AI Ethics Principles Compared: OECD, UNESCO and the EU

AI ethics principles are the short lists of commitments that governments, international bodies and companies publish to say how artificial intelligence ought to be built and used. There are now dozens of them, and most explainers pick one and recite it. This piece does the more useful thing: it sets the major frameworks side by side, shows where they agree and where they quietly part ways, and ranks them by the question that matters most to anyone building, funding or regulating AI. Which of these principles can actually be enforced?

What AI ethics principles are, and why there are so many lists

A principle is a statement of intent, not a procedure. It names a value, such as fairness or transparency, and leaves the method to whoever adopts it. That made principles easy to write, and by 2019 at least 84 public and private initiatives had published their own.

When researchers at ETH Zurich mapped that corpus, they found the lists were less varied than they looked. Anna Jobin, Marcello Ienca and Effy Vayena identified convergence around five principles: transparency, justice and fairness, non-maleficence, responsibility and privacy. They also found substantive disagreement about what each of those words means in practice and how it should be implemented. Agreement on the vocabulary, in other words, arrived well before agreement on the substance. For the plainer question of what ethics in artificial intelligence means as a field, we cover that separately.

The five frameworks that matter

OECD AI Principles (2019, updated 2024)

The OECD set describes itself as the first intergovernmental standard on AI. Adopted in May 2019 and updated in 2024, it contains five values based principles: inclusive growth, sustainable development and well-being; human rights and democratic values, including fairness and privacy; transparency and explainability; robustness, security and safety; and accountability. Five recommendations to governments sit alongside them, covering research investment, an enabling ecosystem, interoperable policy, labour market transition and international cooperation.

Its influence runs further than its text. The OECD notes that the European Union, the Council of Europe, the United States and the United Nations use its definition of an AI system and lifecycle in their own legislation and guidance.

EU Ethics Guidelines for Trustworthy AI (2019)

The European Commission's High Level Expert Group published its guidelines in April 2019 with seven key requirements: human agency and oversight; technical robustness and safety; privacy and data governance; transparency; diversity, non-discrimination and fairness; societal and environmental well-being; and accountability. Beneath those requirements the full document sets out four ethical principles: respect for human autonomy, prevention of harm, fairness and explicability. It also asks readers to acknowledge the tensions between them, which few principle sets admit exist.

UNESCO Recommendation on the Ethics of AI (2021)

In November 2021, 193 Member States adopted UNESCO's Recommendation, the broadest list in circulation. It rests on four values (human rights and dignity; peaceful, just and interconnected societies; diversity and inclusiveness; environmental flourishing) and ten principles: proportionality and do no harm; safety and security; privacy and data protection; multi-stakeholder and adaptive governance; responsibility and accountability; transparency and explainability; human oversight and determination; sustainability; awareness and literacy; fairness and non-discrimination. What sets it apart is that it pairs the principles with policy action areas, from gender to data to the environment.

Council of Europe Framework Convention (2024)

The Council of Europe Committee of Ministers adopted its Framework Convention on artificial intelligence and human rights, democracy and the rule of law on 17 May 2024. It was opened for signature in Vilnius on 5 September 2024 as the first international legally binding treaty on AI, with Andorra, Georgia, Iceland, Norway, Moldova, San Marino, the United Kingdom, Israel, the United States and the European Union among the first signatories. It enters into force three months after five signatories, at least three of them Council of Europe member states, have ratified it.

A corporate example: Google's AI Principles

Companies publish principles too, and Google's are the most studied. They were set in 2018, after the company declined to renew the Project Maven drone imagery contract following an employee petition. We return to what happened to them in 2025 below, because it is the clearest lesson in this whole subject.

Where the lists agree

Map the three public frameworks against Jobin's five themes, add human oversight and well-being, and the overlap is close to complete.

Seven themes, three frameworks, no gaps. If a reader needs a working list of the core AI ethics principles, this table is it: transparency, fairness, safety, accountability, privacy, human oversight and well-being.

Where the lists quietly differ

The differences are in emphasis and in what each body chose to add.

