ISTQB CT-AI: Introduction to AI and Quality Characteristics

Parul Dhingra13+ Years ExperienceHire Me

Senior Quality Analyst

Updated: 1/25/2026

Understanding what artificial intelligence actually is - versus what popular culture suggests - forms the foundation for everything else in AI testing. This article covers the first two chapters of the CT-AI syllabus: Introduction to AI and Quality Characteristics for AI-Based Systems.

These chapters establish essential concepts you'll build upon throughout the certification. You'll learn how to define AI precisely, distinguish between current AI capabilities and theoretical future developments, understand the regulatory landscape shaping AI development, and recognize the unique quality concerns AI systems present.

Chapter 1: Introduction to AI

The first chapter of the CT-AI syllabus establishes foundational AI knowledge that testers need to communicate effectively with AI development teams and understand what they're testing.

Why Testers Need AI Fundamentals

You might wonder why testers need to understand AI theory at all. Can't we just test the system's behavior without understanding how it works internally?

For traditional software, black-box testing works well precisely because deterministic systems behave predictably. Given the same input, you expect the same output. Test oracles are straightforward: compare actual results to expected results.

AI systems break these assumptions. Understanding why helps you design effective tests:

• Probabilistic outputs require different test strategies than deterministic ones
• Training data influences system behavior in ways that requirements don't capture
• Model architecture choices create specific failure modes you should test for
• AI limitations help you identify what can't be tested traditionally

The goal isn't to make you an ML engineer. It's to give you sufficient knowledge to ask the right questions and design appropriate test strategies.

Defining Artificial Intelligence

The term "artificial intelligence" gets thrown around loosely in marketing and media. For testing purposes, you need precise definitions.

The ISTQB Definition

According to ISTQB, artificial intelligence refers to systems that display intelligent behavior by analyzing their environment and taking actions - with some degree of autonomy - to achieve specific goals.

Key elements of this definition:

Intelligent behavior: The system does something that appears intelligent - recognizing patterns, making predictions, understanding language, or making decisions.

Analyzing environment: AI systems perceive and process information from their surroundings, whether that's sensor data, text, images, or structured databases.

Autonomy: AI systems make decisions without explicit human instructions for each action. The degree of autonomy varies widely across systems.

Goal-oriented: AI systems work toward defined objectives, whether that's classifying images correctly, maximizing game scores, or minimizing delivery times.

AI vs Machine Learning vs Deep Learning

These terms are related but distinct:

Artificial Intelligence is the broadest category - any system exhibiting intelligent behavior. This includes rule-based expert systems, statistical models, and learning systems.

Machine Learning is a subset of AI where systems learn patterns from data rather than following explicitly programmed rules. Instead of telling the system exactly what to do, you provide examples and let it discover patterns.

Deep Learning is a subset of machine learning using neural networks with multiple layers. It's particularly effective for complex pattern recognition tasks like image and speech recognition.

AI (broadest)
  └── Machine Learning
        └── Deep Learning

For CT-AI, most content focuses on machine learning systems since they're the most common AI applications today.

Narrow AI vs General AI

One of the most important distinctions in AI is between narrow and general artificial intelligence.

Narrow AI (Weak AI)

Narrow AI systems are designed to perform specific tasks within limited domains. Every AI system deployed today is narrow AI:

• Image classifiers that identify objects in photos
• Speech recognition systems that transcribe audio
• Recommendation engines that suggest products or content
• Language models that generate or translate text
• Game-playing systems that master specific games
• Fraud detection systems that identify suspicious transactions

Characteristics of narrow AI:

• Excels at specific tasks, often surpassing human performance
• Cannot transfer learning to different domains
• Doesn't understand context beyond training
• Has no general reasoning ability
• Cannot set its own goals

A chess AI that beats world champions cannot play checkers without complete retraining. A translation system doesn't understand meaning - it recognizes patterns linking text in different languages.

