OECD AI principles: the five values for trustworthy AI

Each AI system that a business deploys raises the question of whether the people it impacts should trust it. This includes pricing models, fraud detection systems, and chatbots that handle benefit queries. Establishing trust demands early decisions concerning data quality and human oversight. It also calls for a clear process to resolve errors when the system fails. Trust cannot be added as a feature at the end of a project.

The OECD AI Principles constitute the first intergovernmental standard for trustworthy AI, adopted in 2019 and updated in 2024. They outline five values-benefit, fairness, transparency, robustness, and accountability-and present five policy recommendations aimed at governments.

These principles are not legally binding on their own. Nevertheless, they provide the foundation for many regulations and corporate policies created since their introduction.

The OECD AI Principles constitute a formal effort to record standards that governments have committed to follow. These principles were adopted in 2019 and were updated in 2024. They acted as the first intergovernmental standard for AI and still influence current regulations and corporate policies. If you examine an AI governance framework, much of its content is derived from these principles.

Shipshape Data constructs data and AI systems. We encounter the OECD framework on a frequent basis. Clients rarely mention it by name, but the framework addresses the same questions that determine whether an AI project passes a formal audit. This guide explains the principles and identifies their target audience. It also covers the changes in the 2024 update and provides instructions on how to move these concepts from policy documents into team workflows.

An overview of the OECD AI Principles

These principles give a shared definition of the term "trustworthy AI". The Organisation for Economic Co-operation and Development agreed on them in May 2019. It marked the first occasion when a coalition of national governments signed a common standard for artificial intelligence. Before that date, many companies and research organisations had issued ethics charters, yet those documents did not create a consensus among different countries.

The framework is divided into two separate parts. The first part includes five value-based principles that describe the characteristics of a trustworthy AI system. The second part offers five recommendations for governments on how to establish the necessary conditions for responsible AI. These values are relevant to anyone who builds or operates such systems, whereas the recommendations target those who formulate national policy. Distinguishing the two parts is important because practical implementation occurs almost entirely within the first section.

Their origins and the scope of their influence

OECD members negotiated these principles prior to their broader adoption. The nations that adhere to them account for a large share of global trade and investment. In the same year, the G20 employed these principles for its own AI guidance. Such wide reach renders the document relevant to a London consultancy and its clients, as well as to a regulator in Paris. If your business operates in Europe, North America, or much of Asia, the regulations you must comply with were shaped by this document.

These principles are not legally binding. The OECD does not impose fines. Rather, these principles create a standard vocabulary that formal laws later adopt. The EU AI Act, national AI strategies, and many corporate policies employ these same five values. Learning them now avoids having to study the same concepts under different titles later.

How the two parts are connected

We split this material on purpose. Technical teams cannot deliver responsible AI alone, and governments cannot produce it only through mandates. The values set a target for developers. The recommendations give governments ways to back that target with funding, infrastructure, and practical regulations. Most organisations only adopt the first part, but the second part is valuable because it shows why the regulatory landscape keeps evolving.

The five values that define trustworthy AI

You should commit these five values to memory. They are relevant for training large models as well as for creating small classifiers for email filtering. Each value identifies a particular way an AI system can fail. Combined, they address most of the practical issues that arise in these systems.

• Inclusive growth, sustainable development and well-being. The system ought to benefit people and the environment rather than merely the financial records of its creators. This means tackling factors that are often ignored, such as the energy needed for training and whether the benefits reach the individuals who bear the risks.
• Human-centred values and fairness. AI systems must respect basic rights and avoid discrimination. They should aid human decision-making, not replace it. If a model rejects a loan application, a human should be able to understand the reason and contest the decision, instead of having the decision hidden.
• Transparency and explainability. Anyone affected by an AI decision should be told that the technology was used. They should also receive a clear explanation of how that decision was reached. This value aims to prevent the excuse "The algorithm said so" from being used as justification.
• Robustness, security, and safety. The system must operate reliably under all expected conditions and resist attempts to compromise or deceive it. This category covers model drift, adversarial inputs, and unusual edge cases.
• Accountability. A person or team must be answerable when an AI system causes harm. The model itself cannot be held responsible. This value is often missed, but it is the first thing auditors look for.

