CT-AI Questions and Answers

Section4.3 – Test Levels for AI Systemsclearly defines ML model testing as the level at which testers evaluate whether an ML model fulfills itsfunctional performance criteria, including accuracy, precision, recall, F1, robustness, stability, and fairness. Therefore, Option C is the correct and syllabus-aligned statement.

Option A is incorrect because input data testing focuses onvalidity and correctness of data entering the model, not interactions with all system components. Option B is incorrect: acceptance testing in the syllabus focuses primarily onbusiness and stakeholder requirements, not specifically explainability. Explainability testing may occur at multiple levels depending on context. Option D is also incorrect because API testing belongs tointegration testing, not system testing, even when AI is consumed as a service.

Thus,Option Cis the only statement that precisely matches syllabus definitions.

The ISTQB CT-AI syllabus (Section3.2 – Model Evaluation) specifies the correct usage oftraining,validation, andtestdatasets. It emphasizes that thetest dataset must be representative of the real operational dataand must beequivalent in distribution to the training and validation sets, ensuring a fair and unbiased evaluation. Option D precisely matches this requirement.

Option A contradicts the syllabus because validation and test sets servedifferent purposes: validation is for tuning, test is for final evaluation. Combining them undermines the reliability of results. Option B is incorrect because even with limited data, the syllabus recommends maintaining a test set or using techniques such ascross-validationrather than eliminating testing. Option C is wrong because equal distribution (33/33/33) isnot recommended; typically, the training set is much larger (e.g., 70–80%).

Thus, OptionDis the only statement aligned with the syllabus' guidance.

According to the syllabus,conventional AIsystems are limited to specific, pre-defined tasks and do not have generalized intelligence:

"Conventional AI systems are limited in their scope and typically only perform specific tasks within the domain for which they have been designed. They do not exhibit general AI behavior."

(Reference: ISTQB CT-AI Syllabus v1.0, Section 1.2)

Section4.2 – Test Coverage Criteria for AI Modelsof the ISTQB CT-AI syllabus describes neural network–specific coverage methods. Among the techniques,threshold coverageis explicitly noted asadaptable, meaning testers may choose different thresholds to determine whether neuron activation is considered “covered.” This flexibility makes threshold coverage adjustable to the model architecture, problem domain, and required test thoroughness.

Options A and B (Sign–Sign and Sign–Change coverage) are more rigid structural criteria and are not described as adaptable within the syllabus. They focus on sign patterns of neuron activations and do not allow altering thresholds. Option D, neuron coverage, measures the proportion of neurons activated at least once. Although simple, it is not defined as an adaptable criterion. Its limitations are documented: it provides shallow insight and too easily achieves high coverage.

Onlythreshold coverageallows testers to adjust activation thresholds for more refined coverage measurement, makingOption Cthe correct choice.

A. Feature selection

Feature selection is the process of selecting the most relevant features from the data. While important, it is not directly about handling excess data.

B. Data sampling

Data sampling involves selecting a representative subset of the data for training. When there is more data than needed, sampling can be used to create a manageable dataset that maintains the statistical properties of the full dataset.

C. Data labeling

Data labeling involves annotating data for supervised learning. It is necessary for training models but does not address the issue of having excess data.

D. Data augmentation

Data augmentation is used to increase the size of the training dataset by creating modified versions of existing data. It is useful when there is insufficient data, not when there is excess data.

Therefore, the correct answer isBbecause data sampling is the most relevant activity when dealing with an excess amount of data for training​​.

Section1.6.3 – Reinforcement Learningof the ISTQB CT-AI syllabus states that reinforcement learning (RL) is based on anagent interacting with an environment, performing actions, and receivingrewards or penalties. The core concept is thereward function, which guides the agent’s learning process. The syllabus emphasizes that training in RL isdriven by rewards, and the agent aims to maximize cumulative reward over time. Therefore, OptionCdirectly reflects the correct description: the agent learns by being rewarded for successful actions .

Option A is incorrect because RL doesnotuse labeled data; that applies to supervised learning. Option B contradicts the syllabus definition: RL fundamentallyrequiresinteraction with the environment. Option D is incorrect because the reward function isdefined by humans, not learned by the agent; the agent learns apolicy, not the reward function itself.

