEC 25010.
- D. ISO/IEC 25059 replaces ISO/IEC 25010 entirely, so its traditional characteristics no longer apply to AI-based systems.

**Answer:** A

**Explanation:** ISO/IEC 25059 extends the quality model from ISO/IEC 25010 to address AI-specific considerations, evaluating AI-based systems from both product quality and quality in use perspectives with new and modified characteristics. It does not replace ISO/IEC 25010. It adds characteristics to both perspectives, such as AI robustness in product quality and societal and ethical risk mitigation in quality in use. In the 2023 version of ISO/IEC 25010, interaction capability replaced usability, and ISO/IEC 25059 adds user controllability and transparency under interaction capability.

### Question 4 of 10

How is the activation value of a neuron in a hidden layer of a deep neural network calculated?

- A. The neuron's bias is replaced by the loss computed for the current input, and the result is multiplied by each incoming connection weight.
- B. The neuron compares its single input with a threshold learned from the test data and outputs either zero or one.
- C. A weighted sum of the previous layer's activation values plus the neuron's bias is passed through an activation function.
- D. The inputs from the previous layer are averaged with equal weight, and the mean is passed directly to the next layer.

**Answer:** C

**Explanation:** Each neuron after the input layer computes a weighted sum of the activation values from connected neurons in the previous layer, each connection having its own weight, adds its individual bias, and passes the result through a non-linear activation function. Equal-weight averaging ignores the learned weights and the activation function. The loss is computed at the output and fed back during training to adjust weights and biases, not used in place of the bias. Neurons in a deep network receive multiple inputs, and thresholds are not learned from test data.

### Question 5 of 10

A team uses 10-fold cross-validation for a model that detects a rare disease present in only 3% of the records. What does the CT-AI syllabus recommend to keep each fold representative?

- A. Reduce k to 2 so that each fold contains enough records, and report the better of the two fold results
- B. Assign records to folds using stratified sampling so that each fold has a similar share of positive cases
- C. Use leave-one-out cross-validation instead, because it removes the need for any fold to be representative
- D. Assign all positive records to a single fold so that the model can be validated specifically on that disease

**Answer:** B

**Explanation:** The syllabus states that data is typically randomly assigned to folds, often using stratified sampling to make each fold representative, especially with imbalanced data or small datasets. Putting all positive cases in one fold would leave the other training folds without them. Reporting the better of two results is biased, and the syllabus averages fold metrics. Leave-one-out is a special case of k-fold and does not remove the need for a sound evaluation; its average is still optimistically biased without a hold-out set.

### Question 6 of 10

A test manager compares a locked traffic-sign classifier with an adaptive product recommendation engine. Which statement about their testability is CORRECT?

- A. The locked classifier is always perfectly deterministic, so a single execution of each test case is enough to establish its behavior and expected results.
- B. The adaptive engine may change its behavior during or because of testing, so an automated suite of core checks is re-run after significant changes.
- C. The adaptive engine cannot be tested before deployment because its future behavior is unknown, so testing starts in production only.
- D. When the locked classifier is retrained and updated, only the test cases that relate to the changed classes need to be executed again.

**Answer:** B

**Explanation:** Adaptive systems are harder to test because they may change their behavior during or as a result of testing, and unpredictable new behaviors cannot be tested in advance; an automated suite focused on core functionality is run whenever the system changes significantly. Locked systems are largely deterministic, but large neural networks can still behave non-deterministically because of floating-point limits and GPU parallelism. An updated locked system is typically considered a new system requiring a fresh round of testing. Adaptive systems can be rigorously tested before deployment, for example by simulating environmental changes.

### Question 7 of 10

A company plans to red team its customer-service LLM chatbot before launch. Which sequence of activities is consistent with the red teaming approach described in the CT-AI syllabus?

- A. Assemble a diverse team, give it access in a safe test environment, prompt for vulnerabilities, analyze the failures, and build datasets for mitigations.
- B. Run a standard public benchmark suite, compare the scores with competitor models, and publish the ranking as evidence of safety.
- C. Write a fixed set of scripted prompts with expected answers, execute them once against the production system, and report the pass rate to the product owner.
- D. Give the chatbot to real customers, collect complaints for a month, analyze them, and then assemble a team to fix the most frequent defects.