General AI (Strong AI / AGI)

Artificial General Intelligence would possess human-level intelligence across all cognitive domains. It would:

• Learn any task a human can learn
• Transfer knowledge between domains
• Reason about abstract concepts
• Understand context and meaning
• Potentially set its own goals

AGI does not exist today. Despite media hype and science fiction, no current system approaches general intelligence. This distinction matters for testers because:

1. We're always testing narrow AI with specific capabilities and limitations
2. Test strategies must account for the narrow scope of what systems can actually do
3. Claims about AI capabilities should be evaluated against narrow AI realities

Why This Distinction Matters for Testing

When testing narrow AI, you must understand what the system is designed to do and what lies outside its scope. A narrow AI will fail - often unpredictably - when encountering situations outside its training domain.

Effective AI testing includes:

• Testing within the intended domain (does it work as designed?)
• Testing at domain boundaries (how does it handle edge cases?)
• Testing outside the domain (how does it fail when misused?)

The AI Landscape Today

Understanding current AI capabilities and applications helps contextualize what you're likely to test.

Common AI Applications

Computer Vision

• Object detection and recognition
• Facial recognition
• Medical image analysis
• Quality inspection in manufacturing
• Autonomous vehicle perception

Natural Language Processing

• Text classification and sentiment analysis
• Machine translation
• Chatbots and virtual assistants
• Document summarization
• Named entity recognition

Predictive Analytics

• Demand forecasting
• Churn prediction
• Fraud detection
• Risk assessment
• Maintenance prediction

Recommendation Systems

• Product recommendations
• Content suggestions
• Ad targeting
• Personalization engines

Robotics and Automation

• Autonomous vehicles
• Warehouse robots
• Industrial automation
• Surgical robots

AI Limitations

Understanding AI limitations is as important as understanding capabilities for effective testing:

Data dependency: AI systems are only as good as their training data. Biased, incomplete, or low-quality data produces biased, incomplete, or low-quality models.

Brittleness: AI systems can fail unexpectedly with inputs that differ from training data, even when those inputs seem similar to humans.

Lack of common sense: AI systems don't have general world knowledge. They can make errors that seem obvious to humans.

Opacity: Many AI models, especially deep learning, are difficult to interpret. Understanding why a model made a specific decision is challenging.

Adversarial vulnerability: Carefully crafted inputs can fool AI systems while appearing normal to humans.

AI Regulations and Standards

The regulatory landscape for AI is evolving rapidly. Testers need awareness of regulations affecting the systems they test.

EU AI Act

The European Union's AI Act is the most comprehensive AI regulation to date. It takes a risk-based approach:

Unacceptable Risk (Banned)

• Social scoring by governments
• Real-time biometric identification in public spaces (with limited exceptions)
• Manipulation systems that cause harm

High Risk (Strict Requirements)

• Safety components in regulated products
• Critical infrastructure management
• Educational and vocational training
• Employment and worker management
• Essential services access
• Law enforcement applications
• Migration and border control
• Justice administration

Limited Risk (Transparency Requirements)

• Chatbots (must disclose AI interaction)
• Emotion recognition systems
• Biometric categorization
• Deepfakes and synthetic content

Minimal Risk (No Special Requirements)

• AI-enabled video games
• Spam filters
• Most consumer applications

For high-risk systems, the AI Act requires:

• Risk management systems
• Data governance
• Technical documentation
• Record-keeping
• Transparency to users
• Human oversight
• Accuracy, robustness, and cybersecurity

Other Regulatory Frameworks

IEEE Standards IEEE has developed several AI-related standards including:

• IEEE 7000: Addressing ethical concerns
• IEEE 7001: Transparency of autonomous systems
• IEEE 7002: Data privacy
• IEEE 7010: Well-being metrics

ISO/IEC Standards

• ISO/IEC 22989: AI concepts and terminology
• ISO/IEC 23053: ML framework
• ISO/IEC 42001: AI management systems

Industry-Specific Regulations

• Medical AI: FDA regulations, MDR in EU
• Financial AI: Model risk management requirements
• Automotive AI: Safety standards like ISO 26262

Implications for Testing

Regulations create testing requirements:

• Documentation: Tests must be documented to demonstrate compliance
• Bias testing: High-risk systems require fairness evaluation
• Transparency testing: Systems must provide explanations where required
• Robustness testing: Systems must handle errors gracefully
• Security testing: Protection against adversarial attacks
• Human oversight: Systems must support human intervention

How AI Changes the Tester's Role

AI fundamentally changes what testers do and how they do it.

