- CT-AI (Certified Tester AI Testing) is the ISTQB certification for testing AI-based systems, especially those built on machine learning (ML).
- Two versions currently coexist: v1.0 (2021, 11 chapters, at least 25.1 h of training) and v2.0 (2026, 7 chapters, at least 19.5 h). v2.0 is the version going forward.
- v1.0 exams are available in English until April 21, 2027 and in other languages until October 21, 2027; after that, everything is v2.0.
- v2.0 focuses only on testing AI and adds generative AI, LLM testing and red teaming; using AI for testing moved to CT-GenAI.
- Exam for both versions: 40 questions, 60 minutes (75 in a non-native language) and 65% to pass (31/47 points in v1.0; 29/44 in v2.0).
- Prerequisite: the ISTQB Foundation Level (CTFL) certificate. You do not need to be a data scientist.
- We offer the ISTQB AI Testing course on demand for companies.
You cannot test an AI system the way you test a traditional one. Its behavior comes from data, not just code; it can give different answers to the same input; it can degrade over time without anyone touching a line; and sometimes nobody knows exactly why it made a decision. For a tester, that breaks many rules: what is the expected result? What counts as a defect? When is it "good enough"?
The ISTQB CT-AI certification answers those questions. In this guide I explain what it is, what changed between version 1.0 and 2.0, what the syllabus covers chapter by chapter, what the exam looks like and how to choose the right version.
What CT-AI is and where it sits in the ISTQB portfolio
CT-AI is an ISTQB Specialist module. You take it after Foundation, and it goes deep into one domain: the quality and testing of AI-based systems. For the full certification map with prices in your currency, see our ISTQB certification guide.
v1.0 vs. v2.0: what changed
v1.0 was released in October 2021 and was the reference for years. In 2026 ISTQB launched v2.0, which reorganizes the content around the machine learning lifecycle: data, model and system.
| Aspect | CT-AI v1.0 (2021) | CT-AI v2.0 (2026) |
|---|---|---|
| Chapters | 11 | 7 |
| Minimum accredited training | 25.1 hours | 19.5 hours (about 3 days) |
| Scope | Testing AI systems and using AI for testing | Testing AI systems only |
| Generative AI and LLMs | Barely covered | Dedicated section: GenAI testing, red teaming, exploratory testing of an LLM |
| Quality | AI-specific characteristics (bias, ethics, transparency and more) | Aligned with ISO/IEC 25059 |
| Test levels | Several levels adapted to AI | Two ML-specific levels: input data testing and model testing |
| Exam | 40 questions, 47 points, pass with 31 | 40 questions, 44 points, pass with 29 |
| Available until | Apr 21, 2027 (English) / Oct 21, 2027 (other languages) | Current |
Which one should you take? If you plan to certify in the coming months and you already studied v1.0, you can finish with it: the certificate is the same "ISTQB CT-AI" and it does not expire. If you are starting from scratch, I recommend v2.0: it is shorter, updated for generative AI and it is the one that will exist after 2027. When you request a quote for the course, we confirm which version it covers based on when you plan to sit the exam.
Who it is for and prerequisites
The syllabus targets anyone involved in testing AI-based systems: testers, test analysts, data analysts, test engineers, test consultants, test managers, user acceptance testers and developers. It also suits people who need to understand the topic: project managers, quality and development managers, business analysts, operations teams, IT directors and consultants.
- Formal prerequisite: the ISTQB Foundation Level certificate (CTFL v4.0 or earlier). If you do not have it yet, start with our ISTQB Foundation Level 4.0 course.
- Recommended: some development or testing experience (ISTQB mentions, for example, six months as a system or acceptance tester, data scientist or developer).
- Not needed: AI or advanced math expertise. The syllabus explains ML from scratch, although the hands-on exercises are easier if you have seen a little Python or worked with data in spreadsheets.
