4.1 ISO/IEC 25059 AI Quality Characteristics

Why testers reach for ISO/IEC 25059

Conventional software quality models assume you can specify a function tightly enough that a given input has a right answer and a wrong answer. Many AI-based systems, especially those that use probabilistic machine learning, do not behave that cleanly. ISO/IEC 25059 extends the ISO/IEC 25010 SQuaRE quality model so organizations can evaluate AI-based systems without pretending they are ordinary deterministic programs.

The extension uses two perspectives at the same time:

• Product quality — properties of the AI-based product itself, which you can attribute to the built system.
• Quality in use — outcomes when people and organizations actually use the product in a context of use.

From a testing viewpoint these labels are not decoration. They are how you set test objectives, how you write acceptance criteria, and how you interpret results. A model can meet an accuracy bar on a clean evaluation set and still fail robustness, controllability, or societal-risk objectives. CT-AI v2.0 learning objective AI-2.1.1 (K2) asks you to classify behaviors according to the ISO/IEC 25059 characteristics summarized in the syllabus. The ISO publications are references for extra reading. The exam samples the syllabus summary, not the full clause text of 25010 or 25059.

This OpenExamPrep chapter is independent teaching. It walks through the v2.0 characteristic set testers must be able to use. It does not substitute for buying the ISO documents, and it does not treat OpenExamPrep as an ISTQB body.

Study cue: When a scenario describes a behavior, name the characteristic first, then decide whether you are looking at the product or at quality in use.

What changed relative to ISO/IEC 25010

ISO/IEC 25010 already supplied parent characteristics such as functional suitability, reliability, security, and—on the quality-in-use side—satisfaction and Freedom from risk. The 2023 edition of ISO/IEC 25010 replaced the older usability label with interaction capability. That rename matters on the exam: user controllability and product-side transparency hang off interaction capability, not off the retired usability name.

ISO/IEC 25059 adds or adjusts AI-specific sub-characteristics on those parents. The comparison below is the syllabus-level map you need.

Two exam traps sit in this table. First, transparency appears twice—under interaction capability (product) and under satisfaction (quality in use). Second, user controllability and intervenability both mention timely intervention, but they have different parents and different purposes.

Use the v2.0 set in this chapter. Do not import older specialist-syllabus quality labels that are not on this list.

AI functional correctness (product quality)

AI-based systems, especially probabilistic ML systems, cannot guarantee perfect accuracy. A residual error rate is expected. ISO/IEC 25059 therefore evaluates AI functional correctness by considering both correct and incorrect outputs and by defining acceptable thresholds for incorrect results.

That is a testing change of mindset. You do not stop at the function returned a label. You measure how often the output is right, how often it is wrong, and whether the wrong-answer rate stays inside the agreed band on a defined dataset or operational slice. Metric names such as accuracy, recall, precision, or latency to a correct control action belong with this characteristic; the metrics material later in the syllabus develops those measures. Here, remember the quality idea: correctness is thresholded, and misses count.

Worked mapping. A histopathology model scores 95.2% accuracy on a locked evaluation set whose pass bar is 95%. The misses are logged and reviewed against the error budget. Classify the behavior as AI functional correctness. It is not robustness unless you also showed that the same correctness level survived biased, adversarial, or invalid inputs. It is not adaptability: the operational environment did not change after deploy.

Functional adaptability (product quality)

Functional adaptability is a new sub-characteristic of functional suitability. It is the ability of the AI-based system to autonomously adapt to changes in its operational environment after it is deployed.

Autonomously and after deployment are the words that keep this characteristic from collapsing into ordinary change control. A scheduled retraining that a data-science team runs over the weekend is human-driven maintenance. A live system that restyles its own behavior because the session or the environment changed is adaptability.

Worked mapping. A video service watches a subscriber complete three full-length documentaries in one session and, without a new release, shifts the homepage so documentaries occupy at least 40% of the recommendation slots. Classify as functional adaptability.

Second worked mapping. An engine-management system adapts when it crosses a specified altitude threshold, and the criterion allows a maximum of 20 seconds for that adaptation. Classify as functional adaptability, not as AI functional correctness. The syllabus files this example under functional adaptability for the same reason it files the documentary homepage there: the deployed system autonomously changes its behavior because its operational environment changed. The 20 seconds is a time bound on the adaptation, not a separate correctness metric. Do not be pulled toward correctness merely because a number is attached.

On the exam, read what the stem measures. If it asks whether the intended function produced enough right answers against an error threshold, think AI functional correctness. If it asks whether the deployed system autonomously adjusted because the operational environment moved—altitude, viewing history, load—think functional adaptability, even when a deadline is quoted.

