Skip to content

Ch 1: Introduction to Artificial Intelligence ​

Official Reference

ISTQB® CT-AI Syllabus v2.0, Chapter 1 (K2, 120 minutes), section 1.1, Learning Objectives (LOs) AI-1.1.1 to AI-1.1.8. Official syllabus and documents

Exam weight: ~14% of points (6 questions, 6 points, all K2). Chapter 1 has 8 LOs but only 6 questions, so not every LO is tested in every exam. Expect "BEST describes / differentiates" questions rather than calculations.

This chapter is the vocabulary foundation for everything that follows. You will meet these ideas again in later chapters: probabilistic outputs make test oracles hard (Chapter 4), self-learning undermines safety assurance (Chapter 2), and deployment targets drive deployment testing (Chapter 7). If you have no machine learning (ML) background, think of this chapter as the "domain knowledge" a tester needs before testing anything.

Chapter keywords (know name and definition, K1): AI-based system, artificial intelligence, general AI, machine learning, ML development framework, narrow AI, super AI. Each is defined in the section where it first matters.

1.1.1 AI-Based and Conventional Systems (K2) ​

Artificial intelligence (AI): The capability of an engineered system to acquire, process, create and apply knowledge and skills.
AI-based system: A system that incorporates one or more AI components.
Source: ISTQB® Glossary

The syllabus contrasts the two kinds of system on four points:

AspectConventional systemAI-based system (especially ML)
How behavior is definedImperative languages: developers explicitly write step-by-step instructions (if-then-else, loops)Does not follow predefined rules: analyzes patterns in data to decide how to respond to new inputs
Problem-solvingDeterministic: predictable and transparent; easy to see how inputs produce outputsProbabilistic reasoning, statistical inference, pattern recognition: handles uncertainty and ambiguity, but outputs are not always predictable
ExplainabilityLogic can be read in the codeOften a "black box": deep learning models can have billions of parameters
AdaptabilityStatic: manual updates are needed for new knowledgeCan be self-learning: keeps improving with new data, which requires continuous monitoring to stay aligned with core requirements
Tester's view: For a conventional discount calculator you derive the expected result from the rules ("orders over €100 get 10% off"). For an ML image classifier trained to recognize cats, there is no rule to read. It learned from a dataset of cat images and applies the extracted patterns to unseen images. That is why later chapters talk about statistical testing and the test oracle problem.

Why explainability matters: in critical domains such as healthcare, finance, defence and transportation, you need to understand why a model made a decision. Achieving transparency (see 2.1.1) and explainability has become a critical focus of AI regulation.

Exam Distinction

"Learns from data" and "adapts after deployment" are not the same thing. All ML systems learn from data during training, but only self-learning systems keep changing after they are deployed. Locked and adaptive systems are covered in Chapter 4.

1.1.2 Narrow AI, General AI, and Super AI (K2) ​

Narrow AI: AI focused on a single well-defined task to address a specific problem.
General AI: A type of AI that can match human cognitive abilities across most intellectual tasks.
Super AI: A type of AI that far exceeds human capabilities.
Source: ISTQB® Glossary
mermaid
flowchart TB
  N["Narrow AI (weak AI)<br/>specific tasks<br/>all AI deployed today"] -->|"not achieved yet"| G["General AI (strong AI)<br/>most intellectual tasks<br/>a human can do"]
  G -->|"technological singularity"| S["Super AI<br/>improves itself without<br/>human intervention"]
  F["Frontier AI<br/>most advanced narrow AI"] -.->|"subset of"| N

Everything that exists today, frontier AI included, is narrow AI; every step below narrow AI is a capability that has not been achieved yet.

Narrow AIGeneral AISuper AI
Also calledWeak AIStrong AIArtificial superintelligence
ScopeSpecific tasks within a limited domain (image recognition, speech processing, translation)Most intellectual tasks a human can do, without retraining for each new taskSurpasses human intelligence and general AI
StatusAll deployed AI systems in use todayNo AI-based system today has itHypothetical; many believe it would pose an existential risk
Key traitHighly efficient, but cannot generalize: a face recognizer cannot translate unless retrainedHuman-like reasoning and adaptability across domainsContinuously improves itself without human intervention or control

More details the exam can use:

  • Frontier AI is the most advanced form of narrow AI, pushing current capabilities with GenAI models. It includes large-scale systems with highly autonomous decision-making, but it remains task-specific and is not general AI.
  • Super AI does not strictly require Internet access, but such access could significantly expand its capabilities and influence.
  • The point at which AI would move from general AI to super AI, if it ever occurs, is the technological singularity.

