Appearance
Ch 2: Quality Characteristics for AI-Based Systems
Official Reference
ISTQB® CT-AI Syllabus v2.0, Chapter 2 (K2, 45 minutes), sections 2.1–2.2, Learning Objectives (LOs) AI-2.1.1, AI-2.1.2, AI-2.2.1. Official syllabus and documents
Exam weight: ~7% of points (3 questions, 3 points, all K2): exactly one question per LO. Typical forms: matching behaviors or measures to characteristics, the "key safety challenge", and the "LEAST likely valid acceptance criterion".
As a tester with a Certified Tester Foundation Level (CTFL) background, you know ISO/IEC 25010 quality characteristics (functional suitability, reliability, security…). This short chapter adds the characteristics that ISO/IEC 25059 defines for AI. It explains why safety is harder to assure when AI is involved, and shows what testable acceptance criteria for these characteristics look like. The vocabulary comes back throughout the rest of the syllabus: AI robustness in adversarial testing, transparency in model documentation, functional correctness in the metrics chapter.
Chapter keywords (know name and definition, K1): functional adaptability, AI functional correctness, intervenability, AI robustness, safety, societal and ethical risk mitigation, transparency, user controllability.
2.1.1 AI-Specific Quality Characteristics (K2)
ISO/IEC 25059 extends the quality model of ISO/IEC 25010 to address AI-specific considerations. It evaluates AI-based systems from two perspectives: product quality and quality in use. For testers, these characteristics directly influence how test objectives are defined, how acceptance criteria are formulated, and how test results are interpreted.
mermaid
flowchart TB
subgraph PQ["Product quality"]
direction LR
FS["Functional suitability"] --> AFC["AI functional correctness<br/>(adjusted)"]
FS --> FA["Functional adaptability"]
IC["Interaction capability<br/>(replaces usability)"] --> UC["User controllability"]
IC --> TR1["Transparency"]
REL["Reliability"] --> ROB["AI robustness"]
SEC["Security"] --> INT["Intervenability"]
end
subgraph QU["Quality in use"]
direction LR
SAT["Satisfaction"] --> TR2["Transparency"]
FFR["Freedom from risk"] --> SER["Societal and ethical<br/>risk mitigation"]
end
PQ ~~~ QULearn each new characteristic together with its parent: the parent is often the deciding clue in exam questions, and transparency is the only one that appears in both perspectives.
| Characteristic | Perspective and parent | What it means (syllabus) |
|---|---|---|
| AI functional correctness | Product quality (adjusted functional correctness) | AI, especially probabilistic ML, cannot guarantee perfect accuracy; a certain error rate is expected. Evaluated by considering both correct and incorrect outputs and defining acceptable thresholds for incorrect results (see 3.3) |
| Functional adaptability | Product quality; new sub-characteristic of functional suitability | The ability of the system to autonomously adapt to changes in its operational environment after it is deployed |
| User controllability | Product quality; new sub-characteristic of interaction capability | A human or another external agent can intervene in its functioning in a timely manner |
| Transparency | Product quality and quality in use; new sub-characteristic of interaction capability and of satisfaction | The degree to which appropriate information about the system is communicated to stakeholders (see 6.1.2) |
| AI robustness | Product quality; new sub-characteristic of reliability | Maintains its level of AI functional correctness regardless of circumstances: biased, adversarial or invalid data inputs, external interference, adverse environmental conditions, operator misuse |
| Intervenability | Product quality; new sub-characteristic of security | The degree to which an operator can intervene in a timely manner to prevent harm or hazard |
| Societal and ethical risk mitigation | Quality in use; new sub-characteristic of freedom from risk | Mitigates societal and ethical risk across many areas (listed below) |
Interaction capability is the term that replaces usability in the 2023 version of ISO/IEC 25010.
Societal and ethical risk mitigation covers: accountability; fairness and non-discrimination; professional responsibility; promotion of human values; privacy; safety and security; human control of technology; community involvement and development; human-centered design; respect for the rule of law; respect for international norms of behavior; environmental sustainability; and labor practices.
Official glossary definitions (ISTQB® Glossary):
Functional correctness (the glossary entry for the keyword "AI functional correctness"): The degree to which a component or system provides the correct results with the needed degree of precision.
Functional adaptability: The degree to which an AI-based system can acquire information from data or previous actions to make future predictions.
User controllability: The degree to which a user can intervene in a system's functioning in a timely manner.
Transparency: The degree to which appropriate information about an AI-based system is communicated to relevant stakeholders.
Robustness (the glossary entry for the keyword "AI robustness"): The degree to which a component or system maintains its level of performance under any circumstances.
Intervenability: The degree to which an operator can intervene in a system's functioning in a timely manner to prevent harm or hazard.
Societal and ethical risk mitigation: The degree to which an AI-based system avoids risks to society arising from its operation.
