CT-AI Certified Tester AI Testing Exam Question and Answers

A. Regression

Regression is used to predict a continuous value and is not suitable for grouping people based on similar tastes.

B. Association

Association is used to find relationships between variables in large datasets, often in the form of rules (e.g., market basket analysis). It does not directly group individuals but identifies patterns of co-occurrence.

C. Clustering

Clustering is an unsupervised learning method used to group similar data points based on their features. It is ideal for identifying groups of people with similar tastes without prior knowledge of the group labels. This technique will help the company segment its customer base effectively.

D. Classification

Classification is a supervised learning method used to categorize data points into predefined classes. It requires labeled data for training, which is not the case here as we want to identify groups without predefined labels.

Therefore, the correct answer isCbecause clustering is the most suitable method for grouping people with similar tastes for targeted product marketing​​.

The process of developing a machine learning model typically involves the use of three types of datasets:

Training Data:This is used to train the model, i.e., to learn the patterns and relationships in the data.

Validation Data:This is used to tune the model’s hyperparameters and to prevent overfitting during the training process.

Test Data:This is used to evaluate the final model’s performance and to estimate how it will perform on unseen data.

Let's analyze each option:

A. Training data - validation data - test data

This option correctly includes all three types of datasets used in the process of creating and validating a model. The training data is used for learning, validation data for tuning, and test data for final evaluation.

B. Training data - validation data

This option misses the test data, which is crucial for evaluating the model’s performance on unseen data after the training and validation phases.

C. Training data - test data

This option misses the validation data, which is important for tuning the model and preventing overfitting during training.

D. Validation data - test data

This option misses the training data, which is essential for the initial learning phase of the model.

Therefore, the correct answer isAbecause it includes all necessary datasets used during the process of learning and creating the model: training, validation, and test data​​.

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.

References:

ISTQB CT-AI Syllabus, Section 6.1, Neural Networks, discusses the components and functioning of neurons in a neural network.

"Neural Network Activation Functions" (ISTQB CT-AI Syllabus, Section 6.1.1).

Automation bias, also known as complacency bias, occurs when humans over-rely on automated systems and fail to question or validate the system’s output. In this scenario, the auto-text correction feature of the word processing tool incorrectly suggests "Isle of Eight" instead of "Isle of Wight." The issue arises because multiple UAT testers accept the incorrect suggestion without noticing it, demonstrating a reliance on the AI-based system rather than their own judgment.

Automation bias is commonly seen in:

Text correction systems, where users accept incorrect suggestions without verifying them.

Medical diagnosis AI tools, where doctors may rely too much on AI recommendations.

Autonomous driving systems, where drivers become overly dependent on automation and fail to react in critical situations.

Section 7.4 - Testing for Automation Bias in AI-Based Systemsexplains that automation bias occurs when people accept AI-generated outputs without verifying them, often leading to incorrect decisions​.

Reference from ISTQB Certified Tester AI Testing Study Guide:

Themachine learning (ML) workflowfollows a structured sequence of steps. Once stakeholders have agreed on theobjectives, business priorities, and the framework/model selection, the next logical step is toprepare and pre-process the databefore training the model.

Data Preparationis crucial becausemachine learning models rely heavily on the quality of input data. Poor data can result in biased, inaccurate, or unreliable models​.

The process involvesdata acquisition, cleaning, transformation, augmentation, and feature engineering​.

Preparing the dataensures it is in the right format, free from errors, and representative of the problem domain, leading to better generalization in training.

A (Tune the ML Algorithm):Hyperparameter tuning occursafter the model has been trainedand evaluated​.

C (Agree on Acceptance Criteria):Acceptance criteria should already have been defined in theinitial objective-setting phasebefore framework and model selection​.

D (Evaluate the Framework and Model):The selection of the framework and ML model has already been completed. The next step isdata preparation, not reevaluation​.

ISTQB CT-AI Syllabus (Section 3.2: ML Workflow - Data Preparation Phase)

"Data preparation comprises data acquisition, pre-processing, and feature engineering. Exploratory data analysis (EDA) may be performed alongside these activities"​.