Exam Distinction: frontier AI

The section text (1.1.2) calls frontier AI a subset of narrow AI. The glossary entry (also in syllabus Chapter 9) defines it as "a general-purpose AI-based system that exceeds the capabilities of today's most advanced AI systems". For a question about where frontier AI sits on the narrow–general–super spectrum, follow the section text: it is still narrow AI.

1.1.3 Different Types of AI Technologies (K2) ​

Machine learning (ML): The process of using computational techniques to enable systems to learn from data or experience.
Source: ISTQB® Glossary

ML is one of the core branches of AI. It lets systems learn from data and build models without explicit programming. Some ML systems (MLS) keep adapting with new data throughout their lifetime. Others work on what they initially learned and need explicit retraining to update their capabilities.

mermaid
flowchart TB
  AI["Artificial intelligence"] --> ML["Machine learning<br/>supervised / unsupervised / reinforcement"]
  AI --> OT["Other AI technologies<br/>NLP, computer vision, fuzzy logic,<br/>search algorithms, rule-based / expert systems"]
  ML --> DL["Deep learning<br/>CNN, RNN, transformers"]
  DL --> GEN["Generative AI<br/>e.g. LLMs = DNN + NLP"]
  GEN -.-> AG["Agentic AI<br/>autonomous agents that plan, reason, act"]

Read the diagram as nested sets: deep learning is a subset of ML, which is a subset of AI, and GenAI builds on deep learning; it is not a separate branch of AI.

The three ML approaches (more on these in Chapter 3):

ApproachDataExample algorithms and uses from the syllabus
Supervised learningLabeled dataLinear regression, decision trees; prediction and classification
Unsupervised learningUnlabeled dataClustering (with clustering algorithms) to uncover patterns
Reinforcement learningInteraction with an environmentIntelligent agents learn optimal behaviors by trial and error

Key ML technologies: neural networks, Bayesian models, support vector machines (SVM) and random forests.

Deep learning (DL) is a subset of ML that uses deep neural networks (DNN) to solve complex problems:

  • Convolutional neural networks (CNN): image recognition and object detection
  • Recurrent neural networks (RNN): sequential data such as text or time series
  • Transformers: long-range dependencies in sequences. They power natural language processing (NLP) models and vision transformers for images

GenAI builds on these technologies to create new content. It is driven by models such as large language models (LLM), which combine DNNs with NLP to analyze and generate human-like language.

Other specialized AI technologies:

TechnologyUsed for (syllabus examples)
NLPLanguage analysis: sentiment analysis, machine translation
Computer visionVisual data: facial recognition, robotics
Fuzzy logicReasoning under uncertainty
Search algorithmsOptimization problems (e.g., navigation) and strategic decision-making
Rule-based reasoning systems (expert systems)Structured decision support

Full integration of these technologies is still limited, but LLMs show the potential to combine them into unified intelligent systems. Agentic AI extends them through autonomous agents that plan, reason, and act independently to achieve goals in dynamic environments.

Testing analogy

An expert system is "AI" but behaves much like conventional software: its rules are written by humans, so you can derive expected results from them. Do not assume that every AI-based system is probabilistic.

1.1.4 Generative AI (K2) ​

Generative AI (GenAI): A type of AI that uses ML to generate new content that resembles human-created content.
Source: ISTQB® Glossary. Syllabus Chapter 9 words it as "creates new content by learning patterns from existing data".

GenAI creates new content (text, images, video, music or complex data), and many GenAI systems also support classification and prediction. They learn from vast amounts of data and produce outputs that resemble their training data.