Safety: The degree to which a component or system under defined conditions can avoid endangering human life, health, property, or the environment.
Functional correctness (the glossary entry for the keyword "AI functional correctness"): The degree to which a component or system provides the correct results with the needed degree of precision.
Functional adaptability: The degree to which an AI-based system can acquire information from data or previous actions to make future predictions.
User controllability: The degree to which a user can intervene in a system's functioning in a timely manner.
Transparency: The degree to which appropriate information about an AI-based system is communicated to relevant stakeholders.
Robustness (the glossary entry for the keyword "AI robustness"): The degree to which a component or system maintains its level of performance under any circumstances.
Intervenability: The degree to which an operator can intervene in a system's functioning in a timely manner to prevent harm or hazard.
Societal and ethical risk mitigation: The degree to which an AI-based system avoids risks to society arising from its operation.
Safety: The degree to which a component or system under defined conditions can avoid endangering human life, health, property, or the environment.
Exam Distinction: user controllability vs. intervenability
Both are about timely intervention. Intervenability (under security) is an operator intervening to prevent harm or hazard: emergency stops, vetoing a critical action. User controllability (under interaction capability) is a human or external agent taking control or overriding the system in normal use: overriding a blocked transaction, taking over a drone. If the scenario stresses danger or harm prevention, choose intervenability.
Exam Distinction: functional adaptability wording
Section 2.1.1 describes functional adaptability as adapting autonomously to changes in the operational environment after deployment. The glossary wording is "acquire information from data or previous actions to make future predictions". Both describe a system that changes its behavior based on what it learns in operation. Recognize either wording.
Classify the behavior (practice the K2 skill):
- A music app shifts its recommendations within a day of a user starting to play jazz → functional adaptability
- A face-unlock model still works in low light and with sunglasses → AI robustness
- A credit model approves equally qualified applicants at different rates by ethnicity → a failure of societal and ethical risk mitigation (fairness and non-discrimination)
- A spam classifier is accepted at 2% false-positive rate rather than "no errors" → AI functional correctness (acceptable threshold for incorrect results)
2.1.2 AI and Safety (K2)
Safety-related systems have the potential to cause injury or harm to people, property or the environment. Developing and testing non-AI safety-related systems takes a lot of effort but is feasible. AI-based systems add five challenges:
| Challenge | Why it hurts safety assurance |
|---|---|
| Specifications | Traditionally, requirements are defined for the complete system and refined down to code. AI requirements often begin with vague goals and are implicitly provided via the training data, without formalizing every detail, so traceability from requirements to implementation can be inadequate |
| Non-determinism | Hard to guarantee precise behavior: even rigorously tested models can behave unexpectedly because of random number generation or slight variations in input values |
| Self-learning | Pre-deployment testing demonstrates safety integrity, but the behavior progressively moves away from the tested behavior. Mitigations: manage how the model learns and the data it uses, or add safety guards that stop it learning or deciding unsafely (e.g., a content moderation component that filters prompts) |
| Explainability and transparency | Safety needs to understand how and why decisions are made, but AI decision-making is often not transparent. Explainable AI techniques such as LIME (Local interpretable model-agnostic explanations) help, but they are not widely available and may compromise system performance |
| Evolving regulations | AI is not included in mature functional safety standards, and some even prohibit its use. The EU AI Act classifies AI used as a safety component (aviation, medical devices, automotive) as high-risk, with strict development and testing requirements |
mermaid
flowchart LR
T["Tested and certified<br/>behavior at release"] --> L["Self-learning<br/>in operation"]
L --> D["Behavior drifts away<br/>from what was tested"]
D -.->|"mitigate"| M1["Manage how the model<br/>learns and its data"]
D -.->|"mitigate"| M2["Safety guards, e.g.<br/>prompt filtering"]Self-learning breaks the usual chain of "test, then certify, then deploy", because the deployed system is no longer the one you tested; the two dotted mitigations limit how far it can drift.
Testing analogy
Think of a self-learning system as software that patches itself in production without a release process. Your pre-release regression suite proves what the system was, not what it is. This is why Chapter 4 recommends re-running an automated core test suite after significant adaptations.
2.2.1 Acceptance Criteria for AI-Based Systems (K2)
For AI-based systems, acceptance criteria often need to be statistical, probabilistic or threshold-based rather than binary, which introduces additional testing challenges. Consider both functional and non-functional characteristics. ISO/IEC 25010 and ISO/IEC 25059 together give the framework. This section focuses on criteria for the AI-specific characteristics (ISO/IEC 25059) and safety.