"The data used to train, tune, and test the model must be representative of the operational data that will be used by the model"​.

Why Other Options Are Incorrect:Supporting References from ISTQB Certified Tester AI Testing Study Guide:Conclusion:Since the model selection is complete, thenext step in the ML workflow is to prepare and pre-process the datato ensure it is ready for training and testing. Thus, thecorrect answer is B.

AI-based systems oftenexhibit non-deterministic behavior, meaning theydo not always produce the same output for the same input. This makesensuring safety more difficult, as the system's behavior can change based on new data, environmental factors, or updates​.

Why Non-determinism Affects Safety:

In traditional software, the same input always produces the same output.

In AI systems, outputsvary probabilisticallydepending on learned patterns and weights.

This unpredictability makes itharder to verify correctness, reliability, and safety, especially in critical domains likeautonomous vehicles, medical AI, and industrial automation.

A (Simplicity):AI-based systems are typicallycomplex, not simple, which contributes to safety challenges​.

B (Sustainability):While sustainability is an important AI consideration, it doesnot directly affect safety​.

D (Robustness):Lack of robustnesscan make AI systems unsafe, butnon-determinism is the primary issuethat complicates safety verification​.

ISTQB CT-AI Syllabus (Section 2.8: Safety and AI)

"The characteristics of AI-based systems that make it more difficult to ensure they are safe include: complexity, non-determinism, probabilistic nature, self-learning, lack of transparency, interpretability and explainability, lack of robustness"​.

Why Other Options Are Incorrect:Supporting References from ISTQB Certified Tester AI Testing Study Guide:Conclusion:Sincenon-determinism makes AI behavior unpredictable, complicating safety assurance, thecorrect answer is C.

AI-based systems have specific quality characteristics, includingevolution,autonomy, andadaptability. A test objective that evaluates whether an AI systemevolvesto improve performance over time directly aligns with AI quality characteristics​.

Explanation of Answer Choices:

Option A: The robot must evolve to optimize its routing.

Correct.Evolution is an AI quality characteristic that ensures the systemlearns from past experiencesand adapts to improve efficiency​.

Option B: The robot must recharge for no more than six hours a day.

Incorrect.This is an operational constraint rather than an AI-specific quality characteristic​.

Option C: The robot must record the time of each delivery which is compiled into a report.

Incorrect.Logging data does not relate to AI quality characteristics likeadaptability or autonomy​.

Option D: The robot must complete 99.99% of its deliveries each day.

Incorrect.This is a performance target rather than an AI quality characteristic​.

ISTQB CT-AI Syllabus References:

Evolution as an AI Quality Characteristic:"Check how well the system learns from its own experience. Check how well the system copes when the profile of data changes (i.e., concept drift)"​.

Thus,Option A is the best choice as it directly tests an AI quality characteristic (evolution) in the upgraded autonomous robot.

AI test environments have several unique characteristics that differentiate them from traditional test environments. Let’s evaluate each option:

A. Challenges resulting from low accuracy of the models.

Low accuracy is a common challenge in AI systems, especially during initial development and training phases. Ensuring the model performs accurately in varied and unpredictable scenarios is a critical aspect of AI testing.

B. The challenge of mimicking undefined scenarios generated due to self-learning.

AI systems, particularly those that involve machine learning, can generate undefined or unexpected scenarios due to their self-learning capabilities. Mimicking and testing these scenarios is a unique challenge in AI environments.

C. The challenge of providing explainability to the decisions made by the system.

Explainability, or the ability to understand and articulate how an AI system arrives at its decisions, is a significant and unique challenge in AI testing. This is crucial for trust and transparency in AI systems.

D. Challenges in the creation of scenarios of human handover for autonomous systems.

While important, the creation of scenarios for human handover in autonomous systems is not a characteristic unique to AI test environments. It is more related to the operational and deployment challenges of autonomous systems rather than the intrinsic technology-related characteristics of AI.

Given the above points, optionDis the correct answer because it describes a challenge related to operational deployment rather than a technology-related characteristic unique to AI test environments.

