Are AI principles technical specifications?

Usually they are broader commitments that need more concrete requirements.

Can principles apply to purchased AI?

Yes; choices about procurement, configuration and use can implement or undermine them.

Do all principles reduce to numerical metrics?

No. Some require qualitative evidence, judgment and meaningful participation.

AI Glossary / Governance, ethics and safety / Advanced

AI principles

Definition

AI principles are high-level commitments intended to guide how artificial intelligence is developed, deployed and used. They express what an organization or community considers important, such as accountability, fairness, transparency, privacy and safety. Principles provide direction, but their practical meaning must be translated into responsibilities, requirements and evidence for the particular system and people affected.

What it means

A principle such as transparency does not specify one universal explanation that every user must receive. A person using a writing tool, an operator investigating a failure and someone affected by a consequential decision need different information. Applying the principle requires attention to those needs.
There is also no single list called the AI principles that every organization uses identically. The OECD and UNESCO provide influential frameworks with overlapping concerns but different structures. An organization should identify which commitments it has adopted and what they mean in its context, rather than displaying familiar words without explaining their consequences.

How it works

Principles become useful when they influence decisions. A team can connect a stated commitment to a requirement, an accountable owner, an implementation measure and evidence of whether it works. For example, a commitment to accountability could require a named person to investigate reported failures and a record of the resulting action.
The OECD's values-based principles address beneficial development, human rights and fairness, transparency, robustness and accountability. UNESCO's recommendation also emphasizes human dignity, oversight and environmental considerations, with policy actions to support implementation. These frameworks provide guidance rather than a ready-made test for every product.
Some commitments create tensions in particular situations. Explaining a decision need not require exposing another person's private records. Teams should document how they address the competing concerns and what evidence supports their choice, instead of claiming that all principles are automatically satisfied.

Three practical examples

A learning platform turns a transparency commitment into a clear explanation of where AI assists content creation and how readers can report an error. It checks whether that explanation is understandable to its audience.
A document-processing team connects a privacy commitment to defined input permissions and retention practices. It verifies actual handling rather than assuming a statement on a policy page changes system behavior.
A product team interprets a fairness commitment by identifying who could be disadvantaged and examining relevant outcomes. It explains why the chosen measures address those concerns, while acknowledging what they do not measure.

Think of it this way

Think of principles as the reasons behind rules. A principle explains the direction you want decisions to take; a procedure explains what people should do in a specific situation. Evidence shows whether their actions achieved the intended result. Confusing these three levels can leave a strong statement with little practical effect.

Worked example

Fictional Willow Learning adopts three commitments for an AI-assisted exercise generator: understandable disclosure, accountable correction and appropriate oversight. The team writes two operational requirements under each commitment, giving six requirements in total.
For disclosure, it requires a visible notice and a user-comprehension check. For correction, it requires a reporting route and a named response owner. For oversight, it requires subject review before release and a way to withdraw a faulty exercise. These are the team's invented requirements, not a standard prescribed by the cited frameworks.
At review, all six requirements have documents describing them. Only four have implementation evidence: the notice, reporting route, response owner and withdrawal mechanism. The document-completion rate is 6 / 6 = 100%, while the implementation-evidence rate is 4 / 6, approximately 66.7%.
Even the latter number is not an ethics score. It does not show whether users understand the notice or whether an exercise is accurate. The team therefore reports the two missing checks explicitly. It also tests the implemented reporting and withdrawal routes, because evidence that a feature exists differs from evidence that it works reliably when needed.

Illustrative case study

Fictional Cedar Answers publishes a principle promising that users can understand significant AI-assisted recommendations. Its first implementation exposes a long technical explanation filled with internal terminology. Readers cannot tell what information influenced the recommendation or how to question it.
The team revisits the principle with user representatives. It creates an explanation organized around the decision's purpose, relevant input categories, important limitations and an accessible review route. It tests comprehension and checks that the explanation does not reveal another person's information.
The principle itself has not changed; its implementation has become more specific and testable. The case illustrates why a principle should guide an ongoing process of interpretation, feedback and improvement rather than serve as a label attached after development.

Common misconceptions and mistakes

Treating principles as self-executing: written commitments require implementation and accountability.
Assuming identical wording means identical practice: organizations can interpret broad terms differently.
Turning a checklist total into a moral verdict: completing items does not establish every relevant outcome.
Ignoring tensions between commitments: a useful decision explains how concerns were addressed.
Equating adoption with legal compliance: applicable obligations and voluntary commitments need to be distinguished.

Business use

Use AI principles to make expectations explicit before difficult product decisions arise. Connect them to design reviews, procurement questions, release evidence and incident handling. Ask what would change in a real decision because a principle exists. If no one can answer, the commitment may need clearer interpretation, ownership or implementation. Review it with people who experience the system's effects, not only those building it.

Check your understanding

Why is Willow's 100% document rate insufficient?

Answer: Two required checks lack implementation evidence, and documents do not establish outcomes.

How does a principle differ from a procedure?

Answer: It states a guiding commitment rather than a specific sequence of actions.

Why test explanations with users?

Answer: A technically detailed explanation may still fail to inform its intended audience.

Related terms

Responsible AI
Accountability
Transparency
Impact assessment
AI ethics
governance
human oversight
stakeholder engagement

Learn next

Choose one principle and write a practical requirement for a hypothetical AI service. Identify its owner, the evidence you would collect and a realistic situation in which another principle could complicate the decision.

Sources

OECD, AI Principles overview, values-based principles and policy recommendations, updated in 2024. https://oecd.ai/en/ai-principles UNESCO, Recommendation on the Ethics of Artificial Intelligence, values, principles and implementation context. https://www.unesco.org/en/artificial-intelligence/recommendation-ethics NIST AI RMF 1.0, Govern function, connecting organizational principles with risk-management practices. https://airc.nist.gov/airmf-resources/airmf/5-sec-core/ The Willow Learning requirements and Cedar Answers case are original teaching illustrations.

Previous: AI infrastructure Next: AI risk management