• UNESCO adds two principles nobody else lists. Awareness and literacy treats public understanding as an ethical requirement, not a communications task. Multi-stakeholder and adaptive governance makes participation itself a principle.
• UNESCO also adds proportionality. Its first principle says an AI system must not go beyond what is necessary to achieve a legitimate aim, a test closer to human rights law than to engineering.
• The OECD's 2024 update moved towards operations. It now asks for mechanisms that let a system be overridden, repaired or decommissioned safely, and it names misinformation and disinformation amplified by AI.
• The EU treats the environment as a requirement, not a growth objective, and asks for the tensions between principles to be acknowledged rather than resolved by fiat.
• Corporate sets tend to include a balancing test. Google's revised principles say the company will proceed where likely benefits substantially exceed foreseeable risks. That is a decision rule, and it places the weighing in the company's hands.

From principle to law: how binding each one is

This is the axis most comparisons skip, and it is the one that decides whether a principle survives commercial pressure.

1. Corporate pledge. Fully voluntary, and revisable without notice. On 4 February 2025 Google removed a section of its principles that listed four applications it would not pursue, including weapons and surveillance. Human Rights Watch told the BBC the unilateral change showed why voluntary principles are not an adequate substitute for regulation and binding law.
2. Intergovernmental recommendation. The OECD and UNESCO texts carry political weight and shape national policy, but neither is enforceable against a company.
3. Recital in binding law. The EU AI Act's recital 27 recalls the seven HLEG principles explicitly as non-binding, without prejudice to the Act's legal requirements, and names them as a basis for drafting codes of conduct. The principles inform the law; the obligations live in its articles.
4. Treaty. The Council of Europe Convention obliges ratifying states to put its commitments into domestic law, which is the furthest any set of AI ethics principles has travelled towards enforcement.

Whether more of this ladder should be climbed by statute is a live argument, which we set out in should AI be regulated.

Why principles alone are not enough

Brent Mittelstadt of the Oxford Internet Institute made the sharpest case in 2019. AI ethics had converged on something resembling the four classic principles of medical ethics, but medicine has things AI development lacks: common aims and fiduciary duties, a professional history and norms, proven methods for turning principles into practice, and robust legal and professional accountability. Consensus on high level principles, he warned, can hide deep political and normative disagreement.

The answer is not fewer principles but machinery behind them: named owners, impact assessments, oversight with authority, and records that outlast the people. We set out that machinery, and the documented cases where ethics boards failed, in AI ethics and governance, and compare the public scaffolding in AI accountability frameworks from GAO, NIST and the OECD.

Choosing principles for your own organisation

For a research group, startup or agency writing its own list, the comparison above suggests a short discipline.

1. Borrow the wording, do not invent it. The OECD and EU texts are what regulators, auditors and partners already recognise. A bespoke vocabulary only adds translation work.
2. Give every principle an owner and an artefact. Transparency becomes a model card; fairness becomes a bias test with a threshold; oversight becomes a person with the right to pause the system.
3. Mark which principles are already law where you operate. In the EU several of them now sit behind binding articles; elsewhere they may still be voluntary.
4. Write down what you will not do, and how a change would be announced. Google's 2025 revision shows that the exclusions are the part of a principles page most likely to move.
5. Review the list on a fixed cycle. The OECD revised its own text after five years; yours will age faster.

If the case for doing any of this still needs making internally, the documented harms in why AI ethics is important make it more persuasively than any list does.

Where the principles get argued over in person

Principles are drafted by one community and applied by another, and the two rarely share a schedule. EX Future Summit puts them in one: its program runs nine tracks, with AI Ethics, Government and AI and Society alongside research, finance and health tech. It runs on 18 to 20 November 2026 online and in rooms in Las Palmas and Bali, as a single thirty hour broadcast between the two islands, and online attendance is free for verified researchers and students.

FAQ

What are the five core AI ethics principles?

The most cited answer comes from a 2019 review of global AI ethics guidelines, which found convergence on transparency, justice and fairness, non-maleficence, responsibility and privacy. The OECD's own five principles are a different but overlapping list, so it is worth saying which one you mean.

Are the OECD AI Principles legally binding?

No. They are an OECD Council Recommendation, which commits adhering governments politically but cannot be enforced against a company. Their influence comes through national laws and guidance that adopt their definitions and language.

Is there a legally binding international treaty on AI ethics?

Yes. The Council of Europe Framework Convention on AI was opened for signature on 5 September 2024 and is the first legally binding international treaty in the field. It takes effect once five signatories, including at least three Council of Europe member states, have ratified it.

Can a company change its AI principles?

Yes, at any time. Corporate principles are voluntary. Google rewrote its AI principles in February 2025 and removed its commitment not to pursue weapons and surveillance applications, which is why binding law and contracts carry more weight than a published list.