Bias in an AI system occurs when the training data contains inherent prejudices that cause the model to make unfair predictions. Experience-based testing, particularlyExploratory Data Analysis (EDA), helps uncover these biases by analyzing patterns, distributions, and potential discriminatory factors in the training data​.

Option A:“Experience-based testing should be used to confirm that the training data set is operationally relevant. This can include applying exploratory data analysis (EDA) to check for bias within the training data set.”

This is the correct answer. EDA involves examining the dataset for bias, inconsistencies, or missing values, ensuring fairness in ML model predictions​.

Option B:“Back-to-back testing should be used to compare the model created using the training data set to another model created using the test data set. If the two models significantly differ, it will indicate there is bias in the original model.”

Back-to-back testing is used for regression testing and to compare versions of an AI system but is not primarily used to detect bias​.

Option C:“Acceptance testing should be used to make sure the algorithm is suitable for the customer. The team can re-work the acceptance criteria such that the algorithm is sure to correctly predict the remaining applicants that have been set aside for the validation data set ensuring no bias is present.”

Acceptance testing focuses on meeting predefined business requirements rather than detecting and mitigating bias​.

Option D:“A/B testing should be used to verify that the test data set does not detect any bias that might have been introduced by the original training data. If the two models significantly differ, it will indicate there is bias in the original model.”

A/B testing is used for evaluating variations of a model rather than for explicitly identifying bias​.

Bias Testing Methods:"AI-based systems should be tested for algorithmic bias, sample bias, and inappropriate bias. Experience-based testing and EDA are useful for detecting bias"​.

Exploratory Data Analysis (EDA):"EDA helps uncover potential bias in training data through visualization and statistical analysis"​.

Analysis of the Answer Options:ISTQB CT-AI Syllabus References:Thus,Option A is the best choice for detecting bias in the loan applicant model.

A. This is an example of expert system bias.

Expert system bias refers to bias introduced by the rules or logic defined by experts in the system, not by the data distribution.

B. This is an example of sample bias.

Sample bias occurs when the training data is not representative of the overall population that the model will encounter in practice. In this case, the over-representation of ethnicity A (70%) compared to B, C, and D (30%) creates a sample bias, as the model may become biased towards better performance on ethnicity A.

C. This is an example of hyperparameter bias.

Hyperparameter bias relates to the settings and configurations used during the training process, not the data distribution itself.

D. This is an example of algorithmic bias.

Algorithmic bias refers to biases introduced by the algorithmic processes and decision-making rules, not directly by the distribution of training data.

Based on the provided information, optionB(sample bias) best describes the situation because the training data is skewed towards ethnicity A, potentially leading to biased model performance​​.

Data mislabeling occurs for several reasons, which can significantly impact the performance of machine learning (ML) models, especially in supervised learning. According to the ISTQB Certified Tester AI Testing (CT-AI) syllabus, mislabeling of data can be caused by the following factors:

Random errors by annotators– Mistakes made due to accidental misclassification.

Systemic errors– Errors introduced by incorrect labeling instructions or poor training of annotators.

Deliberate errors– Errors introduced intentionally by malicious data annotators.

Translation errors– Occur when correctly labeled data in one language is incorrectly translated into another language.

Subjectivity in labeling– Some labeling tasks require subjective judgment, leading to inconsistencies between different annotators.

Lack of domain knowledge– If annotators do not have sufficient expertise in the domain, they may label data incorrectly due to misunderstanding the context.

Complex classification tasks– The more complex the task, the higher the probability of labeling mistakes​.

Among the answer choices provided, "Lack of domain knowledge" (Option A) is the best answer because expertise is essential to accurately labeling data in complex domains such as medical, legal, or engineering fields.

Certified Tester AI Testing Study Guide References:

ISTQB CT-AI Syllabus v1.0, Section 4.5.2 (Mislabeled Data in Datasets)

ISTQB CT-AI Syllabus v1.0, Section 4.3 (Dataset Quality Issues)

Bias in AI models often originates fromincomplete or non-representative training data. In this case, if the training datasetpurposely excludes specific faulty conditions, the machine learning model willfail to learn and detectthese conditions in real-world scenarios.