**Answer:** A

**Explanation:** The red teaming approach involves assembling a diverse team, providing access in a safe test environment, prompting the system through open-ended exploration or checklists, analyzing the identified failures, and creating datasets from the threats to support mitigations; it is most effective before deployment. Collecting customer complaints after launch is operational monitoring, not proactive red teaming. Benchmarking is a static approach that red teaming complements rather than replaces. Red teaming is adaptive, with prompts updated in response to outputs, unlike a fixed scripted suite run in production.

### Question 8 of 10

A tester wants to determine whether a loan approval model treats applicants differently because of their gender. Which activity applies disparate impact analysis?

- A. Measure the model's overall accuracy on the hold-out test set and compare it with the acceptance criterion of the project.
- B. Create counterfactual applications with gender swapped, run both versions, and test the outcome differences for significance.
- C. Remove the gender attribute from the training data, retrain the model, and compare the new overall accuracy with the old one.
- D. Check that the training dataset contains equal numbers of male and female applicants before the model is trained again.

**Answer:** B

**Explanation:** Disparate impact analysis identifies sensitive attributes, generates counterfactuals in which those attributes are changed, presents them to the model, and analyzes many results to determine statistically whether changing the attribute changes the outcome. Removing the attribute and retraining is a mitigation attempt that does not test the existing model and can leave proxy attributes in place. Checking class counts is a representativeness check, not a comparison of model decisions. Overall accuracy does not reveal differences in treatment between sensitive groups.

### Question 9 of 10 (choose 2)

A security tester is planning adversarial testing of a traffic-sign classifier. Which TWO statements about adversarial testing are consistent with the CT-AI syllabus? (Choose two.)

- A. Adversarial examples are often slightly modified legitimate inputs whose changes are imperceptible to humans.
- B. Adversarial examples always result from malicious attacks and are never encountered by accident in operation.
- C. Adversarial testing can only be performed as white-box testing because it needs access to the model's gradients.
- D. Knowing a model's architecture and parameters typically makes it easier to craft effective adversarial examples.
- E. Generating effective adversarial inputs is simple, and the attack techniques have stopped evolving in recent years.

**Answer:** A, D

**Explanation:** Adversarial testing provides perturbations to inputs that are often imperceptible to humans, so adversarial examples are frequently slightly modified legitimate inputs, and white-box knowledge of the model's internals typically makes crafting them easier. Black-box approaches such as surrogate models or brute force also exist, so white-box access is not required. Adversarial examples may be encountered accidentally as well as used in attacks. The syllabus states that generating effective inputs is technically complex and keeping up with evolving attack techniques is an ongoing challenge.

### Question 10 of 10

A loan approval model has no reliable oracle for individual decisions. A source test case for an applicant was approved and confirmed as correct. Which follow-up test case applies a valid metamorphic relation?

- A. Increase only the applicant's declared income and expect the decision to remain an approval
- B. Change the applicant's postcode to a different region and expect exactly the same approval probability
- C. Reduce the applicant's income and expect the decision to change to a rejection in every case
- D. Double both income and existing debt and expect the approval probability to remain exactly unchanged

**Answer:** A

**Explanation:** Increasing income while all other attributes stay the same should not turn an approval into a rejection, a monotonicity relation derived from domain knowledge that checks the relationship between outputs. Expecting an identical probability after a location change is not a domain property, since location may legitimately influence risk. A reduction in income need not always cause rejection, so that relation is too strong. Changing income and debt together overlooks interactions between variables, which the syllabus warns produces incorrect relations.

Ready for a timed run? Take the free 15-question CT-AI mini exam: https://devopscodelab.com/app/exams/istqb-ct-ai-v2-0/mini/

DevOpsCode Lab is not affiliated with ISTQB. CT-AI and related names are trademarks of their owners. We write every question ourselves; these are not real exam questions.
