Quality Characteristics of AI-based Systems (ISO/IEC 25059)
The model
ISO/IEC 25059 defines a quality model for AI systems by extending ISO/IEC 25010 (product quality). The familiar characteristics remain - functional suitability, performance efficiency, compatibility, interaction capability, reliability, security, maintainability, flexibility - but they are re-interpreted for learned behaviour, and AI-specific sub-characteristics are added or emphasised.
AI-relevant additions commonly examined include:
- Functional suitability: functional correctness plus functional adaptability (the model can be retrained/transferred to new data or contexts).
- Reliability: fault tolerance and recoverability extended by robustness (correct behaviour on perturbed, noisy or adversarial input) and resilience.
- Interaction capability: transparency and interpretability (users can understand why an output occurred) and controllability (users can correct, override or stop the system).
- Security: resistance to adversarial inputs, data poisoning and model extraction.
- Safety: raised to a first-class concern, including freedom from unacceptable harm and from discriminatory outcomes (bias/fairness is usually treated as part of functional suitability/safety rather than a standalone 25010 characteristic).
- Flexibility replaces portability in the 25059 revision.
Note: 25059 gives characteristics and sub-characteristics, not thresholds - you still have to define measurable requirements.
Traps
- Treating safety as relevant only to physical/robotic systems.
- Confusing transparency (the process is disclosed) with interpretability (the output can be explained).
- Assuming fairness is a top-level characteristic in 25010.
- Forgetting that quality in AI is statistical: a characteristic is measured over a distribution of inputs, not one test case.
Worked rule of thumb
Classify each identified risk: output wrong in normal conditions → functional correctness; wrong under unusual/adversarial input → robustness; user cannot see why → transparency; user cannot override → controllability; performance degrades on a subgroup → fairness/functional correctness.