Thus, OptionCis the only statement consistent with RL as defined in the syllabus.

The ISTQB CT-AI syllabus (Section5.3 – AI Support for Defect Analysis) explains that AI can categorize defect reports using natural language processing or classification models. Categorization helps route defects efficiently and determine which areas of the system are affected. Thus, OptionCis correct: AI canidentify defect categories, supporting assignment and triage.

Option A is incorrect because ML cannot infer severity or consequence without domain-specific risk modeling. Option B inaccurately ties duplicate detection to criticality; the syllabus separatesduplicate detectionfromseverity analysis. Option D is incorrect because AI may assist in routing defects using categories, but the syllabus does not claim AI identifies specific developers.

Thus,Option Cis the syllabus-accurate statement.

The syllabus explains that ML models can be used to analyze reported defects and suggest which developers are best suited to fix them based on historical data about defect assignment and resolution speed:

"Assignment: ML models can suggest which developers are best suited to fix particular defects, based on the defect content and previous developer assignments."

(Reference: ISTQB CT-AI Syllabus v1.0, Section 11.2, page 78 of 99)

The question asks which test is least likely to be performed during the ML model testing phase. Let's consider each option:

Testing the accuracy of the classification model (A): Accuracy testing is a fundamental part of the ML model testing phase. It ensures that the model correctly classifies the data as intended and meets the required performance metrics.

Testing the API of the service powered by the ML model (B): Testing the API is crucial, especially if the ML model is deployed as part of a service. This ensures that the service integrates well with other systems and that the API performs as expected.

Testing the speed of the training of the model (C): This is least likely to be part of the ML model testing phase. The speed of training is more relevant during the development phase when optimizing and tuning the model. During testing, the focus is more on the model's performance and behavior rather than how quickly it was trained.

Testing the speed of the prediction by the model (D): Testing the speed of prediction is important to ensure that the model meets performance requirements in a production environment, especially for real-time applications.

The syllabus mentions that experience-based testing and EDA are effective for detecting biases:

“Experience-based testing can be used to verify that the training dataset is operationally relevant and identify potential sources of bias. EDA is also useful for exploring the data and understanding any relationships that might lead to bias in the model.”

(Reference: ISTQB CT-AI Syllabus v1.0, Section 8.3, page 58 of 99)

Section2.4 – Data Labeling Approachesof the ISTQB CT-AI syllabus explicitly definesAI-assisted data labelingas a hybrid process in which an automated tool performs the initial labeling and human annotators subsequently verify, correct, or complete the labels. This two-step process improves efficiency while retaining human oversight to ensure data quality. The syllabus describes this method as an effective compromise when manual labeling alone would be too slow or costly, and when initial automation can identify obvious patterns before a human provides the final authoritative labels .

Option A (internal labeling) refers to labeling conducted by the organization’s own staff but does not imply automation. Option B (crowdsourced labeling) leverages a distributed workforce, typically without automation. Option C (outsourced labeling) transfers labeling tasks to external vendors but similarly does not involve an AI-first step. Only OptionDreflects the two-stage automated-then-human workflow described in the syllabus.

Therefore,AI-assisted data labeling(Option D) is unequivocally correct.

According to Section4.6 – AI Behaviour Testing (AIB Testing)of the ISTQB CT-AI syllabus, AIB testing is used to evaluate changes in the functional behavior of self-learning systems. The core principle iscomparing pre-change and post-change model behavior using the same test inputs, so that any difference in outputs can be attributed to the model’s learning and not to differences in input data. This directly corresponds to OptionC.

Option A is incorrect because the absence of a test oracle does not justify generating new test cases; AIB relies onreusing identical inputsto detect behavioral drift. Option B is invalid because using different inputs prevents meaningful comparison. Option D is incorrect because comparing with an unrelated non-self-learning system does not allow evaluation of the same model’s behavioral evolution.

Thus, OptionCaccurately represents the correct application of AIB testing: assessing model behavior changes by running identical test inputs before and after learning updates.

The question refers to a problem where data used for an object detection ML system was labelled incorrectly. This issue is most closely related to "accuracy issues." Here's a detailed explanation:

Accuracy Issues: The primary goal of labeling data in machine learning is to ensure that the model can accurately learn and make predictions based on the given labels. Incorrectly labeled data directly impacts the model's accuracy, leading to poor performance because the model learns incorrect patterns.

