Last updated July 28, 2026
The ISTQB Certified Tester AI Testing (CT-AI) v2.0 exam has 40 questions worth 44 points in total, and you need 29 points — 65% — in 60 minutes to pass. CTFL is a prerequisite. This guide works from the published v2.0 syllabus and the official exam structure, and it is explicit about one thing most preparation material glosses over: for CT-AI, ISTQB does not publish a per-chapter question count.
40 questions, 44 points — a mix of one-point and two-point items. The heavier ones are the higher cognitive levels, where you apply or analyse rather than recall.
29 of 44 points to pass — 65%. Because the weights differ, counting questions instead of points will mislead you: missing four two-point items costs as much as missing eight one-point items.
60 minutes, or 75 minutes with the standard 25% allowance if you sit the exam in a language that is not your native language.
CTFL is required — this is the one ISTQB specialist exam where candidates most often discover the prerequisite late. No data science degree or programming test is required.
An unanswered question scores exactly what a wrong one does, so never leave one blank.
v2.0 is organised around the lifecycle of an AI-based system rather than around AI theory. Seven areas:
Introduction to Artificial Intelligence — what counts as AI, narrow versus general, AI technologies and the state of the field including generative AI.
Quality Characteristics for AI-Based Systems — the AI-specific characteristics of ISO/IEC 25059 (flexibility, adaptability, autonomy, transparency, explainability, robustness, bias) and how acceptance criteria are written for a probabilistic system.
Machine Learning — supervised, unsupervised and reinforcement learning, data for ML, functional performance metrics for classification, and neural networks.
Testing AI-Based Systems — the test oracle problem under non-determinism, testing generative AI and large language models, and how test levels change for ML systems.
Input Data Testing for Machine Learning Systems — a test level that does not exist outside ML.
Model Testing for Machine Learning Systems — the second ML-specific level.
Machine Learning Development Testing — testing the pipeline that produces the model.
Now the honest part. For CTFL, ISTQB publishes an exam structure table stating exactly how many questions come from each chapter. For CT-AI it does not. Anyone who tells you “chapter 3 is worth eight questions” has invented the number. What you can plan against is the shape of the syllabus: the three ML-specific testing areas plus the machine learning chapter carry most of the applied material, and that is where preparation pays.
Metric arithmetic under time pressure. Given a confusion matrix, compute precision, recall and F1, and say which one matters for the stated goal. Learn to read the matrix in the exam's orientation, not only in the one your textbook drew, and know the accuracy paradox cold: 99% accuracy on a rare-class problem can mean the model never predicts the rare class at all.
Confusing the two ML-specific test levels. Input data testing looks at the data before training — completeness, labelling quality, leakage, representativeness. Model testing looks at the trained model's behaviour. Questions describe a defect and ask which level should have caught it.
The test oracle problem. With a non-deterministic system there is often no single expected result, which is why metamorphic testing, back-to-back testing and A/B comparison exist. Options that assume one fixed expected output are usually the distractor.
Generative AI specifics. Prompt injection versus jailbreaking, groundedness and hallucination, red teaming, why the same prompt can produce different answers, and what an evaluation set for a retrieval-augmented system actually measures.
Bias vocabulary. Algorithmic, sample, inappropriate and historical bias are different things with different mitigations, and the exam separates them.
Areas 1 and 2. Build your own glossary: one line per term, in your words. If you already work with ML this week feels trivial — do it anyway, because the exam tests the ISO/IEC 25059 vocabulary, not your team's.
Area 3, and this is the week that decides the result. Draw confusion matrices by hand until precision, recall and F1 come out in under a minute. Work at least five numeric examples, including one where accuracy is high and recall is terrible. Then read about neural networks only to the depth the syllabus asks: structure and testing implications, not backpropagation mathematics.
Areas 4 to 7 in the first half of the week, keeping the two ML-specific test levels straight in your notes. Then switch entirely to timed 40-question mocks — at least three, on separate days. Stop new material 48 hours out and re-read your glossary and your list of mistakes.
ExamCaliber has 5 CT-AI mock exams drawn from a pool of 220 original questions, each mock 40 questions in 60 minutes at a 65% pass mark, with a written rationale on every option including the wrong ones. Read the rationale for the options you did not pick — that is what transfers to the next question.
The caveat: our mocks weight every question at one point, while the real exam mixes one- and two-point items across 44 points. So read your result as a percentage, not as a point count, and do not try to map “I scored 30” onto the real 29-of-44 threshold. We publish this rather than quietly implying our scoring is identical.
Start with any of the CT-AI mock exams, or read a full mock as a study sheet on its questions and answers page.
They are different certifications and people book the wrong one. CT-AI is about testing AI-based systems — you are the tester, the AI is the system under test. CT-GenAI is about using generative AI to support testing — the AI is your tool. If your job is to assure a model someone else trained, you want CT-AI.
Make sure you are studying v2.0. v1.0 is being retired — the English version remains available until 21 April 2027 and non-English versions until 21 October 2027 — and v2.0 reorganised the material substantially, so a v1.0 question bank tests topics that have moved or gone.
Request the 25% time allowance at booking if it applies to you. It is granted then, not at the desk.
The fee is set by your national member board or exam provider, not by ISTQB, so it varies by country.
Our questions are original, written against the published v2.0 syllabus. They are not real exam questions, and anyone offering you those is offering something ISTQB does not release. A readiness bar that works: 80% or better across three different mocks, taken timed, with no topic area consistently below 65%.