Appearance
Exam Tips (CT-AI Syllabus v2.0)
Official References
- Syllabus: ISTQB® CT-AI Syllabus v2.0 (2026-04-17)
- Exam rules: Exam Structures and Rules v1.2 and Exam Structures & Rules Tables v1.21
- Official practice exam: Sample Exam A v2.2: questions and answers (PDF downloads from ISTQB®)
About the Exam
| Detail | Value |
|---|---|
| Duration | 60 minutes (75 minutes if the exam language is not your native language: +25%) |
| Questions | 40 |
| Total points | 44 |
| Passing score | 65% (29 of 44 points) |
| Question types | Single-select (one correct answer out of four); multiple-select ("Select TWO answers" out of five) |
| Allowed aids | None (closed book) |
Source: Exam Structures & Rules Tables v1.21 (CT-AI v2.0 row), and Exam Structures and Rules v1.2 § 6.2.1 for the 25% time extension.
Question Types and Points
Single-select (most questions)
Four options (a–d), and one of them is correct. K2 questions are worth 1 point.
Multiple-select
The question says "Select TWO answers" and offers five options. Sample Exam A awards its 1 point for the correct pair, so treat these as all-or-nothing: two correct choices earn the point, and one correct choice earns nothing.
K3 (apply) questions: 2 points each
The exam rules say that "a K2 question is to be allocated 1 point. In general, a K3 question should be allocated 2 points" (Exam Structures and Rules v1.2, § 5.2.2). CT-AI v2.0 has four K3 Learning Objectives (LOs). The blueprint arithmetic (40 questions, 44 points) means one 2-point question per K3 LO:
| K3 LO (verbatim) | What you must be able to do |
|---|---|
| AI-3.3.1: Calculate common ML functional performance metrics from a given set of confusion matrix data | Compute accuracy, precision, recall and F1-score from raw counts (whichever way round the table is drawn), and interpret them against an acceptance criterion |
| AI-4.2.2: Implement red teaming for GenAI systems | Choose the next step of the red-teaming approach, security vs. safety test scenarios, the right timing, and coverage strategies (crowd-sourced, LLM-generated attacks, hybrid) |
| AI-5.1.5: Apply dataset constraint testing | Choose and evaluate single-value, multi-value and comparison constraints for given rules and records, and plan their automation in the pipeline |
| AI-6.1.5: Use metamorphic testing to derive test cases for a given scenario | Derive and judge metamorphic relations (MRs) and follow-up test cases; spot incorrect MRs (e.g., interacting variables, absolute expected values) |
There is no negative marking: always give an answer.
Question Weighting by Chapter
| Chapter | Questions | Points | Weight |
|---|---|---|---|
| Ch 1: Introduction to Artificial Intelligence | 6 | 6 | 14% |
| Ch 2: Quality Characteristics for AI-Based Systems | 3 | 3 | 7% |
| Ch 3: Machine Learning | 7 | 8 | 18% |
| Ch 4: Testing AI-Based Systems | 7 | 8 | 18% |
| Ch 5: Input Data Testing for ML Systems | 6 | 7 | 16% |
| Ch 6: Model Testing for ML Systems | 9 | 10 | 23% |
| Ch 7: ML Development Testing | 2 | 2 | 5% |
| Total | 40 | 44 | 100% |
Chapter 1 has 8 LOs but only 6 questions, and Chapter 3 has 7 K2 LOs for 6 K2 questions. Some of their LOs are therefore "selectable": not every LO appears in every exam set (Exam Structures and Rules v1.2, § 2.1). Chapters 3–6 together are 75% of the points, and each contains one 2-point K3 question. Prioritize them.
Key Strategies
Read the Question Carefully
- Watch for qualifiers that change the meaning: NOT, BEST, MOST likely, always, only, must.
- For scenario questions, decide what is being asked (a risk, a mitigation, a test level, a characteristic) before reading the options.
- For "Select TWO", eliminate the clearly wrong options first; usually three are wrong for clear syllabus reasons.
Know the Official Definitions
The exam rules define the correct answer as "the best possible answer… according to the relevant syllabus (or Glossary for keywords)" (§ 2.1.4). Learn the keyword definitions in the Glossary. Where the ISTQB® Glossary and syllabus Chapter 9 word a definition differently, the glossary page shows both. Examples:
- Concept drift: "A change in user expectations, behavior or the operational environment over time that negatively impacts the ML functional performance of an operational ML model" (Glossary). Section 6.1.7 explains it as a change in the relationship between the input data and the correct output, while the inputs may look the same.