Why Not Other Options:

Security Issues: This pertains to data breaches or unauthorized access, which is not relevant to the problem of incorrect data labeling.

Privacy Issues: This concerns the protection of personal data and is not related to the accuracy of data labeling.

Bias Issues: While bias in data can affect model performance, it specifically refers to systematic errors or prejudices in the data rather than outright incorrect labeling.

From the confusion matrix:

True Positives (TP) = 15

False Positives (FP) = 5

False Negatives (FN) = 15

True Negatives (TN) = 65

Accuracy= (TP + TN) / Total

= (15 + 65) / 100

=80%

Precision= TP / (TP + FP)

= 15 / (15 + 5)

= 15 / 20

=75%

Section3.2 – Functional Performance Criteriain the syllabus explains accuracy and precision exactly these ways when evaluating ML classification performance.

Option B is therefore the only correct pair of values.

The syllabus defines overfitting as:

“Overfitting is when the ML model learns the training data so well that it is unable to generalize to accommodate new data.”

This occurs when the model memorizes the training data, including noise, instead of learning the general patterns.

(Reference: ISTQB CT-AI Syllabus v1.0, Section 3.5.1, page 31 of 99)

The question asks which deficiency is most likely to be discovered by the test expert given the scenario of poor real-world performance despite good isolated accuracy.

A lack of similarity between the training and testing data (A): This is a common issue in ML where the model performs well on training data but poorly on real-world data due to a lack of representativeness in the training data. This leads to poor generalization to new, unseen data.

The input data has not been tested for quality prior to use for testing (B): While data quality is important, this option is less likely to be the primary reason for the described issue compared to the representativeness of training data.

A lack of focus on choosing the right functional-performance metrics (C): Proper metrics are crucial, but the issue described seems more related to the data mismatch rather than metric selection.

A lack of focus on non-functional requirements testing (D): Non-functional requirements are important, but the scenario specifically mentions issues with detecting real cancer cases, pointing more towards data issues.

 Testing Autonomy:Testing for human handover when it should not be relinquishing control is the least appropriate because it contradicts the very definition of autonomous systems. The other tests are relevant to ensuring smooth operation and transitions between human and AI control.

 Reference:ISTQB_CT-AI_Syllabus_v1.0, Sections on Testing Autonomous AI-Based Systems and Testing for Human-AI Interaction​​​​. ​

A. Black box attacks based on adversarial examples create an exact duplicate model of the original.

Black box attacks do not create an exact duplicate model. Instead, they exploit the model by querying it and using the outputs to craft adversarial examples without knowledge of the internal workings.

B. These attack examples cause a model to predict the correct class with slightly less accuracy even though they look like the original image.

Adversarial examples typically cause the model to predict the incorrect class rather than just reducing accuracy. These examples are designed to be visually indistinguishable from the original image but lead to incorrect classifications.

C. These attacks can't be prevented by retraining the model with these examples augmented to the training data.

This statement is incorrect because retraining the model with adversarial examples included in the training data can help the model learn to resist such attacks, a technique known as adversarial training.

D. These examples are model specific and are not likely to cause another model trained on the same task to fail.

Adversarial examples are often model-specific, meaning that they exploit the specific weaknesses of a particular model. While some adversarial examples might transfer between models, many are tailored to the specific model they were generated for and may not affect other models trained on the same task.

Therefore, the correct answer isDbecause adversarial examples are typically model-specific and may not cause another model trained on the same task to fail​​.

The syllabus defines reward hacking as:

"Reward hacking can result from an AI-based system achieving a specified goal by using a 'clever' or 'easy' solution that perverts the spirit of the designer’s intent."

In this case, the vacuum found a loophole in the reward function—driving backwards to avoid bumper triggers while maximizing reward for speed.

(Reference: ISTQB CT-AI Syllabus v1.0, Section 2.6, page 24 of 99)

The syllabus defines data acquisition as consisting of three steps:

“Data acquisition: The activity of acquiring data relevant to the business problem to be solved by an ML model, typically involving the activities of identifying, gathering and labelling data.”