Main technologies behind GenAI:

TechnologyHow it works (syllabus wording)
Generative adversarial networks (GANs)Two neural networks in competition to create highly realistic synthetic data
Diffusion modelsGenerate content by gradually adding and then removing noise from data
TransformersUnderpin LLMs; use self-attention to generate coherent, contextually relevant text; increasingly adapted for multimodal tasks

Societal and ethical concerns (know all three and the benefits):

ConcernWhat the syllabus says
MisuseDeepfakes, misinformation, convincing fraudulent content, which undermine trust in digital media and public discourse. Synthetic content amplifies privacy, security and manipulation risks, but also brings benefits in entertainment, marketing, education and research
EmploymentEspecially white-collar work: writing, design, coding, legal and medical documentation. GenAI boosts productivity and creative opportunities, but may disrupt the workforce and require widespread reskilling
SustainabilitySubstantial compute leads to high energy consumption and a significant carbon footprint, prompting calls for more efficient model designs and greener infrastructure

Most practical GenAI tools today are based on foundation models, which are fine-tuned for specific applications. The field is moving toward multimodal models that process and generate text, images and audio. Regulatory frameworks such as the EU AI Act are emerging to guide responsible development and use.

1.1.5 Hardware for Machine Learning Systems (K2) ​

Different hardware may be used for training and for inference (making predictions). For example, a speech recognition model may run on a low-end smartphone, although cloud computing may be needed to train it.

ML typically benefits from hardware that supports:

  1. working with large data structures;
  2. massively parallel (concurrent) processing, e.g., for matrix multiplication;
  3. low-precision arithmetic (quantization): fewer bits (e.g., 4 instead of 32), which gives faster processing, lower power consumption, smaller and more cost-effective chips, and reduced bandwidth requirements.
HardwareCharacteristics in the syllabusTypical fit
CPU (central processing unit)Complex, high-precision operations that ML does not typically need; few cores; usually higher clock speedsLess efficient for training and running ML models
GPU (graphics processing unit)Thousands of cores for massively parallel, relatively simple processingTypically outperforms CPUs for ML; generally the best option for small-scale ML work
AI-specific hardware: Application-Specific Integrated Circuits (ASIC), System-on-a-ChipMultiple cores, specialized data management, in-memory processingBest suited to edge computing, with training done in the cloud on specialized hardware
Neuromorphic processorsNot von Neumann: brain-inspired designs mimicking neuronal structuresStill being developed

Exam Distinction

CPUs have the higher clock speed, yet GPUs win for ML because of parallelism. ASICs and Systems-on-a-Chip are for edge use; training still happens in the cloud.

1.1.6 Development and Hosting of AI Models (K2) ​

AI as a Service (AIaaS): A software licensing and delivery model in which AI and AI development services are centrally hosted.
Source: ISTQB® Glossary

Acquire or build?

  • Third-party AI-based systems usually come as pretrained models or AIaaS, which means faster deployment and quicker time-to-market.
  • Private AI-based systems (on-premises or customized cloud setups) can be better tailored to specific requirements, but need specialized skills, from in-house experts or outsourced teams.

The syllabus treats development and hosting as two separate decisions, each with local, cloud and hybrid options:

LocalCloudHybrid
DevelopmentDirect control and privacy. Small models (decision trees, compact neural networks) on personal computers; mid-sized need dedicated GPUs; large ones need on-premises server clusters (energy, cooling, hardware costs)Public: pre-configured environments, pay-as-you-go, low initial hardware investment, easy scaling. Private: enhanced security and privacy, but greater upfront infrastructure investmentPrototype locally, then scale to the cloud; keep sensitive components on-premises (e.g., preparation of private data); use the cloud for compute-intensive tasks
HostingSmaller models on PCs or smartphones: privacy, no cloud licensing costs, limited hardware. Larger models on dedicated servers: significant upfront investment, enhanced controlPublic: scalable, powerful, no maintenance concerns, ideal for fluctuating workloads. Private: similar benefits plus enhanced security and customization, in-house or via dedicated providers, at higher costSome operations run locally, with cloud elasticity for intensive tasks

Decision factors (development and hosting are typically decided separately): model size, complexity, performance requirements, budget constraints, security and data privacy, deployment needs, and regulatory requirements. Some organizations adopt multi-tiered strategies to balance efficiency and control.

Example: An insurer anonymizes claim data on its own servers (sensitive data stays on-premises), trains a fraud model on rented cloud GPUs (compute-intensive), and hosts the model behind a cloud API. Development and hosting are both hybrid here, and they were chosen separately.