These points outline a system built to help users rather than to extract data from them. The system permits questioning, stays stable when faced with complex situations, and has a designated owner. These are governance requirements, not technical ones. They produce technical effects, which makes them hard to add to an existing system later. More details on the underlying ideas can be found in our glossary entries on what AI ethics means and responsible AI.

Transparency and accountability need specific decisions about data and ownership. You must make those decisions early in a project. If you did not decide at the start, you must now modify your existing systems.

The five policy recommendations, and who they are for

The second half of the framework is aimed at governments rather than engineering teams. You will not apply these points directly unless you work in public policy. They remain relevant because they shape the environment in which your systems function. Moreover, public-sector clients are increasingly required to show that their work complies with these standards.

There are five, and they read as a to-do list for a state that wants responsible AI without strangling the innovation that produces it:

• Invest in AI research and development. Public funds ought to back the development of technical capabilities. These funds should likewise back research on safety and trustworthiness so that those capabilities do not cause harm.
• Foster a digital ecosystem for AI. Governments are urged to construct the data infrastructure, provide compute access, and create the open standards that enable more than a handful of giant firms to participate.
• Shape an enabling policy environment. Rules ought to promote experimentation and adoption while still safeguarding the public, a balance that is more difficult than it appears and explains why regulation is repeatedly rewritten.
• Build human capacity and prepare for labour market change. AI alters the nature of work. This recommendation emphasises education, retraining, and support for transitions instead of overlooking the impending disruption.
• Cooperate internationally. If each country creates incompatible AI rules, safety and trade will suffer. Governments ought to keep communicating with each other to preserve consistency.

Governments need to balance two objectives: promoting AI development and shielding the public from its risks. When a government follows this guidance, the benefits for businesses become practical and significant. You obtain clearer regulations and more predictable enforcement. This lowers the risk of creating a product that a regulator could later prohibit. If you need a broader view of how these organisations interact, our note on what AI governance is outlines the current landscape.

Changes in the 2024 update

The 2019 principles were released before the phrase "large language model" became widely known. By 2024 this time gap turned into a significant issue, prompting the OECD to revise the text. The primary update assists organisations that based their governance on the original version because the five core values stay the same. The changes concentrate on interpreting those values to encompass systems that were uncommon five years ago.

The revised framework explicitly includes generative AI, foundation models, and reinforcement learning. Those systems simultaneously break several of the previous comfortable assumptions. A model trained on a large slice of the public internet has murky provenance. A system that can generate convincing text, images, or audio on demand can also generate convincing lies at scale. And an architecture whose reasoning is genuinely difficult to inspect places real strain on the transparency and explainability value.

The parts worth flagging

A few of the updates are worth calling out because they change what a governance review has to check:

• Misinformation and synthetic content. This update tackles the risk of AI generating false information and the need to track the origin of content. Provenance and watermarking are now expected features instead of optional additions.
• Data governance under scrutiny. When a model trains on vast, loosely sourced datasets, questions about consent, copyright, and bias in the training data cease to be academic. The revision places greater emphasis on getting the data foundation right, which is the part we spend most of our time on.
• Environmental cost. Training and operating large models uses energy. The update acknowledges this fact rather than assuming that computing resources are cost-free.
• Safety across the whole lifecycle. The focus widens from deployment time to the entire lifespan of a system. This includes the processes of retraining and repurposing the technology.

This is the practical interpretation. If you have already set up governance based on the 2019 principles, you do not need to discard your work. You must determine how each value applies to the generative systems you use, since the original language did not address these systems in detail. Governance frameworks have generally failed to keep pace with current technology. The 2024 revision is the OECD's acknowledgment of this gap and an attempt to update the standards.