(Reference: ISTQB CT-AI Syllabus v1.0, Section 4.1, page 33 of 99)

In a neural network, the activation value of a neuron is determined by a combination of inputs from the previous layer, the weights of the connections, and the bias at the neuron level. Here’s a detailed breakdown:

Inputs for Activation Value:

Activation Values of Neurons in the Previous Layer:These are the outputs from neurons in the preceding layer that serve as inputs to the current neuron.

Weights Assigned to the Connections:Each connection between neurons has an associated weight, which determines the strength and direction of the input signal.

Individual Bias at the Neuron Level:Each neuron has a bias value that adjusts the input sum, allowing the activation function to be shifted.

Calculation:

The activation value is computed by summing the weighted inputs from the previous layer and adding the bias.

Formula: z=∑(wi⋅ai)+bz = \sum (w_i \cdot a_i) + bz=∑(wi​⋅ai​)+b, where wiw_iwi​ are the weights, aia_iai​ are the activation values from the previous layer, and bbb is the bias.

The activation function (e.g., sigmoid, ReLU) is then applied to this sum to get the final activation value.

Why Option A is Correct:

Option A correctly identifies all components involved in computing the activation value: the individual bias, the activation values of the previous layer, and the weights of the connections.

Eliminating Other Options:

B. Activation values of neurons in the previous layer, and weights assigned to the connections between the neurons: This option misses the bias, which is crucial.

C. Individual bias at the neuron level, and weights assigned to the connections between the neurons: This option misses the activation values from the previous layer.

D. Individual bias at the neuron level, and activation values of neurons in the previous layer: This option misses the weights, which are essential.

The ISTQB CT-AI syllabus emphasizes that testing AI-based systems requirescross-functional collaborationandexperience-based testingwhen parts of the team lack domain knowledge. In this scenario, the ML expert understands ML and dataset preparation but lacks knowledge ofcamera system behavior, the device’s operational data pipeline, and end-user workflows. The remainder of the team understands the domain and system testing but not ML. Section4.4 – Human Factors and AI Testingand4.3 – System Testing of AI Componentshighlight that when domain understanding is unevenly distributed,experience-based testing conducted by the full team(testers, developers, domain experts) is the most effective approach. This ensures that AI outputs align with actual user expectations and system behavior. OptionCaligns exactly with this principle.

Option A is too limited and does not address the need to validate ML integration. Option B is incorrect because reusing old test cases overlooks AI-specific risks in the operating data pipeline. Option D is useful but focuses only on data representativeness, not system-level user validation. Therefore,Option Cis the best, syllabus-aligned test approach.

 Coverage for Neuron Activation Values:Sign change coverage is used to ensure that test cases cause each neuron to achieve both positive and negative activation values. This type of coverage ensures that the neurons are thoroughly tested under different activation states.

 Reference:ISTQB_CT-AI_Syllabus_v1.0, Section 6.2 Coverage Measures for Neural Networks, which details different types of coverage measures, including sign change coverage​​.

Section5.3 – AI Support for Defect Predictionof the ISTQB CT-AI syllabus explains that AI-based defect prediction models rely onhistorical patterns, including past defects, code behavior, and similar system configurations. ML models trained on prior defect data can identifycomponents likely to contain defectswhen new changes resemble previous defect-inducing patterns. This directly supports OptionA, which states that defect prediction is most effective when based on previous similar constellations.

Option B is incorrect: ML can predictwhich componentsare likely to fail, not only whether defects exist. Option C is incomplete; code metrics help, but defect prediction relies onmanycontextual features (historical defects, code churn, commit frequency, etc.). Option D is wrong because defect prediction isnotbased on formal principles and typically requiresmany features, not just a few.

Thus,Option Ais the correct and syllabus-consistent answer.

 AI-Based Autonomous Functions:An AI-based autonomous system is one that can respond to its environment without human intervention. The other options either involve human decisions or do not use AI at all.

 Reference:ISTQB_CT-AI_Syllabus_v1.0, Sections on Autonomy and Testing Autonomous AI-Based Systems​​​​.

The ISTQB CT-AI syllabus states in Section2.10 – Explainability, Transparency, and Interpretabilitythat interpretability isuser-dependent, meaning different users understand explanations differently. This is because interpretability depends not only on the ML algorithm but also on the user’s domain knowledge, experience, and expectations. OptionBdirectly reflects this syllabus principle: interpretability testing must consideruser background, and explanations must be comprehensible to the intended user group.