What you will be able to do: business outcomes
v2.0 defines eight business outcomes. A certified person can:
| Code | Business outcome (v2.0) |
|---|---|
| BO1 | Understand the current state of AI, including generative AI |
| BO2 | Experience the implementation and testing of machine learning models |
| BO3 | Understand how simple neural networks work and how they are tested |
| BO4 | Understand the AI-specific quality characteristics defined by ISO/IEC 25059 |
| BO5 | Calculate and interpret ML functional performance metrics |
| BO6 | Recognize the scope and importance of the two test levels specific to ML systems |
| BO7 | Contribute to an effective test strategy for an ML system |
| BO8 | Design and execute test cases for ML systems |
Learning objectives go up to K4 (analyze), and there are hands-on exercises (H0 demo, H1 guided, H2 with hints): building an ML model, preparing data, evaluating a model with metrics, implementing a perceptron, applying metamorphic testing and exploratory testing of an LLM, among others.
Syllabus content chapter by chapter
Syllabus v2.0 (current)
| Chapter | Minimum time | Key topics |
|---|---|---|
| 1. Introduction to Artificial Intelligence | 120 min | AI-based vs. conventional systems, narrow/general/super AI, technologies, generative AI, hardware, model hosting, frameworks, regulation |
| 2. Quality Characteristics for AI-Based Systems | 45 min | AI-specific characteristics (ISO/IEC 25059), safety, acceptance criteria |
| 3. Machine Learning | 375 min | Forms of ML, workflow, pretrained models, fine-tuning and RAG, data preparation, training/validation/test datasets, metrics, neural networks and coverage |
| 4. Testing AI-Based Systems | 195 min | Locked vs. adaptive systems, statistical approach, test oracles, GenAI and LLM testing, red teaming, test levels, risk-based testing |
| 5. Input Data Testing for ML Systems | 180 min | Data risks, bias, data pipelines, representativeness, dataset constraints, label correctness |
| 6. Model Testing for ML Systems | 225 min | Model documentation and review, performance of probabilistic systems, adversarial, metamorphic, drift, overfitting/underfitting, A/B, back-to-back |
| 7. ML Development Testing | 30 min | ML development risks and deployment testing |
Chapter 1. Introduction to Artificial Intelligence
It starts with what sets an AI-based system apart from a conventional one: in a conventional system someone programmed the rules; in an ML system the rules are learned from data. It distinguishes narrow AI (what exists today, specialized in one task), general AI and super AI, and reviews technologies such as rule-based systems, ML and deep learning.
v2.0 adds a generative AI section and practical topics every AI tester runs into: hardware for ML (GPUs and specialized processors), where models are developed and hosted (including AI as a service), development frameworks and the regulations and standards that apply to these systems.
Chapter 2. Quality Characteristics for AI-Based Systems
Short, but very important for the exam. AI systems have quality characteristics that the traditional ISO 25010 model does not fully cover; v2.0 aligns them with ISO/IEC 25059. Think adaptability, autonomy, transparency, explainability, robustness and freedom from bias.
It also covers safety in AI systems, such as a semi-autonomous vehicle, and how to define acceptance criteria. Example: "the fraud classifier must reach a recall of at least 0.90 on the agreed test dataset and show no significant performance differences across customer groups." Without criteria like that, nobody can say whether the system passed.
Chapter 3. Machine Learning
The longest chapter, at more than six hours. It explains the three forms of ML: supervised (learns from labeled examples, such as emails marked spam or not spam), unsupervised (finds groups or patterns without labels) and reinforcement learning (learns from rewards). Then it walks through the ML workflow: understand objectives, choose a framework and algorithm, prepare data, train, evaluate, tune, test, deploy and monitor.
It covers pretrained models, fine-tuning and RAG, which matter a lot now that so many products are built on third-party models. The data part covers preparation (acquisition, pre-processing, feature engineering) and splitting into training, validation and test datasets. A classic mistake you will learn to catch: testing the model on data it already saw during training, which artificially inflates results.
Next come functional performance metrics for classification, based on the confusion matrix: accuracy, precision, recall and F1-score. This is K3: the exam can give you a matrix and ask you to calculate. Quick example: if a fraud model flags 100 transactions, 80 are real fraud and it missed 20 frauds, precision is 80/100 = 0.80 and recall is 80/100 = 0.80.