New Testing Challenges

The Oracle Problem Traditional testing compares actual results to expected results. For AI systems, determining the "correct" answer is often impossible:

• What's the correct sentiment of a nuanced review?
• What's the correct translation of an ambiguous phrase?
• What's the correct next frame an autonomous vehicle should predict?

Testers must develop alternative approaches when exact expected results aren't available.

Non-Determinism The same input might produce different outputs across runs due to:

• Random initialization in training
• Floating-point precision differences
• Model updates in production
• Environmental factors

Testing strategies must accommodate acceptable variation.

Emergent Behavior AI systems can exhibit behaviors not explicitly programmed. Testing must explore what the system might do, not just what it was designed to do.

Data as Specification Training data effectively specifies AI behavior. Testers need to evaluate data quality, not just code quality.

New Testing Skills

CT-AI prepares testers to:

• Understand AI/ML concepts to communicate with development teams
• Evaluate training data quality and representativeness
• Design tests for non-deterministic systems
• Test AI-specific quality characteristics
• Use specialized AI testing techniques
• Evaluate AI-powered testing tools

Collaboration with AI Teams

Testers working with AI systems interact with new roles:

Data Scientists: Develop models and analyze data. Testers provide feedback on model performance in realistic scenarios.

ML Engineers: Build and deploy ML systems. Testers verify end-to-end system behavior, not just model accuracy.

Data Engineers: Manage data pipelines. Testers evaluate data quality and pipeline reliability.

Domain Experts: Provide business knowledge. Testers help translate domain requirements into testable criteria.

Chapter 2: Quality Characteristics for AI Systems

The second chapter extends software quality concepts to AI-specific concerns.

Traditional Quality Characteristics Applied to AI

ISO 25010 defines standard software quality characteristics. These still apply to AI systems but require adaptation:

Functional Suitability

Does the system do what it should?

For AI systems, this means:

• Does the model achieve its intended purpose?
• Does accuracy meet requirements across relevant scenarios?
• Are predictions useful for the business problem?

Performance Efficiency

How well does the system use resources?

For AI systems:

• Inference time (how fast are predictions?)
• Training time (how long to develop/update models?)
• Resource consumption (memory, compute, energy)
• Scalability under load

Compatibility

Does the system work with other systems?

For AI systems:

• Integration with data pipelines
• API compatibility
• Model format standards
• Interoperability with existing systems

Usability

Can users effectively use the system?

For AI systems:

• Are outputs understandable?
• Can users provide feedback for improvement?
• Are confidence levels communicated appropriately?
• Does the interface support appropriate trust calibration?

Reliability

Does the system perform consistently?

For AI systems:

• How consistent are predictions?
• How does the system handle unexpected inputs?
• What happens when the model encounters data drift?
• How does the system degrade under stress?

Security

Is the system protected from threats?

For AI systems:

• Model theft protection
• Adversarial attack resistance
• Training data privacy
• Secure model deployment

Maintainability

Can the system be modified effectively?

For AI systems:

• Model retraining processes
• Version control for models and data
• Monitoring and alerting
• Debugging and diagnosis capabilities

Portability

Can the system move between environments?

For AI systems:

• Model format portability
• Environment independence
• Hardware compatibility
• Deployment flexibility

AI-Specific Quality Characteristics

Beyond traditional characteristics, AI systems have unique quality concerns.

Explainability

Definition: The degree to which a system can explain how and why it reached a particular decision or prediction.

Why it matters:

• Users need to understand recommendations to trust and act on them
• Regulators require explanations for certain high-risk decisions
• Developers need interpretability to debug and improve models
• Affected individuals have rights to explanations in many jurisdictions

Levels of explainability:

• Global: Overall model behavior patterns
• Local: Explanation for specific predictions
• Intrinsic: Models that are inherently interpretable (decision trees, linear models)
• Post-hoc: Explanations generated after predictions (LIME, SHAP)

Testing explainability:

• Are explanations provided when required?
• Are explanations understandable to intended audiences?
• Are explanations accurate (do they reflect actual model reasoning)?
• Are explanations consistent for similar predictions?