This results in:

Sample bias, where the training data is not fully representative of all possible faulty conditions.

Algorithmic bias, where the model prioritizes certain defect types while ignoring others.

B. Selecting training data by purposely including all known faulty conditions→ This would help reduce bias by improving model generalization.

C. Selecting testing data from a different dataset than the training dataset→ This is a good practice to evaluate model generalization but does not inherently introduce bias.

D. Selecting testing data from a boat manufacturer's bolt longevity data→ While using unrelated data can createpoor model accuracy, it does not directly introduce bias unless systematic patterns in the incorrect dataset lead to unfair decision-making.

Section 8.3 - Testing for Algorithmic, Sample, and Inappropriate Biasstates thatsample bias can occur if the training dataset is not fully representative of the expected data space, leading to biased predictions​.

Why are the other options incorrect?Reference from ISTQB Certified Tester AI Testing Study Guide:

Setting model and algorithm hyperparameters is an essential step in the machine learning workflow, primarily occurring during the tuning phase.

Evaluating the model (A): This stage involves assessing the model's performance using metrics and does not typically include the setting of hyperparameters.

Deploying the model (B): Deployment is the stage where the model is put into production and used in real-world applications. Hyperparameters should already be set before this stage.

Tuning the model (C): This is the correct stage where hyperparameters are set. Tuning involves adjusting the hyperparameters to optimize the model's performance.

Data testing (D): Data testing involves ensuring the quality and integrity of the data used for training and testing the model. It does not include setting hyperparameters.

Hence, the most appropriate stage of the ML workflow to set model and algorithm hyperparameters isC. Tuning the model.

References:

ISTQB CT-AI Syllabus Section 3.2 on the ML Workflow outlines the different stages of the ML process, including the tuning phase where hyperparameters are set.

Sample Exam Questions document, Question #31 specifically addresses the stage in the ML workflow where hyperparameters are configured.

Narrow AI (also known as Weak AI) is designed to perform specific tasks without possessing general intelligence or consciousness. The AI system in the question is capable of recognizing speech and responding to specific booking-related requests but fails when asked about unrelated topics (such as weather or labor markets)​.

Option A:“General AI”

Incorrect. General AI (AGI) refers to an AI system that can perform any intellectual task a human can. The described system is task-specific and does not exhibit general intelligence​.

Option B:“Narrow AI”

Correct. The AI system is limited to a predefined domain (ticket booking) and cannot process unrelated questions. This is characteristic of Narrow AI, which excels at specific tasks but lacks broader cognitive abilities​.

Option C:“Super AI”

Incorrect. Super AI surpasses human intelligence, exhibiting advanced reasoning and creativity. The AI in the scenario is far from this level​.

Option D:“Conventional AI”

Incorrect. Conventional AI is a broader term that may include rule-based systems. The described system relies on machine learning and natural language processing, making it more aligned with Narrow AI​.

Definition of Narrow AI:"Narrow AI refers to AI systems that are designed to perform a single task or a limited set of tasks, without general intelligence"​.

General vs. Narrow AI:"General AI remains an area of research, while most current AI applications fall into the category of Narrow AI"​.

Analysis of the Answer Options:ISTQB CT-AI Syllabus References:Thus,option B is the correct categorization for the AI-based ticket booking system.

ï‚· Labelling Approaches:Among the options provided, pre-labeled datasets require the least time and effort because the data has already been labeled, eliminating the need for further manual or automated labeling efforts.

 Reference:ISTQB_CT-AI_Syllabus_v1.0, Section 4.5 Data Labelling for Supervised Learning, which discusses various approaches to data labeling, including pre-labeled datasets, and their associated time and effort requirements​​.

ï‚· Type of Testing for Stock-Price Prediction Models:Concept drift refers to the change in the statistical properties of the target variable over time. Severe weather storms causing massive losses in coffee production and affecting stock prices would require testing for concept drift to ensure that the model adapts to new patterns in data over time.

 Reference:ISTQB_CT-AI_Syllabus_v1.0, Section 7.6 Testing for Concept Drift, which explains the need to test for concept drift in models that might be affected by changing external factors​​.