It closes with neural networks: the structure of a deep network, an exercise implementing a perceptron and coverage measures specific to neural networks (for example, neuron coverage).
Chapter 4. Testing AI-Based Systems
This is where testing proper begins. It distinguishes locked systems (the model does not change in production) from adaptive ones (they keep learning), explains why a statistical approach is needed (a single run proves nothing in a probabilistic system) and tackles the test oracle problem: how do you know the right answer for a recommendation model?
The big new piece in v2.0 is the section on testing generative AI and LLMs, including red teaming: deliberately trying to make the system do what it should not (leak information, bypass its rules, generate harmful content) to find vulnerabilities before an attacker does. There is an exploratory testing of an LLM exercise, for example probing a customer service chatbot with ambiguous, out-of-scope or malicious questions.
Finally, it defines test levels for ML systems and how to apply risk-based testing.
Chapter 5. Input Data Testing for ML Systems
In ML, data is part of the "code". This chapter makes data testing a test level of its own: input data risks and mitigations, bias testing (does the credit dataset represent every customer group?), data pipeline testing (does a transformation break something?), representativeness (does training data look like production data?), dataset constraint testing (ranges, types, null values) and label correctness (a model learns labeling mistakes). It includes a hands-on data testing exercise.
Chapter 6. Model Testing for ML Systems
The second ML-specific level. It covers model risks, model documentation and review, and performance testing of probabilistic systems. It brings together the techniques this certification is known for:
- Adversarial testing: small intentional changes to the input (a few pixels in an image) that make the model fail.
- Metamorphic testing: when you have no oracle, you check relationships. If a system averages a list of numbers, changing the order must not change the result, and doubling every value must double it. In AI it is used for image recognition, search engines and speech recognition.
- Drift testing: detecting when the world changes and the model stops being accurate (concept drift). The classic example is how the COVID-19 pandemic affected sales forecasting models.
- Overfitting and underfitting: a model that memorizes the training data or does not learn enough.
- A/B testing: comparing two variants of the system to see which performs better, for example after retraining.
- Back-to-back testing: using another implementation as a pseudo-oracle and comparing results to find defects.
Chapter 7. ML Development Testing
The short final chapter looks at the process: ML development risks and their mitigations, and deployment testing of the ML system, that is, checking that what reaches production is the right model, with the right configuration, in the right environment.
Syllabus v1.0: how it is organized
If you will sit v1.0 before it is retired, these are its 11 chapters:
| Chapter (v1.0) | Minimum time |
|---|---|
| 1. Introduction to AI | 105 min |
| 2. Quality Characteristics for AI-Based Systems | 105 min |
| 3. Machine Learning (ML) Overview | 145 min |
| 4. ML Data | 230 min |
| 5. ML Functional Performance Metrics | 120 min |
| 6. ML Neural Networks and Testing | 65 min |
| 7. Testing AI-Based Systems Overview | 115 min |
| 8. Testing AI-Specific Quality Characteristics | 150 min |
| 9. Methods and Techniques for the Testing of AI-Based Systems | 245 min |
| 10. Test Environments for AI-Based Systems | 30 min |
| 11. Using AI for Testing | 195 min |
Much of the content is shared with v2.0 (ML, data, metrics, neural networks, and adversarial, metamorphic, A/B and back-to-back testing). What is unique to v1.0: a chapter on characteristics such as flexibility, autonomy, evolution, ethics, side effects and reward hacking; topics such as automation bias (trusting system recommendations too much); pairwise and exploratory testing applied to AI; virtual test environments; and a full chapter on using AI for testing: defect analysis, test case generation, regression suite optimization, defect prediction and GUI testing. That last block now lives, updated, in CT-GenAI.
Download the official syllabus
Here are both versions in English. Read the one you will sit in full: everything is examinable except the introduction, hands-on objectives, references and appendices.