Fairness

Definition: The degree to which a system treats different groups equitably and avoids unjust discrimination.

Why it matters:

• Legal requirements prohibit discrimination in many domains
• Ethical obligations to treat people fairly
• Business risk from reputational damage
• Actual harm to individuals from biased decisions

Types of fairness:

• Individual fairness: Similar individuals receive similar treatment
• Group fairness: Different demographic groups receive comparable outcomes
• Statistical parity: Equal positive prediction rates across groups
• Equalized odds: Equal true positive and false positive rates across groups

Testing fairness:

• Define protected attributes and relevant fairness metrics
• Measure model performance across different groups
• Test for proxy discrimination (bias through correlated features)
• Evaluate historical bias in training data

Freedom from Bias

Definition: The degree to which a system is free from systematic errors that favor certain outcomes.

Bias types:

• Selection bias: Training data doesn't represent the target population
• Measurement bias: Systematic errors in data collection
• Aggregation bias: Model assumptions don't fit all subgroups
• Evaluation bias: Benchmark data doesn't represent real-world usage
• Deployment bias: System used differently than designed

Testing for bias:

• Evaluate training data representativeness
• Test model performance on subgroups
• Compare predictions against unbiased baselines
• Monitor for bias in production data

Transparency

Definition: The degree to which system operations, including inputs, outputs, and processes, are visible and understandable.

Components of transparency:

• What data was used to train the model?
• What features influence predictions?
• How was the model validated?
• What are the known limitations?
• Who is responsible for the system?

Testing transparency:

• Is documentation complete and accurate?
• Are data sources identified?
• Are model limitations communicated?
• Can users access relevant information about the system?

Robustness

Definition: The degree to which a system maintains performance under varying conditions, including unexpected inputs and adversarial attacks.

Robustness concerns:

• Input variation: How does the system handle noisy or unusual inputs?
• Distribution shift: How does performance change when data differs from training?
• Adversarial inputs: Can crafted inputs fool the system?
• Edge cases: How does the system handle boundary conditions?

Testing robustness:

• Test with noisy and corrupted inputs
• Test with out-of-distribution data
• Perform adversarial testing
• Evaluate graceful degradation

Autonomy

Definition: The degree to which a system can operate and make decisions without human intervention.

Testing autonomy:

• What decisions does the system make independently?
• When does it escalate to humans?
• Can humans override system decisions?
• Are autonomy boundaries appropriate for the risk level?

Balancing Quality Characteristics

Quality characteristics often trade off against each other. Understanding these trade-offs is essential for testers.

Common Trade-offs

Accuracy vs Explainability More complex models (deep learning) often achieve higher accuracy but are harder to explain. Simpler models (decision trees) are interpretable but may sacrifice accuracy.

Fairness vs Accuracy Enforcing fairness constraints may reduce overall accuracy. The model performs worse overall to ensure equitable treatment across groups.

Performance vs Robustness Highly optimized models may be more vulnerable to adversarial attacks. Robustness techniques often add computational overhead.

Autonomy vs Human Oversight More autonomous systems are efficient but reduce human control. Less autonomous systems are safer but may be slower or more expensive.

Making Trade-off Decisions

Testers should understand that quality decisions aren't purely technical. They involve:

• Business priorities and risk tolerance
• Regulatory requirements
• Ethical considerations
• User needs and expectations
• Cost and resource constraints

Your role is to:

1. Test relevant quality characteristics
2. Communicate trade-offs to stakeholders
3. Provide evidence for decision-making
4. Verify that chosen trade-offs are implemented correctly

Connecting to Other CT-AI Topics

Chapters 1 and 2 provide foundations that later chapters build upon:

• Machine learning overview (Chapter 3) expands on AI concepts introduced here
• Testing AI-specific characteristics (Chapter 8) operationalizes the quality characteristics
• Methods and techniques (Chapter 9) provides concrete approaches for testing these characteristics
• Regulations influence what quality characteristics are mandatory versus optional

Master these foundational chapters before moving to more specialized testing content.

Test Your Knowledge

Frequently Asked Questions

Frequently Asked Questions (FAQs) / People Also Ask (PAA)