Option A is incorrect because explainability testing requires substantial information about the model, data, and expected behavior—not just exploratory effort. Option C is incorrect because explainability isnotgenerally quantifiable through dynamic testing alone, and the syllabus does not assert model-type specificity in this way. Option D exaggerates LIME’s capabilities. LIME offers approximate local explanations, but cannotpreciselystate root causes; the syllabus emphasizes itslimitationsand that explanations areapproximations, not exact reasons.

Therefore,Option Bis the most syllabus-aligned and correct statement.

The syllabus defines back-to-back testing as a method to compare a modified AI system to the previous version, which is ideal in this scenario:

"Back-to-back testing is performed by comparing the outputs of two systems that are supposed to provide the same outputs, one being a known and trusted system and the other being the test system. This approach can be used to test ML systems after re-training to verify that improvements have not introduced regressions."

(Reference: ISTQB CT-AI Syllabus v1.0, Section 9.3, page 67 of 99)

The syllabus states:

“A/B testing can be used to test updates to an AI-based system where there are agreed acceptance criteria, such as ML functional performance metrics, as described in Chapter 5. A/B testing is used to compare the updated variant with the previous variant.”

(Reference: ISTQB CT-AI Syllabus v1.0, Section 9.4, page 68 of 99)

A. Identifying suitable tests by looking at the complexity of the test cases.

While complexity analysis can help in selecting important test cases, it does not directly address the issue of optimizing the entire regression suite effectively.

B. Using a random subset of tests.

Randomly selecting test cases may miss critical tests and does not ensure an optimized regression suite. This approach lacks a systematic method for ensuring comprehensive coverage.

C. Automating test scripts using AI-based test automation tools.

Automation helps in running tests efficiently but does not inherently optimize the selection of tests to prevent the suite from growing too large.

D. Using an AI-based tool to optimize the regression test suite by analyzing past test results.

This is the most effective approach as AI-based tools can analyze historical test data, identify patterns, and prioritize tests that are more likely to catch defects based on past results. This method ensures an optimized and manageable regression test suite by focusing on the most impactful test cases.

Therefore, the correct answer isDbecause using an AI-based tool to analyze past test results is the best option to optimize regression test selection and manage the size of the regression suite effectively​​.

The ISTQB CT-AI syllabus explains inSection 4.5 – Testing AI-Specific Risksthat adversarial testing is a structured test activity in which testers applyadversarial attacks—crafted or perturbed inputs—to intentionally expose weaknesses in the ML model. The purpose is to identify vulnerabilities that could be exploited throughdata poisoning,evasion attacks, orinput manipulation. OptionCcorrectly reflects this syllabus definition: adversarial testing is aboutusing attacks to locate weaknesses so they can be removed or mitigated.

Option A is incorrect because regression testing does not verify data sourcing policies; it verifies unchanged functionality after modifications. Option B is incorrect because adversarial examples are oftenaddedto training datasets to improverobustness(a practice called adversarial training), not excluded. Option D is incorrect because AIB testing is not described as superior to exploratory data analysis in outlier detection; both have different purposes, and EDA remains essential for data quality assessment.

Thus,Option Cis consistent with syllabus-defined adversarial testing.

The ISTQB CT-AI syllabus introducesAI-specific quality characteristics, includingevolution,functional safety,compatibility, andbias-related data quality. Section5.1 – AI-Specific Test Objectivesexplains thatevolutionrefers to an AI system’s capability to continue improving or at least maintain performance as it undergoes additional training. GPT_Legal is explicitly described as aself-learning systemexpected to:

continuously reduce false positives,

achieve weekly accuracy improvements of 10%,

reach and maintain 90% accuracy,

adapt to new environments (patent law firm → corporate legal department).

This aligns perfectly with the syllabus definition ofevidence of evolution: ensuring the model doesnot degradeas additional training data is introduced. OptionBtherefore directly supports the described acceptance criteria for this evolving, self-learning application.

Option A (functional safety) is irrelevant because patent searching and drafting do not constitute safety-critical domains. Option C (compatibility) is necessary but not the primary AI-specific objective. Option D addresses bias, which is important but not central to the described performance and continuous-learning expectations.

Thus,Option Bis the most appropriate AI-specific test objective.