User controllability (product quality)

User controllability is a new sub-characteristic of interaction capability. A human or another external agent can intervene in the AI-based system's functioning in a timely manner.

This is an interaction property: can the user or an authorized external agent take over, override, or steer the AI while it is running, soon enough for the intervention to matter?

Worked mapping. When an autonomous inspection drone broadcasts a distress signal because it lost GPS, a supervisor takes control within 0.5 seconds. Classify as user controllability. A farm-vision controller that notifies the grower at 30% visual degradation and deactivates at 50% if nobody responds is the same characteristic: the human is given a timely path to remain in control of the function.

Keep the parent in mind. Interaction capability replaced usability in ISO/IEC 25010:2023. If a question still says usability as the parent of user controllability, the parent name is stale.

Transparency (product quality and quality in use)

Transparency is the degree to which appropriate information about the AI-based system is communicated to stakeholders. ISO/IEC 25059 treats it as a new sub-characteristic of interaction capability (product) and of satisfaction (quality in use).

Product-side tests check whether the build actually exposes the information: version identifiers, documentation endpoints, training-data provenance. Quality-in-use tests check whether stakeholders receive information they can use to understand, accept, or challenge the system.

Worked mapping. An operations dashboard and an API both return the unique version identifier of the currently deployed prediction model and a link to its documentation. A company standard requires enough information about a third-party model and the provenance of its training data. Both are transparency. Do not treat a specific explainability algorithm as the definition of this characteristic. Communication of appropriate information is the syllabus idea. Techniques such as LIME appear in the safety section as methods that may support understanding of decisions; they are not the definition of transparency.

AI robustness (product quality)

AI robustness is a new sub-characteristic of reliability. It is the ability to maintain AI functional correctness regardless of circumstances such as biased, adversarial, or invalid data inputs, external interference, adverse environmental conditions, and operator misuse.

Clean-set accuracy is correctness. Robustness asks whether that correctness holds up when inputs or the environment are hostile, broken, or abused.

Worked mapping. A penetration-alert model keeps prediction response time below one second while access to the central vulnerability database is disrupted for 30 seconds. An edge device whose internal temperature exceeds 85°C for 10 continuous seconds switches to lower-fidelity, reduced-power inference instead of crashing. Both are AI robustness: the system maintains an acceptable level of AI functional correctness under interference or adverse environment.

Intervenability (product quality)

Intervenability is a new sub-characteristic of security. It is the degree to which an operator can intervene in a timely manner to prevent harm or hazard.

The harm-prevention purpose and the security parent are what separate it from user controllability. Controllability is whether a human or external agent can intervene in functioning in time. Intervenability is whether an operator can intervene in time to stop harm.

Worked mapping. If a packing robot breaches its safety zone, the production line can be shut down within 0.5 seconds after shutdown is initiated. A power-grid manager has a 30-second confirmation window to veto any AI-proposed action classified as critical. Both are intervenability. The same 0.5-second drone takeover after GPS loss remains user controllability, because the story is timely human control of the vehicle, not a security intervention whose stated purpose is preventing harm to people or plant.

Societal and ethical risk mitigation (quality in use)

This new sub-characteristic of Freedom from risk considers a broad set of societal and ethical concerns, including 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.

You are not expected to audit that list as if it were a statute. You are expected to recognize when a behavior is about societal and ethical outcomes of use, not about raw predictive accuracy.

Worked mapping. An automated sentencing aid must not discriminate among racial groups on a specified fairness metric. A customer-support chatbot must pass an internal red-teaming assessment at 95% or higher by refusing content that promotes violence, self-harm, or hate speech. Classify both as societal and ethical risk mitigation, even if next-token accuracy is excellent.

Classification method you can reuse on the exam

1. Decide product versus quality in use.
2. Note the parent (functional suitability, interaction capability, reliability, security, satisfaction, Freedom from risk) if the scenario hints at it.
3. Match the behavior to one of the seven syllabus characteristics.
4. Separate correctness (thresholded right and wrong on the intended function) from robustness (that correctness under adverse conditions).
5. Separate user controllability (timely human or external-agent interaction) from intervenability (timely operator action to prevent harm).

Work one mixed scenario. A warehouse vision model is 97% accurate on the daytime evaluation set (AI functional correctness), keeps that accuracy when labels on cartons are scuffed (AI robustness), lets a floor supervisor freeze the sorter within a second (user controllability), exposes the model version on the shift dashboard (transparency), and is checked so pick-priority does not disadvantage a protected group on a named fairness metric (societal and ethical risk mitigation). One product, five characteristics, five different test objectives.