Exam Distinction

To spot a hybrid option, look for a mix of locations. "Everything on a public cloud", "everything on local servers" or "everything on a private cloud" are not hybrid, however attractive the stated benefit.

1.1.7 Machine Learning Development Frameworks (K2) ​

ML development framework: A software platform that provides tools and libraries to build, train, and deploy ML models.
Source: ISTQB® Glossary

Typical functionality (five areas):

FunctionWhat the framework provides
Data handlingLoading, preprocessing and managing training and test data: cleaning, formatting, transforming
Model buildingLibraries of ML algorithms and tools to design the model architecture: model type (e.g., neural network, decision tree), number of layers and connections, mathematical operations
Training and optimizationAlgorithms that iteratively adjust the model's internal parameters to optimize performance; some support distributed training and enhancing or fine-tuning pretrained models
EvaluationTools to measure performance on unseen data: accuracy, precision, recall (classification), error rates (ML regression)
DeploymentSome frameworks convert the trained model into a format for web applications, mobile, edge devices or embedded systems

Levels of abstraction: a lower-level API gives more control but needs more coding expertise. A higher-level API simplifies model creation but offers fewer customization options.

Application domains: some frameworks are general purpose. Others are specialized (image recognition, speech recognition, language translation).

Selection factors (seven):

  1. the application area
  2. the need for a user-friendly interface for rapid prototyping
  3. configurability for complex models
  4. the expertise of the users
  5. deployment considerations (some suit resource-constrained environments better)
  6. level of (community) support
  7. ecosystem maturity

Supplementary

TensorFlow and PyTorch, which the syllabus mentions in Chapter 7, are well-known general-purpose ML development frameworks.

1.1.8 Regulations and Standards for AI (K2) ​

Core aim: to foster trust in AI and help realize its benefits while mitigating potential harms. Ideally, compliance should guarantee that AI-based systems are safe, fair, transparent, sustainable, accountable, ethical, and used responsibly.

InstrumentTypeWhat the syllabus says
OECD AI Principles, UN "Governing AI for Humanity" reportInternational soft lawA shared understanding of responsible AI stewardship; a "compass" for governments and organizations; emphasize human-centric AI, ethics, international cooperation
EU AI ActBinding regulation (landmark)Risk-based: AI systems are categorized from minimal to unacceptable risk. High-risk systems (impacting fundamental rights or safety) face stringent requirements: rigorous testing, data governance, human oversight. Substantial penalties based on a percentage of global turnover
Many nations outside the EUNational approachesMore permissive, lighter-touch regulation to encourage innovation
Technical standards (ISO, IEEE)StandardsTurn high-level aspirations into concrete specifications and best practices. ISO/IEC TR 29119-11 gives detailed guidance on testing AI-based systems, a critical element in demonstrating regulatory compliance. The ISO/IEC 42119 series is being developed to cover AI testing
Sector-specific regulationsEmergingE.g., healthcare and finance

Because AI changes quickly, the syllabus stresses continuous dialogue and collaboration between governments, industry, academia and civil society, and says regulations and standards must be regularly reviewed and updated.

Supplementary

The EU AI Act's four tiers are usually named unacceptable, high, limited and minimal risk. The syllabus itself only says "from minimal to unacceptable".

Exam Distinction

Soft law (OECD, UN) guides but does not penalize. Standards (ISO/IEC TR 29119-11, ISO/IEC 25059) give guidance and are not enforced by penalties. The EU AI Act is the instrument with severe penalties.

Unit Quiz ​

Unit Quiz: Chapter 1

Q1. Which statement BEST describes how most ML-based systems differ from conventional systems in the way they produce outputs?

AThey execute step-by-step instructions written in an imperative language, so their outputs are fully predictable
BThey are always faster, because they skip explicit programming and rely on hardware acceleration
CThey analyze patterns in data to decide how to respond to new inputs, using probabilistic reasoning, so outputs are not always predictable
DThey are easier to explain than conventional systems, because their decisions are based on real data

Q2. An AI-based recommendation system keeps learning from new user data after go-live. According to the syllabus, what does this adaptability ALSO require?

AContinuous monitoring to maintain ongoing alignment with core requirements
BManual updates whenever new knowledge must be incorporated
CA switch to imperative programming for the parts that learn
DRemoving explainability features so that the model can learn faster

Q3. Which TWO statements about narrow AI are CORRECT according to the syllabus?

