What kind of document it is
The Recommendation is an OECD Council instrument. The OECD says its Recommendations are not legally binding but represent a political commitment, with an expectation that adherents will do their best to implement them (p. 12). Countries that agree to it are called "Adherents", and the text also invites non-adherents to take it into account (p. 10). The document in this read is the consolidated text, which the OECD published with background information (pp. 3-6).
It says it complements existing OECD standards that already affect AI, such as those on privacy and data protection, digital security risk management and responsible business conduct, and that it concentrates on issues specific to AI (p. 4).
Why it was written
The OECD describes AI as a general-purpose technology that can improve welfare, support growth and innovation, and help with global challenges, while also raising challenges around economic shifts, inequality, competition, labour markets, democracy and human rights (p. 3). It treats trust as a key enabler of digital transformation, and says the trustworthiness of AI systems is a key factor in whether they are adopted. It also describes trustworthiness as something that can become a competitive parameter in the global marketplace (p. 7).
The text was developed through a participatory process. A group of more than 50 experts from government, industry, civil society, trade unions, the technical community and academia met four times between September 2018 and February 2019. A draft was then agreed in the Digital Policy Committee and adopted by the Council at ministerial level in May 2019 (p. 3).
What the preamble says about existing law
The preamble matters for any business. It underlines that existing national and international legal, regulatory and policy frameworks already have relevance to AI, and it names human rights, consumer and personal data protection, intellectual property rights, responsible business conduct and competition. It adds that the appropriateness of some of these frameworks may need to be assessed and new approaches developed (p. 7). It also recognises that AI transformations may have disparate effects within and between societies, notably on the labour market, inequalities, democracy, human rights, privacy and digital security, and it calls for a stable, human-centric policy environment that fosters research and preserves economic incentives to innovate (p. 7).
How the operative text is built
After the preamble, the operative paragraphs do different jobs (pp. 7-11). The Council agrees on the definitions. It recommends that Adherents promote and apply the principles, calls on all AI actors to do the same according to their roles, and underlines that the principles are complementary. It then recommends that Adherents carry out the five national-policy recommendations, invites the Secretary-General and Adherents to spread the text and invites non-Adherents to take it into account, and finally instructs the Digital Policy Committee on follow-up.
Key terms
The Recommendation defines five terms (pp. 7-8):
- AI system: in summary, a machine-based system that pursues stated or implied objectives and works out from the inputs it receives how to produce outputs, such as predictions, content, recommendations or decisions, that can affect physical or virtual settings. The OECD notes that systems differ in how autonomous they are and how far they keep adapting once in use.
- AI system lifecycle: seven phases, from planning and design through data, model building, testing, deployment and operation, to retirement. A decision to retire a system can come at any point during operation.
- AI actors: those who play an active role in the lifecycle, including organisations and individuals that deploy or operate AI.
- AI knowledge: the skills and resources, such as data, code, models, know-how and governance, needed to understand and take part in the lifecycle, including managing risks.
- Stakeholders: everyone involved in or affected by AI systems, directly or indirectly. AI actors are a subset of stakeholders.
The five principles
The principles are for all stakeholders, and AI actors are called on to promote and apply them according to their roles. The text stresses that they are complementary and should be considered as a whole (p. 8).
1.1 Inclusive growth, sustainable development and well-being. Stakeholders should actively pursue beneficial outcomes for people and the planet, such as augmenting human abilities and creativity, including under-represented groups, reducing inequalities, and protecting the environment (p. 8).
1.2 Respect for the rule of law, human rights and democratic values, including fairness and privacy. AI actors should respect these throughout the lifecycle, including non-discrimination, dignity, privacy and data protection, fairness and labour rights. This includes addressing misinformation and disinformation amplified by AI while respecting freedom of expression. To make this real, AI actors should build in safeguards such as the capacity for human agency and oversight, including for use outside the intended purpose and for intentional or unintentional misuse (p. 8).
1.3 Transparency and explainability. AI actors should commit to transparency and responsible disclosure. That means giving meaningful information, appropriate to the context, that helps people understand AI systems, including their capabilities and limitations, makes people aware that they are interacting with AI, including at work, explains where feasible the data, factors and logic behind an output in plain terms, and lets people adversely affected challenge the result (pp. 8-9).
1.4 Robustness, security and safety. AI systems should function appropriately in normal use, foreseeable use or misuse and other adverse conditions without posing unreasonable safety or security risks. There should be ways for a system that risks undue harm or shows undesired behaviour to be overridden, repaired or decommissioned safely, and, where technically feasible, ways to bolster information integrity while respecting freedom of expression (p. 9).
1.5 Accountability. AI actors should be accountable for the proper functioning of AI systems and for respecting the principles, based on their role and context. They should ensure traceability of datasets, processes and decisions so that outputs can be analysed and inquiries answered. They should apply a systematic risk management approach in every phase of the lifecycle on an ongoing basis, and adopt responsible business conduct, including co-operation with suppliers of AI knowledge and resources, users and other stakeholders. The risks named include harmful bias, human rights, safety, security, privacy, labour rights and intellectual property (p. 9).
The five recommendations for governments
Governments that adhere are asked to carry out five recommendations in national policy and international co-operation, with special attention to small and medium-sized enterprises (p. 9).
- 2.1 Invest in research and development, including public investment and encouragement of private investment in interdisciplinary research on technical and on social, legal and ethical issues, and in open-source tools and representative, privacy-respecting open datasets (p. 9).
- 2.2 Foster an inclusive AI-enabling ecosystem, covering data, technologies, infrastructure and ways to share AI knowledge. Governments are asked to consider mechanisms such as data trusts to support safe, fair, legal and ethical data sharing (p. 10).
- 2.3 Shape an interoperable governance and policy environment. Governments should promote an agile environment, consider controlled experimentation, use outcome-based approaches, co-operate across jurisdictions, and review and adapt their own rules to encourage innovation and competition (p. 10).
- 2.4 Build human capacity and prepare for labour market change. This covers skills, a fair transition for workers through training and social protection, and promoting responsible AI at work, with attention to worker safety, job quality, entrepreneurship and broadly shared benefits (p. 10).
- 2.5 Co-operate internationally, including with developing countries, to share AI knowledge, support multi-stakeholder initiatives, promote consensus-driven global technical standards and develop internationally comparable indicators (p. 10).
The 2023 and 2024 revisions
The 2023 revision updated the AI system definition. It clarified that objectives may be explicit or implicit, stressed that inputs may come from humans or machines, made clear that the text covers generative systems that produce content, replaced "real" with "physical" environments and recognised that some systems keep evolving after deployment (p. 4).
The 2024 revision followed a report to the Council that found the Recommendation still highly relevant but in need of clarification. It added attention to misinformation, disinformation and information integrity in the context of generative AI, to uses outside the intended purpose, intentional misuse and unintentional misuse, and to the information AI actors should give for transparency and responsible disclosure. It also added safety language on overriding, repairing or decommissioning systems, emphasis on responsible business conduct with suppliers and users throughout the lifecycle, a call for interoperable policy, and an explicit reference to environmental sustainability. It moved the text on traceability and risk management into the accountability principle (pp. 4-5).
Follow-up
The OECD's Digital Policy Committee, working through a Working Party on AI Governance, is instructed to keep working on AI, develop practical implementation guidance, run a forum for exchanging information and report to the Council no later than five years after the revision and at least every ten years thereafter (p. 11). The OECD runs an AI Policy Observatory (OECD.AI) with a live database of national AI strategies and policies, and an informal network of AI experts (pp. 5-6).