Pre-processing is an essential step in data preparation that ensures data is clean, formatted correctly, and structured for effective machine learning (ML) model training. One common issue that can be resolved during pre-processing isnumbers stored as strings.

Explanation of Answer Choices:

Option A: Insufficient data

Incorrect. Pre-processing cannot resolve insufficient data. If data is lacking, techniques like data augmentation or external data collection are needed​.

Option B: Invalid data

Incorrect. While pre-processing can identify and handle some forms of invalid data (e.g., missing values, duplicate entries), it does not resolve all invalid data issues. Some cases may require domain expertise to determine validity​.

Option C: Wanted outliers

Incorrect. Pre-processing usually focuses on handling unwanted outliers. Wanted outliers may need to be preserved, which is more of a data selection decision rather than pre-processing​.

Option D: Numbers stored as strings

Correct. One of the key functions of data pre-processing isdata transformation, which includes converting incorrectly formatted data types, such as numbers stored as strings, into their correct numerical format​.

ISTQB CT-AI Syllabus References:

Data Pre-Processing Steps:"Transformation: The format of the given data is changed (e.g., breaking an address held as a string into its constituent parts, dropping a field holding a random identifier, converting categorical data into numerical data, changing image formats)"​.

A. Add 10% of the rows randomly and create another model and compare the R-Square scores of both the models.

Adding more data to the training set can affect the R-Square score, but it does not directly verify the correctness of the implementation.

B. Train various models by changing the order of input features and verify that the R-Square score of these models vary significantly.

Changing the order of input features should not significantly affect the R-Square score if the implementation is correct, but this approach is more about testing model robustness rather than correctness of the implementation.

C. Compare the R-Square score of the model obtained using two different implementations that utilize two different programming languages while using the same algorithm and the same training and testing data.

This approach directly compares the performance of two implementations of the same algorithm. If both implementations produce similar R-Square scores on the same training and testing data, it suggests that the new implementation "X" is correct.

D. Drop 10% of the rows randomly and create another model and compare the R-Square scores of both the models.

Dropping data can lead to variations in the R-Square score but does not directly verify the correctness of the implementation.

Therefore, optionCis the most appropriate strategy because it directly compares the performance of the new implementation "X" with another implementation using the same algorithm and datasets, which helps in verifying the correctness of the implementation​​.

To calculate the accuracy, recall, and specificity from the confusion matrix provided, we use the following formulas:

Confusion Matrix:

Actually Rotten: 45 (True Positive), 8 (False Positive)

Actually Fresh: 5 (False Negative), 42 (True Negative)

Accuracy:

Accuracy is the proportion of true results (both true positives and true negatives) in the total population.

Formula: Accuracy=TP+TNTP+TN+FP+FN\text{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN}Accuracy=TP+TN+FP+FNTP+TN​

Calculation: Accuracy=45+4245+42+8+5=87100=0.87\text{Accuracy} = \frac{45 + 42}{45 + 42 + 8 + 5} = \frac{87}{100} = 0.87Accuracy=45+42+8+545+42​=10087​=0.87

Recall (Sensitivity):

Recall is the proportion of true positive results in the total actual positives.

Formula: Recall=TPTP+FN\text{Recall} = \frac{TP}{TP + FN}Recall=TP+FNTP​

Calculation: Recall=4545+5=4550=0.9\text{Recall} = \frac{45}{45 + 5} = \frac{45}{50} = 0.9Recall=45+545​=5045​=0.9

Specificity:

Specificity is the proportion of true negative results in the total actual negatives.

Formula: Specificity=TNTN+FP\text{Specificity} = \frac{TN}{TN + FP}Specificity=TN+FPTN​

Calculation: Specificity=4242+8=4250=0.84\text{Specificity} = \frac{42}{42 + 8} = \frac{42}{50} = 0.84Specificity=42+842​=5042​=0.84

Therefore, the correct combinations of accuracy, recall, and specificity are 0.87, 0.9, and 0.84 respectively.

