OECD AI Principles Guide: Five Values Applied

The OECD AI Principles are the most widely adopted intergovernmental baseline for trustworthy AI, yet many organisations cite them without turning them into controls. First adopted in 2019 and updated in May 2024 to address general-purpose and generative AI, the principles shape national AI strategies, the G20 AI Principles and the definition of an AI system used in later regulation. They are not binding on companies, but regulators, customers and auditors increasingly use them as the reference for what responsible AI looks like. This guide explains the five values-based principles and five policy recommendations, their current status, how they apply to AI agents and the non-human identities they use, and a practical way to map them to your security and governance programme. It is written for security, risk and AI platform teams and the executives who sponsor them.

Key takeaways

Five values, five recommendations. The principles set five values for all AI actors and five recommendations addressed to governments.
Non-binding but influential. The OECD Recommendation on Artificial Intelligence has 47 adherents and the G20 AI Principles are drawn from it, so its language appears in many national and regional frameworks.
The 2024 update is operational. Principle 1.4 now expects AI systems to be capable of being overridden, repaired or decommissioned safely where they risk undue harm, and the update strengthens attention to information integrity, privacy, intellectual property and safety.
Accountability means traceability. Principle 1.5 expects records of datasets, processes and decisions across the lifecycle and a systematic risk management approach at each phase.
Agents make the principles concrete. Override, traceability and transparency for an AI agent depend on its identity, its credentials and its audit trail.

Scope: who and what is covered

The OECD AI Principles are contained in the OECD Council Recommendation on Artificial Intelligence, adopted at ministerial level on 22 May 2019. Adherents are governments: OECD members and non-member countries that choose to adhere. The OECD lists 47 adherents. G20 members committed to the G20 AI Principles, which draw on the OECD text.

Who the principles address

AI actors: the values-based principles (section 1) address all actors who play an active role in the AI system lifecycle, including organisations and individuals that develop, deploy or operate AI. This is the part that matters to companies.
Governments: the recommendations for policy makers (section 2) are addressed to adherent governments and shape funding, regulation and international cooperation.

What counts as an AI system

The OECD defines an AI system as a machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments. Systems vary in their levels of autonomy and adaptiveness after deployment. The definition was revised in late 2023: "content" was added to cover generative AI, "implicit" objectives were added for systems that learn their goals, and adaptiveness after deployment was recognised. The OECD said the revision aimed at broad alignment with policy work internationally, including in the European Union and Japan. The definition covers large language models, agentic systems built on them and classic machine learning alike.

Key obligations

For companies the principles are expectations rather than legal obligations. The table summarises each value and turns it into control objectives a security and governance team can own.

Principle	What the OECD expects	Control objective for an organisation
1.1 Inclusive growth, sustainable development and well-being	Responsible stewardship towards beneficial outcomes, reducing economic, social and gender inequalities and protecting natural environments	Impact assessment covering affected groups; record of intended benefits and environmental cost of training and inference
1.2 Respect for the rule of law, human rights and democratic values, including fairness and privacy	Respect for non-discrimination, privacy and data protection, labour rights and other values throughout the lifecycle; address AI-amplified misinformation and disinformation; safeguards for human agency and oversight, misuse and uses outside intended purpose	Bias testing, data protection impact assessment, human oversight design, intended purpose statement with misuse controls
1.3 Transparency and explainability	General understanding of AI capabilities and limits; awareness of interacting with AI, including at work; plain-language information on data, factors and logic where feasible; information enabling those adversely affected to challenge the output	AI disclosure notices, model and system cards, explanation and contest process
1.4 Robustness, security and safety	Robust, secure and safe throughout the lifecycle under normal use, foreseeable use or misuse and adverse conditions; mechanisms to override, repair and/or decommission safely; information integrity where technically feasible	Security testing including adversarial testing, kill switch and rollback, incident response for AI, provenance controls
1.5 Accountability	Accountability for proper functioning according to role and context; traceability of datasets, processes and decisions; systematic risk management at each lifecycle phase; responsible business conduct and cooperation with suppliers and users	Named owners, AI inventory, audit trails, risk register, supplier assurance

Reading the principles precisely

Transparency has limits: the OECD notes transparency generally does not require releasing source code or proprietary datasets, and that explanations can take different forms depending on context. Plan explanations of main factors and logic, not code disclosure.
Accountability is not liability: the OECD distinguishes accountability (an ethical expectation to answer for outcomes and improve) from liability (legal consequences) and responsibility (causal links). For legal exposure see the AI Liability and Product Liability Guide.
Proportionality: principle 1.5 qualifies risk management by role, context and ability to act, which matters for deployers who do not control the underlying model.

