How we built AIEI-STD-RAIP-01: from 171 coded obligations to 17 duties

A professional standard is easy to write badly. Gather some principles, phrase them as commitments, publish. The result reads well, flatters everyone who signs it and assesses nothing. We wanted AIEI-STD-RAIP-01 to survive an examination room, which meant starting with evidence rather than with a drafting session.

The constraint behind AIEI-STD-RAIP-01

Two rules governed the whole exercise. Every duty had to be traceable to something real – an existing obligation, a documented failure, or an established practice in a mature profession. And every duty had to be assessable at the level of an individual, because AIEI-STD-RAIP-01 was designed for conformity with ISO/IEC 17024 from the outset. A requirement that cannot be examined has no place in a certification standard, however admirable it sounds.

Six research streams

The research ran as six parallel streams rather than one literature review, so that no single source family could quietly set the agenda.

• The standards corpus. ISO and IEEE work on AI management, risk, bias, transparency and ethically aligned design.
• Regulation. The EU AI Act, the Council of Europe framework convention, and the places where obligations fall on people rather than on systems.
• Competency frameworks. Public and private competence models already in use for adjacent roles.
• Credentials and benchmarks. Existing AI credentials, read alongside the codes of conduct of mature professions – engineering, accountancy, medicine – for how duties are written when someone has to enforce them.
• Jurisdictional sampling. The EU, United States, United Kingdom, Council of Europe, China, India, Japan, Korea, Singapore, Brazil, the African Union, Gulf states and UN and UNESCO instruments.
• Failure evidence. Documented AI incidents and harms, read with one question in mind: what could the person in the room have done differently?

The obligation register

The streams produced a register of 171 coded rows. Each row recorded the source, the obligation as actually stated, who it falls on, whether it is binding or advisory, and whether an individual could be examined on it.

The coding mattered more than the collecting. A great deal of what circulates as responsible AI guidance turns out, once coded, to be addressed to organisations, or to be unassessable at the individual level, or to be the same obligation appearing in nine jurisdictions in nine wordings.

From 171 to 17

Four filters did the reduction:

• Merge the same obligation where it recurs across jurisdictions, keeping the clearest formulation.
• Remove obligations that fall on an organisation and cannot be discharged by a person.
• Remove anything that cannot be assessed in an examination.
• Remove anything that would require the professional to make a legal judgment they are not qualified to make.

Sixteen core candidate duties survived. The drafting stage brought the published figure to 17, organised in six parts: Duty, Competence, Responsible practice, Honesty and openness, Accountability, and Integrity under pressure.

Why the standard is flat

The obvious alternative was a tiered family: foundation, advanced, expert. We looked at it and rejected it. Tiers in this field mostly measure seniority rather than duty, and the duties themselves do not scale – a junior engineer who spots a problem owes the same obligation to raise it as the head of the department. One standard, one examination, with role differences handled in a guidance annex.

What we deliberately did not do

• No reproduced standard text. Every duty is written from scratch. Each carries a Sources line for provenance only, not as a requirement to hold or quote those documents.
• Not a job analysis. The register maps obligations, not tasks. A job analysis is a separate instrument and is being handled separately.
• Not another organisational framework. Nothing in the standard tells a company how to govern AI. It states what a person owes, wherever they work.

Still under review

The document moved from working draft 0.1 through 0.5 and 0.6, each version shortened rather than extended, and is now in structured review with comments logged per section and ranked by priority. The version number on the published standard is the real one. We do not round it up.

Read next

• Why responsible AI needs a standard for people – the argument AIEI-STD-RAIP-01 was written to answer.
• ISO 17024 and ISO 42001 – how certifying a person and certifying a management system fit together.

AIEI-STD-RAIP-01 is the standard behind the Responsible AI Professional certification.