References:

ISTQB CT-AI Syllabus, Section 5.1, Confusion Matrix, provides detailed formulas and explanations for calculating various metrics including accuracy, recall, and specificity.

"ML Functional Performance Metrics" (ISTQB CT-AI Syllabus, Section 5).

The problem described in the question is a classic case ofconcept drift. Concept drift occurs when the relationship between input variables and the output variable changes over time, leading to a decline in model accuracy​.

In this scenario, theaverage passenger and baggage weightsused in the model changed due to seasonal variations, but the model was not updated accordingly. This resulted in inaccurate predictions for fuel requirements in the winter season. This is an example ofseasonal drift, where model behavior changes periodically due to recurring trends (e.g., higher luggage weights in winter compared to summer).

To prevent such problems:

Themodel should be regularly testedfor concept drift against agreed ML functional performance criteria​.

Exploratory Data Analysis (EDA)should be performed periodically to detect gradual changes in input distributions.

Retraining of the modelwith updated training data should be done to maintain accuracy.

If drift is detected, mitigation techniques such asincremental learning, retraining with new data, or adjusting model parametersshould be employed.

Option B (Easing the performance standard instead of addressing drift): Lowering the performance standard is not a solution; it only masks the problem without fixing it. Instead, regular testing and retraining should be used to handle drift properly​.

Option C (Corruption and reloading the model): Model corruption is unrelated to this issue. Corruption refers to accidental or malicious damage to the model or data, whereas this case is due to a changing data environment​.

Option D (Lack of transparency): Transparency refers to how understandable the model’s decisions are, but the problem here is a change in data distributions, making drift the primary concern​.

ISTQB CT-AI Syllabus (Section 7.6: Testing for Concept Drift)

"The operational environment can change over time without the trained model changing correspondingly. This phenomenon is known as concept drift and typically causes the outputs of the model to become increasingly less accurate and less useful."​

"Systems that may be prone to concept drift should be regularly tested against their agreed ML functional performance criteria to ensure that any occurrences of concept drift are detected soon enough for the problem to be mitigated."​

ISTQB CT-AI Syllabus (Section 7.7: Selecting a Test Approach for an ML System)

"If concept drift is detected, it may be mitigated by retraining the system with up-to-date training data followed by confirmation testing, regression testing, and possibly A/B testing where the updated system must outperform the original system."​

Why Other Options Are Incorrect:Supporting References from ISTQB Certified Tester AI Testing Study Guide:Conclusion:Since the question describes a situation whereseasonal variations affected input data distributions, the correct answer isA: The model suffers from drift and therefore should be regularly tested to ensure that any occurrences of drift are detected soon enough for the problem to be mitigated.

Whentesting an AI-based chat assistantfor e-commerce customers, thelack of sufficient specificationsmakes it difficult to use structured test techniques. TheISTQB CT-AI Syllabusrecommendsexploratory testingin such cases:

Why Exploratory Testing?

Exploratory testing is usefulwhen specifications are incomplete or unclear.

AI-based systems, particularly those usingnatural language processing (NLP),may not behave deterministically, making scripted test cases ineffective.

Thetester interacts dynamicallywith the system, identifying unexpected behaviorsnot documented in the specification​.

Analysis of Answer Choices:

A (Exploratory testing)→Correct, as it is the best approach when specifications are incomplete.

B (Static analysis)→ Incorrect, as static analysis checks code without execution, which isnot helpfulfor AI chatbots.

C (Equivalence partitioning)→ Incorrect, asthis technique requires well-defined inputs and outputs, which are missing due toinsufficient specifications.

D (State transition testing)→ Incorrect, as state-based testingrequires knowledge of valid and invalid transitions, which is difficult with a chatbot lacking a clear specification​.

Thus,Option A is the correct answer, asexploratory testing is the best approach when dealing with insufficient specifications in AI-based systems.

Certified Tester AI Testing Study Guide References:

ISTQB CT-AI Syllabus v1.0, Section 7.7 (Selecting a Test Approach for an ML System)

ISTQB CT-AI Syllabus v1.0, Section 9.6 (Experience-Based Testing of AI-Based Systems)​.