How to apply the principles in your organisation

A framework that stays on a slide does not protect anyone. You must translate the five values into actions that your teams perform automatically. Clients often struggle with this task because they see it as a single large compliance project rather than a sequence of small, achievable steps. You should not try to fix every issue at once. Create a map first, then address the most significant gaps.

Begin with a map instead of a policy

Identify every AI system currently in use before drafting new regulations. You will likely find undocumented systems within marketing tools or spreadsheet macros. Evaluate each system against the five established values and record any deficiencies. For example, a customer-facing model that lacks an explanation process does not meet transparency standards. A pipeline without a designated owner fails the accountability requirement. This inventory ensures you direct your resources towards actual gaps rather than focusing on the most visible issues.

Assign an owner to every principle

The most common failure we observe is the diffusion of responsibility. When people assume someone else is managing transparency, the task is often neglected. You should assign each value to a specific person or team and record these assignments. For example, the data governance team might manage transparency while the security team manages robustness. A senior leader should be responsible for accountability, as this role cannot be handled by a committee that meets only quarterly. If you do not name specific owners, these principles remain goals that will not hold up during an audit.

Build the checks into how you ship, not after

Principle alignment works best when integrated into the development lifecycle rather than acting as a final hurdle that teams lack time to clear. This requires carrying out fairness testing during the model training phase instead of after deployment. It involves documenting a system's limitations before release so users understand its constraints. It also requires establishing human oversight for any AI function that significantly impacts individuals. When engineers can link specific architectural decisions to defined values, abstract concepts become practical and governance is no longer seen as an administrative burden added to technical tasks. Our glossary entry on ethical AI provides more detail regarding the practical application of fairness testing.

A practical point from this work. Most of what enables these principles is basic data management: understanding the origin of your data, who is responsible for it, how it moves through systems, and whether you can trace a result back to its original source. Teams that have not established these fundamentals often find the OECD values very difficult to meet. These values are not excessive, and you cannot explain a decision if you cannot explain the data it relies on. If you establish a solid data foundation, implementing most of the framework simply involves documenting your existing processes.

Where the principles fit alongside real regulation

It is important to grasp the specific purpose of the OECD framework. It does not function as a law that imposes fines. It serves as a shared foundation for developing stricter rules. The EU AI Act provides the actual enforcement via its risk tiers and penalties. National strategies from the UK, the US, and other nations introduce their own specific requirements. Reading the OECD principles first makes the later rulebooks easier to comprehend. You will see the same five core ideas appear in various formats.

Most organisations ought to follow this sequence. Become familiar with the five values sufficiently to identify them in any context. Apply these values to audit your current operations. When mandatory regulations appear in your market, you will be expanding an existing governance structure instead of launching a new project under a deadline. The main advantage of adopting a non-binding framework early lies in the preparation it offers before compliance becomes a legal requirement.

Where to go next

The OECD AI Principles describe trustworthy AI in a practical manner. These principles contain five values: benefit, fairness, transparency, and accountability, together with policy recommendations for governments. The 2024 update adds generative systems while preserving the original framework. These principles are not legally binding, yet they influence the development of regulations.

Begin by auditing your current AI systems against the five values. Prioritise systems that interact with customers or affect significant decisions. Designate a specific individual to be responsible for each principle. Incorporate these evaluations into your development process instead of treating them as a final check. Evaluate your data infrastructure to confirm it can meet the requirements of this framework.

We devote our time to working on that final section.

Transitioning AI from a pilot phase to a system capable of passing an audit demands more than effective models. You require data that is reliable and traceable. If you need to meet this requirement, talk to us to obtain an objective assessment of how your current foundation compares to the standards you must meet.

When the reading becomes a roadmap question, our AI strategy consultants will honestly tell you what the work requires and what it depends on.

Frequently asked questions