Definition · Scope of AIMSS
What is
artificial intelligence?
Before you can certify AI, you have to define it. Etika's working definition is deliberately broad — narrow definitions let consequential systems escape oversight.
Source
Adapted from OECD AI Principles, ISO/IEC 22989, and the EU AI Act (Article 3), with clarifications adopted by the Etika Policy Institute.
AIMSS § 2.1
Working definition
Artificial intelligence is the science and engineering of building intelligent machines — computer systems that can perform tasks with human-like intelligence, such as understanding language, recognising images, learning from data, reasoning, and making decisions. Modern AI typically learns by finding patterns in large amounts of data and using those patterns to generate predictions or responses. It can be narrow, suited to a specific task, or more general-purpose, as with large language models that handle many tasks.
Etika's definition is intentionally broad — narrow definitions let consequential systems escape oversight. A system is in scope because of what it does and the consequences it produces, not because it uses a particular algorithm or architecture.
Families of systems in scope
Six families. One standard.
01
Predictive ML
Statistical models trained on historical data to predict an outcome.
e.g. Credit scoring, churn prediction, diagnostic classifiers.
02
Computer Vision
Models that interpret images, video, or sensor data.
e.g. Facial recognition, medical imaging, autonomous perception.
03
Natural Language
Models that read, classify, or generate human language.
e.g. Translation, summarisation, sentiment, search ranking.
04
Generative & Foundation Models
Large pre-trained systems that produce new text, images, audio, code, or actions.
e.g. LLMs, diffusion models, multimodal assistants, agents.
05
Decision & Optimisation
Systems that rank, recommend, or allocate at scale.
e.g. Recommender systems, dynamic pricing, allocation, routing.
06
Autonomous Control
Systems that take real-world actions with limited human approval per decision.
e.g. Robotics, vehicles, industrial control, trading systems.
Why AI needs its own standard
Five properties that make AI different from ordinary software.
- Trained, not programmed
- Behaviour is learned from data rather than written as rules. The training data is part of the system.
- Statistical, not deterministic
- Outputs are probabilistic. Identical inputs can produce different outputs across versions or sessions.
- Opaque by default
- The path from input to output is not directly inspectable without dedicated interpretability work.
- Consequential at scale
- Decisions are made or influenced for many people simultaneously, often without per-decision review.
- Lifecycle-bound
- Risk is created across data sourcing, training, evaluation, deployment, monitoring, and retirement — not at one point.
Out of scope
Pure rules-based software, deterministic calculators, basic data pipelines, and conventional analytics dashboards are not AI for the purposes of AIMSS, even when marketed as such.
However, when these systems are wrapped around an AI model — for example, a rules engine that filters LLM output — the combined system is in scope.
AIMSS § 2.4 — Definitional disputes are resolved by the Standards Committee.
From definition to certification