The recommendations for policy makers

Recommendation	Focus	Why it matters to organisations
2.1 Investing in AI research and development	Public and private R&D for trustworthy AI	Funding programmes often require alignment with the principles
2.2 Fostering an inclusive AI-enabling ecosystem	Data, compute and knowledge sharing infrastructure	Shapes data access regimes and sandboxes
2.3 Shaping an enabling interoperable governance and policy environment for AI	Regulation, experimentation and interoperability of frameworks	Explains why national rules reuse OECD definitions and terms
2.4 Building human capacity and preparing for labour market transition	Skills and workforce transitions	Links to AI literacy expectations in regulation
2.5 International co-operation for trustworthy AI	Cross-border cooperation, metrics and standards	Drives shared tools such as the OECD AI Incidents and Hazards Monitor and the Hiroshima reporting framework

Timeline and current status

Date	Event
22 May 2019	OECD Council adopts the Recommendation on Artificial Intelligence at ministerial level
2019	G20 commits to the G20 AI Principles, drawn from the OECD Principles
Late 2023	Revised definition of AI system approved by OECD members (OECD article of 29 November 2023)
May 2024	Principles updated at the 2024 Ministerial Council Meeting to address general-purpose and generative AI, more directly addressing privacy, intellectual property rights, safety and information integrity
7 February 2025	OECD launches the voluntary Hiroshima AI Process Reporting Framework for organisations developing advanced AI systems
2025	First reporting round produces 25 reports, analysed in the OECD report "How are AI developers managing risks?"
30 September 2026	Cut-off for submissions feeding the next analytical review of the Hiroshima reporting framework

As checked on oecd.ai on 24 October 2026, the 2024 text is current and the OECD lists 47 adherents. The adherent count changes as countries join, so check oecd.ai before quoting it.

Enforcement and penalties

There are no penalties. The Recommendation is a non-binding OECD legal instrument: adherents commit politically and the OECD monitors implementation and reports to its Council. The principles gain force indirectly:

Through binding law: regulations that adopt OECD concepts, such as the AI system definition and risk-based lifecycle duties, turn the principles into obligations. See the EU AI Act Compliance Guide and the Council of Europe AI Convention Guide.
Through voluntary reporting: the Hiroshima AI Process Reporting Framework invites developers, deployers and providers of advanced AI systems to report publicly on risk management, building on the Hiroshima Code of Conduct.
Through procurement and contracts: public bodies and large customers reference the principles in AI procurement terms and supplier questionnaires.
Through existing law: privacy, consumer protection, equality and product liability law already apply to AI outputs. See the GDPR Compliance Guide for Security Teams.

What it means for AI agents and non-human identities

AI agents call tools and APIs with service accounts, API keys and OAuth tokens. Each principle has an identity consequence that is easy to miss when the principles are treated as an ethics statement.

Principle	Agent and NHI requirement	Control reference
1.2 human agency and oversight	Human approval gates on high-impact tool calls; agents cannot grant themselves new permissions	NIST SP 800-53 AC-6
1.3 awareness of interacting with AI	Agents identify themselves as AI in messages, tickets and commits; distinct agent identities rather than borrowed human accounts	SP 800-53 IA-9
1.4 override, repair, decommission	Credential revocation and kill switch per agent; decommissioning removes keys, tokens and service accounts	SP 800-53 IR-6, CM-3; CSF 2.0 RS.MA-01
1.4 security under misuse	Prompt injection and tool abuse testing; least-privilege scopes limit blast radius	SP 800-53 SA-15, SI-2
1.5 traceability	Each action attributable to agent identity, model version and human sponsor; integrity-protected logs	SP 800-53 AU-10, AU-12
1.5 cooperation with suppliers	Inventory of models, tools and MCP servers with supplier assurance	SP 800-53 SR-3, SR-6; CSF 2.0 GV.SC-05

Override needs an off switch per identity: if many agents share one credential, you cannot stop one without stopping all. See the AI Agent Kill Switch and Containment Guide.
Traceability needs owners: every agent identity should have a named human or team accountable for it. See the AI Agent Authorisation Guide.
Oversight must be real: approval steps that agents can bypass through broad scopes do not meet principle 1.2. See the Human-in-the-Loop for AI Agents Guide.