- Data drift: "A change in the distribution of input data over time that negatively impacts the ML functional performance of an operational ML model" (Glossary).
- Intervenability: the degree to which an operator can intervene in a system's functioning in a timely manner to prevent harm or hazard (a sub-characteristic of security).
- Shadow testing: comparing a new version with the current one by running the new version in the background.
Distinguish Similar Concepts
The exam's distractors are usually real terms from the wrong context. These pairs come up again and again:
| Pair | The difference |
|---|---|
| ML regression vs. regression testing | ML regression predicts a continuous number; regression testing checks for change-related defects |
| Precision vs. recall | Precision = TP / (TP + FP): how sure you can be about positive predictions. Recall = TP / (TP + FN): how sure you can be that no positives are missed |
| Validation vs. test dataset | Validation data is used to evaluate and tune; the independent (hold-out) test dataset is used to test the tuned model |
| Model vs. algorithm hyperparameters | Model hyperparameters define the structure (layers, tree depth); algorithm hyperparameters control training (iterations) |
| Neuron coverage vs. NBC vs. kMNC | Output above a threshold / output outside the training min–max / share of the k sections of each neuron's output range |
| Locked vs. adaptive | Locked systems keep fixed behavior after deployment (an update counts as a new system); adaptive systems change it in operation |
| Security vs. safety red teaming | Security red teaming uses malicious inputs (e.g., indirect prompt injection via RAG); safety red teaming looks for harmful outputs under ordinary use |
| Red teaming vs. Blue Teaming | Red teaming is proactive and before deployment; Blue Teaming is real-time defensive monitoring and filtering in operation |
| Data bias vs. algorithmic bias | Defects in the training data vs. defects in the algorithm, model or framework (e.g., a credit-score threshold) |
| Single-value vs. multi-value vs. comparison constraints | Missing, range, type / sum, count, duplicate, useful, outlier / greater than, correlate |
| Static vs. dynamic drift testing | Static testing compares input and output distributions (e.g., Kolmogorov-Smirnov) without ground truth; dynamic testing needs user feedback (the current ground truth) |
| Overfitting vs. underfitting | Overfitting: much worse on test than on validation data. Underfitting: consistently low on training and validation data; learning curves stay high and close together |
| A/B vs. back-to-back testing | A/B testing compares two variants using metrics and statistics to choose the better one; back-to-back testing uses a pseudo-oracle to detect defects |
| Performance testing vs. ML functional performance testing | Speed and resources vs. accuracy, precision, recall and F1-score |
| Shadow vs. canary testing | In shadow testing the new model's responses never reach users; in canary testing a small subset of real users receives them |
| User controllability vs. intervenability | Timely intervention by a human or external agent vs. operator intervention to prevent harm or hazard |
Statistics: what you do and do not need to compute
You must calculate confusion-matrix metrics (AI-3.3.1, K3). The sample-size and margin-of-error figures in 6.1.3 (601, 589, ±1.1%, 299 test cases) illustrate the concepts only. The syllabus states that "candidates will not be required to derive or compute these values using statistical formulae in the exam". Know what increasing the number of tests does instead: with the confidence level fixed, the margin of error shrinks; with the margin of error fixed, the confidence level rises.
Time Management
- 40 questions in 60 minutes is 1.5 minutes per question (about 1:50 with the 75-minute extension).
- Answer the K2 recall questions quickly, and bank the time for the four 2-point K3 questions, which typically need a calculation or several careful comparisons.
- Flag anything that takes longer than 2 minutes and come back to it. Never leave a question unanswered.
About This Guide's Practice Tiers
| Tier | Where | Difficulty |
|---|---|---|
| Unit Quiz | End of every lesson page | Checks the main points of the page |
| Topic Drills | One page per syllabus section (43 in total) | Covers every point of the section, mostly scenario-based |
| Official Sample Exam A | External PDFs from ISTQB® (see the links above) | As published; mark yourself with the answer PDF |
| Mock Exams 1–4 | Mock Exam 1 · 2 · 3 · 4 | Same format and chapter split as the real exam, but harder; every LO appears in at least 3 of the 4 mocks |
The mock exams group their questions by chapter so that you can review weak areas. The real exam does not.