OECD Framework for AI Systems

OECD Framework for Classification of AI Systems

• This document provides the OECD framework for classifying AI systems.
• It was approved and declassified by the Committee on Digital Economy Policy (CDEP) on January 13, 2022.
• The framework is intended to aid policymakers, regulators, legislators, and others in characterizing AI systems from a policy perspective.
• It links AI system characteristics with the OECD AI Principles.

AI Systems and Their Varied Impact

• Different AI systems offer varied benefits and risks, necessitating distinct policy and governance approaches.
• The OECD’s framework evaluates AI systems across dimensions that include People & Planet; Economic Context; Data & Input; AI Model; and Task & Output.
• The dimensions have properties and attributes that define and assess policy implications, guiding trustworthy AI as outlined in the OECD AI Principles.

Framework Utility

• It allows zooming in on AI-typical risks like bias or explainability while remaining generic.
• It facilitates nuanced policy debates and helps policymakers develop policies and regulations.
• The framework aids identifying AI system features that matter most to tailor policies to specific AI applications
• Identify metrics to assess wellbeing impact or promote a common understanding of AI.
• It also informs registries and inventories, supporting sector-specific frameworks and risk assessments by providing a basis for detailed criteria catalogs, de-risking/mitigation strategies, and incident reporting.
• It supports risk management, informing compliance and enforcement along the Al system lifecycle and pertaining to corporate governance.

Dimensions of AI System Characteristics

• The framework classifies AI systems and applications along dimensions including People & Planet, Economic Context, Data & Input, Al Model, and Task & Output.
• These dimensions assess policy considerations of AI systems.

Key Dimensions Explained

• People & Planet: Focuses on promoting human-centric, trustworthy AI that benefits people and the planet, identifying individuals and groups interacting with or affected by the AI system. Core characteristics include the optionality of the application.
• Economic Context: Describes the economic and sectoral environment, type of organization, and functional area where an AI system is implemented including the sector (e.g., healthcare, finance, manufacturing), business function/model, criticality, deployment, its impact/scale, and technological maturity.
• Data & Input: Describes the data and expert input used to build an AI model’s representation of the environment, including the provenance of data and inputs (machine or human), data format/structure, and data properties.

AI Model

• It is the computational representation of an AI system's external environment including processes, objects, ideas, people, and interactions.
• Core characteristics include technical type, how the model is built (expert knowledge, machine learning, or both), and how the model is used (objectives and performance measures).

Task & Output

• Tasks refer to what the system performs (personalization, recognition, forecasting, etc.)
• Outputs refer to resulting actions that influence the overall context and characteristics include system task(s), action autonomy, and evaluation methods.

AI Lifecycle Integration

• The AI system lifecycle covers these (non-sequential) phases: planning and design, data collection and processing, model building and usage, verification and validation, deployment, and operating and monitoring.
• Linking the OECD Framework to stages in the AI system lifecycle helps identify relevant Al actors for each dimension, supporting clear accountability.