Compliance steps

Because the principles are voluntary for companies, the aim is alignment you can evidence, usually as part of a wider AI management system.

Adopt the principles in policy: reference the five principles in your AI policy and assign an executive owner for each. See the AI Governance Guide.
Build the AI inventory: list every AI system and agent using the OECD definition, with owner, purpose, data, model, supplier, autonomy level and the identities it uses.
Map principles to controls: use the tables above to link each principle to controls you already run under NIST CSF 2.0, ISO/IEC 27001 or ISO/IEC 42001, and record gaps. See the ISO/IEC 42001 Guide and the NIST AI RMF Guide.
Assess risk per system: run a lifecycle risk assessment that covers bias, privacy, security, safety, information integrity and intellectual property, scaled to role and context. See the AI Risk Assessment Guide.
Pilot on one agent: implement override, traceability and disclosure controls on a single production agent, test the kill switch in a controlled exercise, and confirm logs answer "who did what, with which model, approved by whom".
Roll out in rings: extend to high-impact systems first, then the rest; keep previous model and prompt versions deployable so changes can be rolled back.
Consider external reporting: developers of advanced AI systems can submit a Hiroshima framework report; others can publish a short transparency statement aligned to the five principles.
Review annually: recheck the OECD text and adherent list, and update mappings as binding law based on OECD concepts takes effect.

Decision guidance

Principles vs management system standard: use the OECD principles as the "why" and the outcome statement; use ISO/IEC 42001 or the NIST AI RMF as the "how" when you need auditable processes.
Small organisations: a principle-to-control table, an inventory and a kill switch for each agent cover most of the value without a full management system.
Large or regulated organisations: integrate the principles into the existing control framework and report against them to the board, with evidence drawn from the same audit trails used for regulators.
Developers vs deployers: developers carry most of 1.3 and 1.4 for the model; deployers carry oversight, disclosure to users, access scoping and incident response for their use.

Detection and measurement

Percentage of AI systems and agents in the inventory with an owner and principle mapping.
Percentage of agents with a unique identity and a tested revocation path.
Time to stop a misbehaving agent in exercises (override capability).
Percentage of high-impact actions with a complete audit record (traceability).
Number of user challenges to AI outputs received and resolved within target (contestability).
AI incidents and near misses recorded per quarter, categorised by principle as the OECD AI Incidents and Hazards Monitor does. See the AI Incident Reporting Guide.

Common failure patterns

Principles as a poster: citing the five values in a policy without mapping them to controls, owners and evidence.
Quoting the 2019 text: missing the 2024 changes on override and decommissioning, information integrity and the revised AI system definition.
Override that does not work: no tested way to stop one agent, or decommissioning that leaves its API keys and tokens live.
Disclaiming the AI: treating an AI channel as separate from the organisation. In the Air Canada Chatbot Ruling 2024 the airline was held responsible for its chatbot's wrong answer.
Traceability without identity: logs that record a shared service account rather than the agent, model version and human sponsor.

Practitioner checklist

Reference the 2024 OECD AI Principles in your AI policy with an owner per principle.
Inventory AI systems and agents using the OECD AI system definition.
Map each principle to controls and record gaps.
Give each agent its own identity, scoped credentials and a named human sponsor.
Test override, rollback and decommissioning, including credential revocation, at least annually.
Log agent actions with identity, model version and approver in tamper-evident storage.
Publish AI disclosure notices and a route for affected people to challenge outputs.
Track incidents by principle and review them with the AI governance board.

Standards and references

This guide summarises the OECD Recommendation on Artificial Intelligence and is not legal advice. Related NHI Mgmt Group resources: AI Governance Guide · EU AI Act Compliance Guide · Global AI Regulation Map · Privacy by Design Guide for Engineering Teams

OECD AI Principles Guide: The Five Values and How to Apply Them