ISTQB ISTQB-CT-AI - Questions & Answers

The ISTQB CT-AI syllabus clearly explains the structure of a deep neural network (DNN). A deep neural network consists of three types of layers: an input layer that receives the input data, one or more hidden layers composed of artificial neurons (also referred to as nodes), and an output layer that produces the final result. The defining characteristic of a deep neural network is the presence of at least one hidden layer between the input and output layers.

This directly supports Option B, which correctly states that a deep neural network has at least one internal hidden layer.

The other options are incorrect. A network with only input and output layers would not be considered deep.
There is no minimum requirement of five total layers for a network to be classified as deep. The output layer must be connected to the preceding layers in order to receive processed information; it is not isolated
to maintain integrity.

Thus, Option B is the correct statement.

The ISTQB CT-AI syllabus explains that data pre-processing includes activities such as data cleaning and data transformation. Data transformation covers changes to the format or representation of data, for example converting data types or encodings so that they can be correctly processed by machine learning algorithms.

Specifically, the syllabus states that transformation may include converting categorical values to numerical values or changing data representations, such as numbers stored as strings. This directly corresponds to Option D, which describes a dataset issue that can be resolved through pre-processing.

Option A is incorrect because insufficient data is a data availability problem that cannot be solved by pre-processing.
Option B is incorrect because invalid data is typically addressed through data cleaning rather than transformation.
Option C is incorrect because wanted outliers are intentionally retained and should not be removed or altered during pre-processing.

Thus, Option D is the correct answer.

The ISTQB CT-AI syllabus states that A/B testing can be used to evaluate updates to an AI-based system when there are agreed acceptance criteria, such as machine learning functional performance metrics. In this approach, an updated variant of the system is statistically compared with a previous variant to determine whether the change provides a measurable improvement.

This scenario exactly matches that description: users rate the accuracy of both the old and the updated NLP algorithm, and a statistical test is used to compare the results to decide whether the update should be retained. Therefore, this is an example of A/B testing.

The machine learning workflow typically proceeds iteratively, but once the objectives, business priorities, framework, and model have been agreed upon, the next step is data preparation. The ISTQB CT-AI syllabus explains that data preparation is required to provide suitable input data for training an ML algorithm and for evaluating the resulting model.

The syllabus notes that ML models cannot be trained or evaluated without properly prepared data and that data preparation is therefore a prerequisite for subsequent steps such as training, tuning, and evaluation.
This includes collecting, cleaning, transforming, and splitting the data for training and testing purposes.

This directly supports Option A as the correct next step in the machine learning workflow.

The ISTQB CT-AI syllabus explicitly highlights challenges related to handling personal and confidential data when testing AI-based systems. Testing often requires realistic data sets that reflect real-world usage, but such data frequently contains personal or sensitive information. The syllabus notes that it is difficult to fully anonymize data while preserving its usefulness for testing purposes, and that privacy, legal, and regulatory constraints must be carefully managed during test data handling.

Therefore, managing personal data or confidential data represents a significant challenge in AI-based testing, which directly corresponds to Option A.

The other options are not identified by the syllabus as primary challenges related to test data handling.
Output or intermediate data management is a general testing concern, video frame speed or aspect ratio relates to media processing rather than data governance, and data or ML frameworks concern tooling rather than test data handling.

The ISTQB CT-AI syllabus explains that exploratory testing is particularly suitable when specifications are incomplete or unclear, or when there are test oracle problems. This situation frequently occurs with AI-based systems, such as chat assistants, where behavior may be probabilistic, data-driven, or insufficiently specified in advance.

The syllabus explicitly states that exploratory testing is especially useful when there are poor specifications or difficulties in defining expected results. This aligns directly with the described scenario, where the team recognizes that the specifications are insufficient.
Therefore, exploratory testing is the most appropriate testing approach, making Option A the correct answer.

According to the ISTQB CT-AI syllabus, pre-trained models may inherit biases, limitations, or defects from the data sets, assumptions, and processes used during their original training. Section 1.8.3 - Pre-trained Models explains that users of a pre-trained model typically do not have full visibility into or control over the original training data or training process. As a result, such models can carry forward unknown biases or inaccuracies that may not be appropriate for a new context or use case.

In this scenario, although the pre-trained model appears technically suitable for detecting logos based on words and colors, the most significant risk is that it may have inherited unknown defects or biases from its original training, which could affect recognition accuracy or reliability when applied to this specific brand and market context.

Therefore, Option D correctly identifies the most likely associated risk.

The ISTQB CT-AI syllabus explains that experience-based testing and exploratory data analysis (EDA) are effective techniques for identifying potential bias in training data and models. Experience-based testing can be used to verify that the training data set is operationally relevant and representative of the real-world population. EDA supports this by enabling testers to explore distributions, correlations, and patterns in the data that may reveal imbalances or relationships leading to bias.

This directly supports Option A, which combines experience-based testing with EDA to assess bias in the training data and, by extension, in the resulting model.

The other options are incorrect. Back-to-back testing and A/B testing are not appropriate methods for detecting bias in training data; differences between models do not necessarily indicate bias. Acceptance

testing focuses on stakeholder requirements and suitability, not on identifying bias in data or models.

The ISTQB CT-AI syllabus defines automation bias as a cognitive bias in which people tend to favor or trust the outputs of an automated system over their own judgment or other information sources. This bias is also commonly referred to as complacency bias, where users accept automated decisions or suggestions without sufficient critical review.

In this scenario, the automatic text correction system produces an incorrect suggestion ("Isle of Eight" instead of "Isle of Wight"), and several UAT testers accept this output without noticing the error. This behavior directly reflects automation/ complacency bias, as the testers rely on the automated system's output rather than verifying its correctness.

Therefore, Option B correctly identifies the type of bias described.

The ISTQB CT-AI syllabus explains that several factors directly influence machine learning functional performance, including the data pipeline, the quality of data labeling, and the presence of biased data.
Problems in the data pipeline can introduce noise or inconsistencies, poor labeling quality can mislead the learning process, and biased data can distort model outcomes and reduce accuracy or fairness.

The number of classes, however, is primarily a characteristic of the problem formulation and model design.
While it may influence model complexity or architecture, it is not considered a direct factor affecting functional performance metrics such as accuracy, precision, recall, or bias in the sense described by the syllabus.

Therefore, Option D is the correct answer.