A quality model for AI — ISO/IEC 25059

Lesson 5 of 5 in Trustworthiness, Bias, and AI Quality: The TR Series and 25059.

For twenty years, software teams have specified and evaluated products against the SQuaRE family (Systems and Software Quality Requirements and Evaluation, the ISO/IEC 25000 series), whose centrepiece — ISO/IEC 25010 — defines the quality characteristics every requirements document borrows: functional suitability, reliability, security, usability, maintainability and the rest. Procurement contracts cite it; acceptance tests are structured by it.

Then AI broke its quietest assumption. 25010 presumes deterministic correctness: given a specification, an output is right or wrong, and testing can prove which. An ML model is statistically correct — right 94% of the time, wrong in patterned ways, and possibly different next month. "Functional correctness" as classically defined is unachievable and unmeasurable for such a system. ISO/IEC 25059:2023 exists to repair exactly this: it extends the 25010 model with AI-specific characteristics so that quality requirements, acceptance criteria, and evaluations can be written for systems that are probabilistic by nature. It is a full International Standard — the taxonomy is standardisable even where the thresholds are not.

What 25059 changes relative to base 25010
CharacteristicNew or adaptedWhat it captures for AI systems

Functional correctness

Adapted

Redefined for probabilistic outputs: correctness with degrees and confidence bounds, specified statistically (e.g. per TS 4213 metrics), not as pass/fail determinism

Functional adaptability

New

The system’s capacity to perform in contexts and on data beyond its original specification — the quality-model home of generalisation and distribution shift

Robustness

New

Maintaining performance under perturbation and adversarial pressure — the 24028 attribute promoted into a specifiable product quality, measurable via the 24029 series

User controllability

New

Whether users and operators can direct, correct, and override the system — including intervenability, the ability to intervene in an operation in progress

Transparency

New

Whether appropriate information about the system — capabilities, limitations, data, logic — is available to the right stakeholders as a product property, not a policy afterthought

Societal & ethical risk mitigation (quality in use)

New

Quality experienced in real use now includes whether the system mitigates harm to individuals and society — TR 24368’s themes arriving as evaluable quality-in-use criteria

Closing the loop: from guidance to AIMS artifacts. You now hold all four documents plus the measurement ring. Here is how a competent AIMS lead wires them together — the pattern this whole module has been building toward:

  • Annex C objectives (fairness, robustness, transparency, safety…) are selected and justified using TR 24028’s attribute definitions and TR 24368’s themes.
  • Clause 6 risk assessment draws its AI-specific risk sources from TR 24028’s threat catalogue and TR 24027’s bias taxonomy — that is what "informed by the state of the art" looks like in evidence.
  • A.6 lifecycle controls state requirements and acceptance criteria in 25059’s vocabulary; A.7 data controls lean on the 5259 series for data-quality process.
  • Clause 9 monitoring measures what the objectives promised, with methods from TS 4213 and 24029 — numbers with pedigrees, reviewable at management review.

An auditor who sees this chain — objective, risk source, control, metric, each traceable to published state of the art — sees an organisation that knows why its management system looks the way it does. That traceability, not any individual document, is the professional skill this module certifies.

The guidance-to-artifact pipeline

  1. TR 24028 / 24027 / 24368

    Informative state of the art: attributes, threats, bias taxonomy, ethics themes.

  2. Annex C objectives

    The organisation selects and justifies AI objectives — fairness, robustness, transparency — for its context.

  3. Clause 6 risk & impact assessment

    Threats and bias types become identified risk sources; 42005 adds the affected-party lens.

  4. A.6 / A.7 controls with 25059 criteria

    Lifecycle and data controls state acceptance criteria in quality-model vocabulary.

  5. Clause 9 metrics (TS 4213, 24029, 5259)

    Measurement standards supply valid methods; results feed management review.

  6. Management review → improvement

    Clause 9.3 and 10: results drive corrective action and updated objectives — the loop closes.

Key terms: quality model, SQuaRE series, functional adaptability, intervenability, quality in use

Tool: Fairness Lab — Feel the TR 24027 trade-offs yourself: tune per-group thresholds on a toy classifier and watch demographic parity, equalized odds, and accuracy fight each other.

Interactive checkpoint quiz (1 questions) — open this page in a browser to take it.