Select 2 options.

ANarrow AI can generalize to new tasks without retraining
BAll deployed AI systems in use today are narrow AI
CNarrow AI is also known as strong AI
DFrontier AI is described in section 1.1.2 as a subset of narrow AI
ENarrow AI continuously improves itself without human intervention

Q4. What is the "technological singularity"?

AThe moment a narrow AI system is first deployed to production
BThe point at which an AI system has been trained on all available Internet data
CThe point at which general AI is first achieved by a frontier AI model
DThe point at which AI-based systems would transition from general AI to super AI, if it were to occur

Q5. Which matching of deep learning architectures to their typical strengths is CORRECT?

ACNN: sequential data such as time series; RNN: image recognition; transformer: rule-based reasoning
BCNN: image recognition and object detection; RNN: sequential data such as text or time series; transformer: long-range dependencies in sequences
CCNN: long-range dependencies; RNN: object detection; transformer: fuzzy reasoning under uncertainty
DCNN: natural language processing only; RNN: clustering; transformer: search and navigation

Q6. Which of the following is NOT one of the deep learning examples given in the syllabus?

AConvolutional neural network
BRecurrent neural network
CFuzzy logic
DTransformer

Q7. Which GenAI technology generates content by gradually adding and then removing noise from data?

ADiffusion models
BGenerative adversarial networks (GANs)
CTransformers
DRecurrent neural networks

Q8. A company notes that training its large GenAI model consumed a lot of energy and caused a significant carbon footprint. Which GenAI concern from the syllabus does this illustrate?

AMisuse
BImpact on employment
CPrivacy
DSustainability

Q9. Why do GPUs typically outperform CPUs for ML applications?

AGPUs run at higher clock speeds than CPUs
BGPUs have thousands of cores designed for massively parallel, relatively simple processing, such as matrix multiplication
CGPUs provide high-precision arithmetic for complex operations, which ML requires
DGPUs use the neuromorphic architecture instead of the von Neumann architecture

Q10. Which is a benefit of low-precision arithmetic (quantization) for ML hardware, as given in the syllabus?

AHigher numeric precision for training
BElimination of the need for parallel processing
CAbility to run on the von Neumann architecture only
DFaster processing, lower power consumption and reduced bandwidth requirements

Q11. Which of the following is an example of a HYBRID approach to developing and hosting AI models?

APreparing sensitive private data on-premises, then using cloud resources for compute-intensive training
BDeveloping and hosting everything on a public cloud with pay-as-you-go pricing
CRunning all development and hosting on dedicated local servers
DHosting a small model on a smartphone that was also trained on that smartphone

Q12. What does the syllabus give as the main advantage of acquiring a third-party AI-based system as a pretrained model or AI as a Service (AIaaS)?

AIt can be tailored more closely to specific requirements than a private system
BIt guarantees the highest data privacy
CFaster deployment and quicker time-to-market
DIt removes the need to test the system

Q13. An ML development framework offers a higher-level API. What trade-off does the syllabus describe compared with a lower-level API?

AIt gives more control over model building but requires more coding expertise
BIt simplifies model creation but offers fewer customization options
CIt supports only image recognition and speech recognition
DIt removes the need for data handling and evaluation

Q14. How does the syllabus characterize the EU AI Act?

AAs a soft law instrument that acts as a compass for national AI strategies
BAs a technical standard giving detailed guidance on testing AI-based systems
CAs a lighter-touch regulation designed mainly to encourage innovation
DAs a landmark regulation with a risk-based approach, imposing stringent requirements on high-risk systems and substantial penalties

Q15. Which document does the syllabus name as providing detailed guidance on testing AI-based systems, helping to demonstrate regulatory compliance?

AISO/IEC TR 29119-11
BThe OECD AI Principles
CThe UN report "Governing AI for Humanity"
DISO/IEC 25059

0 / 15 answered

Want more? Each section has a scenario-based Topic Drill: start with 1.1.1 AI-Based and Conventional Systems.

Unofficial study guide for the ISTQB® Certified Tester AI Testing (CT-AI) exam. Not affiliated with International Software Testing Qualifications Board (ISTQB®).