Example acceptance criteria from the syllabus:
| Characteristic | Example acceptance criteria (syllabus) |
|---|---|
| AI functional correctness (see 3.3) | Accuracy of 95% for an image recognition system. Recall of 90% for a defect prediction system |
| Functional adaptability | Engine management system adapts within 20 seconds of crossing a specified altitude threshold. A streaming service recommends at least 40% documentaries after a user watches three full-length documentaries in one session |
| User controllability | A supervisor can take control of an autonomous drone within 0.5 s when it sends a distress signal after losing GPS. A farm control system notifies the farmer when sensor visual performance degrades by more than 30% (manual override), and fully deactivates above 50% without user response |
| Transparency | Sufficient information about a third-party ML model and the provenance of its training data to meet a company standard. A dashboard/API endpoint returns the deployed model's unique version identifier and a link to its documentation |
| AI robustness | A penetration alert system's predictions stay below 1 s while access to its vulnerabilities database is disrupted for 30 s. An edge AI device switches to a lower-fidelity, reduced-power mode instead of crashing above 85 °C for 10 s |
| Intervenability | A production line can be shut down within 0.5 s after a robot breaches its safety zone. A power-grid system gives a 30-second window for an engineer to veto "critical" AI-proposed actions |
| Societal and ethical risk mitigation | A sentencing system does not discriminate between racial groups based on the specified fairness metric. A chatbot passes an internal red teaming assessment with ≥ 95%, refusing violent, self-harm or hate content |
| Safety | Non-AI components of a steering control system comply with ISO 26262-6 at ASIL C. 100% of input–output relationships of a nuclear plant ML model can be mapped with ≥ 99.9% average accuracy by an explainability tool. Control signals exceeding safety limits by more than 10% are regulated within 0.15 s |
Pattern you can reuse (supplementary, not syllabus wording):
e.g., "The parcel-sorting robot shall keep a misrouting rate below 1% while conveyor vibration is above the specified level" (AI robustness). Every syllabus example has a measurable number and a condition. Criteria without them are hard to test.
[System/actor] shall [observable behavior] [measurable threshold] [under condition / within time]e.g., "The parcel-sorting robot shall keep a misrouting rate below 1% while conveyor vibration is above the specified level" (AI robustness). Every syllabus example has a measurable number and a condition. Criteria without them are hard to test.
Exam Distinction: AI-specific vs. generic criteria
A criterion such as "easy to set up and maintain" or "the admin page loads in 2 seconds" is a generic quality criterion (usability, performance efficiency). It is valid for any system, but it is not an acceptance criterion for an ISO/IEC 25059 AI-specific characteristic. Exam questions asking for the criterion "LEAST likely" to be AI-specific usually hide one of these.
Unit Quiz
Unit Quiz: Chapter 2
Q1. In a warehouse, a remote operator can halt an AI-controlled forklift within one second when it is about to collide with a worker. Which ISO/IEC 25059 characteristic does this behavior demonstrate?
AUser controllability
BAI robustness
CIntervenability
DTransparency
Q2. In ISO/IEC 25059, AI robustness is a new sub-characteristic of which quality characteristic?
AReliability
BSecurity
CFunctional suitability
DInteraction capability
Q3. Which TWO AI-specific characteristics belong to the QUALITY IN USE perspective of ISO/IEC 25059?
Select 2 options.
ATransparency
BAI robustness
CIntervenability
DSocietal and ethical risk mitigation
EFunctional adaptability
Q4. Why does self-learning make it harder to demonstrate the safety of an AI-based system?
ASelf-learning systems cannot be tested before deployment
BSelf-learning requires the system to be connected to the Internet
CSelf-learning systems always become less accurate over time
DTesting before deployment shows safety integrity, but the system's behavior progressively moves away from the behavior that was tested
Q5. The requirements for an AI-based driver-assistance function start as "keep the car safely in its lane" and are then implicitly provided through the training data. Which safety challenge does this illustrate?
ANon-determinism
BSpecifications: vague goals implicitly provided via training data, so traceability from requirements to implementation can be inadequate
CEvolving regulations
DExplainability and transparency
Q6. Why are acceptance criteria for AI-based systems often statistical, probabilistic or threshold-based rather than binary?
ABecause AI-based systems, especially those using probabilistic ML, cannot guarantee perfect accuracy, so a certain error rate has to be accepted
BBecause ISO/IEC 25059 prohibits binary acceptance criteria
CBecause AI-based systems never need to meet functional requirements
DBecause statistical criteria are easier to test than binary ones
Q7. Which acceptance criterion BEST targets the TRANSPARENCY of an AI-based loan system?
AThe model approves loans within 200 ms for 99% of requests
BEach decision letter states the deployed model version and links to the model's documentation, including the provenance of its training data
CA loan officer can override any automated decision within 30 seconds
DThe model keeps an accuracy of 90% when 5% of input fields are missing
0 / 7 answered
Want more? Each section has a scenario-based Topic Drill: start with 2.1.1 AI-Specific Quality Characteristics.