Trustworthy AI: the same principles, three official dialects
Lesson 4 of 5 in Core Ethics Principles and Trustworthy AI.
"Trustworthy AI" is what the principles became when institutions needed something an engineer could build toward and an auditor could check. Three official vocabularies dominate, and professionals are expected to switch between them fluently — they are dialects of one language, not rival philosophies.
The EU High-Level Expert Group got there first (Ethics Guidelines for Trustworthy AI, 2019): trustworthy AI must be lawful, ethical, and robust, unpacked into seven requirements — human agency and oversight; technical robustness and safety; privacy and data governance; transparency; diversity, non-discrimination and fairness; societal and environmental well-being; accountability. These guidelines are the EU AI Act’s intellectual ancestry — you can trace Article 14 (oversight) and Articles 8–15 straight back to them.
NIST’s AI RMF (2023) speaks American: seven trustworthiness characteristics — valid and reliable; safe; secure and resilient; accountable and transparent; explainable and interpretable; privacy-enhanced; fair with harmful bias managed. Note the engineering accent: "valid and reliable" leads the list, because a system that doesn’t work is untrustworthy before ethics even enters.
The OECD values-based principles (2019, revised 2024) speak diplomatic: inclusive growth and well-being; human rights and democratic values; transparency and explainability; robustness, security and safety; accountability. Broadest strokes, widest signatures.
| Principle | EU HLEG (2019) | NIST AI RMF (2023) | OECD (2019/2024) |
|---|---|---|---|
Fairness | Diversity, non-discrimination and fairness | Fair — with harmful bias managed | Within “human rights and democratic values”, naming fairness explicitly |
Transparency / explainability | Transparency (incl. traceability and communication) | Accountable and transparent; explainable and interpretable — split into two characteristics | Transparency and explainability (one principle) |
Safety & robustness | Technical robustness and safety | Valid and reliable; safe; secure and resilient — three separate characteristics | Robustness, security and safety |
Privacy | Privacy and data governance | Privacy-enhanced | Within “human rights and democratic values” |
Human oversight | Human agency and oversight — its own requirement | Folded into governance culture (Govern function) rather than a named characteristic | Within human rights values; “human determination” language |
Accountability | Accountability (incl. auditability, redress) | Paired with transparency | Accountability — providers answerable for proper functioning |
Societal well-being | Societal and environmental well-being — its own requirement | Absent as a named characteristic (appears in context guidance) | Inclusive growth, sustainable development and well-being — listed first |
Read the differences as signatures of purpose. The OECD leads with well-being because it writes for governments steering economies. NIST splits robustness into three characteristics and leads with valid and reliable because it writes for engineers who must measure things. The HLEG gives human agency and environmental well-being standalone billing because it writes for a rights-based polity that was about to legislate. Same consensus, three audiences.
One more distinction the exam (and real work) will test: these frameworks describe properties of systems and organizations — they are not yet obligations. The HLEG requirements became binding only where the AI Act translated them into articles. NIST’s characteristics bind only those who adopt them. Trustworthy-AI language in a contract or statute inherits force from the contract or statute — never from the framework itself.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.