The CT-AI exam
Verified structure
| Item | v1.0 | v2.0 |
|---|---|---|
| Questions | 40 | 40 |
| Total points | 47 | 44 |
| Passing score | 31 points (65%) | 29 points (65%) |
| Duration | 60 min (75 in a non-native language) | 60 min (75 in a non-native language) |
| Cognitive levels | K1 to K4 | K1 to K4 |
Source: official ISTQB exam structure tables and the CT-AI v2.0 FAQ on istqb.org. Expect confusion matrix calculations and scenario questions where you must pick the right technique.
Exam options
- Online from home (iSQI FLEX): remote proctoring 24/7. You need a webcam, a microphone, your phone as a second camera and a government-issued ID. Book ahead.
- Pearson VUE: at a test center or online with OnVUE.
- At your company: for 6 or more people, Cool Testers brings a proctor for an on-site exam.
Paper-based exams are no longer offered. When booking, confirm the version (v1.0 or v2.0) and the language.
Where to buy it
The exam is not included in the course. You can buy it directly from iSQI, our partner and official exam provider, with a 20% coupon for your first attempt that we send when you enroll (best if you pay yourself), or through Cool Testers with our certification exam with 1 or 2 attempts, ideal if your employer pays or you need an invoice. You have 1 year to take it.
How to prepare
- Choose your version first. Do not mix materials: v2.0 questions come from the v2.0 syllabus, not from v1.0.
- Get your CTFL. It is required to sit the exam.
- Master the confusion matrix. Practice accuracy, precision, recall and F1 until they are second nature, and understand when each matters most (in medical diagnosis, a false negative costs far more than a false positive).
- Do the hands-on exercises. Building a simple model and testing it with real data locks in the concepts better than any reading. No-code tools or a basic notebook are enough.
- Learn each technique with your own example: a metamorphic relation, an adversarial case, a drift signal.
- Work through the official sample exam for the version you will sit.
- Take an accredited course. Our ISTQB AI Testing course is delivered on demand for companies, online or on site, on your team's schedule. Request a quote or message us on WhatsApp to hear when we open a group.
Benefits and career paths
- Real differentiation: many testers can use a chatbot; very few know how to test an ML model. It is a scarce profile.
- AI products everywhere: banking (fraud, credit), healthcare, e-commerce (recommendations), automotive, chatbot-based customer service. All of them need data, model and bias testing.
- Regulation: frameworks such as the EU AI Act require evidence of quality and risk control in AI systems; people who can produce that evidence play a key role.
- A bridge to data science: you will speak the same language as data scientists and can challenge their metrics with solid arguments.
- For companies: a team that knows how to test AI lowers the risk of shipping biased or unstable models.
Related certifications and next steps
- Before: CTFL 4.0 (mandatory).
- Natural complement: CT-GenAI, for using generative AI in your testing work. Together they cover both sides of AI in QA.
- Technical: CTAL-TTA or CTAL-TAE to go deeper into technical testing and automation.
- Risk and strategy: CTAL-TM for those who define the test strategy.
- Security: CT-STE, if red teaming caught your interest.
Frequently asked questions
Which CT-AI version should I take, v1.0 or v2.0?
If you are starting from scratch, v2.0: it is shorter, includes generative AI and it is the one that will remain. v1.0 only makes sense if you already studied it and sit the exam before it is retired.
Until when can I take v1.0?
In English until April 21, 2027; in other languages until October 21, 2027, retakes included.
Does my v1.0 certificate lose its value?
No. ISTQB certifications do not expire. Your CT-AI certificate remains valid even though the syllabus version changes.
Do I need programming or advanced math?
It is not required. The syllabus explains ML from scratch. The metrics only need simple calculations from the confusion matrix.
What is the difference between CT-AI and CT-GenAI?
CT-AI teaches you to test systems that use AI or machine learning. CT-GenAI teaches you to use generative AI to test. They complement each other.
What does the exam look like?
40 multiple-choice questions in 60 minutes (75 in a non-native language). You need 65% to pass: 31 of 47 points in v1.0 and 29 of 44 in v2.0.
Is the exam included in the course?
No. You can buy it from iSQI with our 20% coupon, or through Cool Testers if you need an invoice or your employer pays.