Frameworks, Standards and Competing Principles: Ethics, Law and Policy

Principles tell you what to care about. Frameworks, standards and laws tell you how organisations organise that care, and sometimes what they must do. These are easy to mix up, and the mix-up has real costs: a team can think a voluntary framework is a legal duty, or that a legal duty is the whole of ethics. This page separates them, introduces three that you will hear about often, and then looks at what to do when two good principles point in opposite directions.

Facts about laws and standards on this page come from official pages checked on 2026-10-12. They change, so each one is dated, and the last section shows you how to check them yourself.

What you will learn

• What the NIST AI Risk Management Framework is for and what its four functions (Govern, Map, Measure, Manage) ask a small team to do.
• How the EU takes a risk-based approach, what is adopted and what changed in the dates, and what ISO/IEC 42001 is meant to be.
• How voluntary principles, technical standards and laws differ, and how a company's stated principles can differ from its practice.
• How to reason when two principles conflict, and how to compare a precautionary approach with an experimenting one without picking a winner.
• How to find the primary source of a framework or law and check its version and date.

The NIST AI Risk Management Framework: purpose

NIST is the National Institute of Standards and Technology, a United States government agency. It released the AI Risk Management Framework (AI RMF) 1.0 on 2023-01-26. It is intended for voluntary use: no law makes anyone follow it. Its purpose is to help organisations build trustworthiness into how they design, develop, use and evaluate AI systems. It was written in the United States, but nothing stops an organisation in any country from using it as a checklist for its own thinking.

Two corrections to things you may read online. There is no "AI RMF 2.0": on 2026-10-12 the NIST page said version 1.0 is being revised, with no timeline and no new version published. And a concept note from April 2026 concerns a separate profile for critical infrastructure, not a new version of the framework. NIST also published a Generative AI Profile (document number NIST AI 600-1) in July 2024, which applies the framework to generative AI and is also voluntary. A framework like this does not give you answers; it gives you a structure to ask good questions in an orderly way.

Govern, Map, Measure and Manage

The framework is organised around four functions. The one-line descriptions below are our plain-language wording, not NIST's text; read NIST's own page for the official description.

Notice how closely this matches the eight-step worksheet from Everyday AI Ethics: Who Benefits and Who Could Be Harmed: roughly, the worksheet's steps 2 and 3 are mostly Map, step 5 is Measure, steps 6 and 8 are Manage, and step 7 is Govern. You do not need to adopt the NIST structure to run the worksheet. But when a client or employer says "we use the NIST framework", you now know the four words and can ask which function their evidence belongs to.

The EU's risk-based approach

According to the European Commission, the European Union has a binding law on AI, the AI Act (Regulation (EU) 2024/1689); dates and details are in the Legal note below. Its main idea is that the rules depend on the risk of the use, not on the technology. The European Commission describes four levels, and the best way to remember them is by example of use.

Legal note (as of 2026-10-12; not legal advice): According to the European Commission's pages, the AI Act entered into force on 1 August 2024 and applies in stages. The Commission's timeline lists: 2 February 2025 for definitions, AI literacy and the first prohibitions; 2 August 2025 for obligations on general-purpose AI models; 2 August 2026 for most rules and their enforcement, including the transparency duties; 2 December 2026 for a further prohibition on AI systems that generate non-consensual sexual deepfakes and child sexual abuse material; 2 December 2027 for high-risk rules for systems in the listed areas; and 2 August 2028 for high-risk rules for AI in products already covered by other EU product law. The last two dates were moved by amendments, known as the AI omnibus. The Commission's pages describe those amendments as adopted and in force since 27 July 2026, with a Commission proposal in November 2025 and a political agreement in May 2026. Not every Commission page is up to date: one FAQ page still called the changes a proposal when we checked. A separate, wider proposal touching other EU laws was reported by a law firm on 1 July 2026 as still pending; treat it as not law. The Act can apply to organisations outside the EU that place AI on the EU market or use it there; scope is a legal question. The Act is binding law; the Commission's guidelines and codes of practice are separate documents. Older summaries show different dates for the high-risk rules. Read the current official text (the EU Official Journal via EUR-Lex, and the Commission's AI Act Service Desk) and check the current official text for your country and sector.

Two lessons for a beginner. First, a hiring tool or an exam-marking tool is an example of an area the Act names, not an automatic classification; the details sit in the text. Second, dates move. When you quote one, quote the source and the date you checked it, as this page does.

ISO/IEC 42001: a management-system standard

ISO and IEC are international standards bodies. ISO/IEC 42001 is described by standards bodies as a standard for an AI management system. A management system is not a piece of technology. It is the set of policies, roles, processes and records an organisation uses to run something in a disciplined way and to improve it over time. For AI, it means the organisation can show who is responsible for AI use, how risks are assessed, how changes are controlled and how problems are handled.

What this does not mean: a standard about an organisation's processes does not certify that any particular model is safe, fair or accurate. We could not read the standard's page at the ISO site while preparing this page, so we do not give its edition details or certification arrangements here. If an organisation says it follows ISO/IEC 42001, ask what exactly is in scope (which teams, which AI uses), what evidence exists, and who checked it. Look up the current edition and status in the ISO catalogue.

Voluntary principles, technical standards and laws

These three are different tools. Mixing them up is the most common error on this topic.

A useful sentence to remember: a principle tells you why, a standard tells you how, and a law tells you whether you may. Ethical concern, legal obligation and company policy are three separate things; a use can be legal and still ethically doubtful, or ethically sound and not yet covered by any rule.

A company's principles and its practices

Many organisations publish AI principles. That is a good start, but a published statement is a promise, not a result. To compare principles with practice, look for evidence of a different kind: what the organisation does, records and lets others check.

Invented scenario:

Northwind Learning (an invented company) sells an AI homework helper. Its website says:
  "We are transparent. We protect your privacy. People always stay in control."
A teacher, Lena in Berlin, wants to know if this is true. She finds:
  - The help page says nothing about when chats are deleted.
  - There is no way to reach a person about a wrong answer.
  - A short document lists who at the company is responsible for the product.
  - There is a public page where the company reports problems it found and fixed.

The invented company is not "lying" here. Lena has found a gap between a statement and the evidence she can see, and she knows which questions to ask. That is the fair way to put it: describe what you found, ask for what is missing, and decide in proportion to the stakes. A company that answers these questions clearly and in writing is showing something real; one that cannot answer is telling you something too. Do the same checking for your own organisation's principles.

When two principles conflict

Principles are not ranked in advance. Sometimes honouring one costs another. The practical method is to name the conflict, state what each principle protects, consider more than one response, and write down which you chose and why.

Invented scenario:

Elif runs a small online language school. Learners practise speaking, and an AI feature gives
feedback on pronunciation. The feature works less well for some regional accents because few
examples of them were used to build it. To improve, it would need recordings of learners'
voices. Many learners did not expect their voices to be kept.

Two principles are in tension. Privacy says: do not keep voice recordings that people did not expect to give. Fairness and inclusion say: learners with less-represented accents deserve feedback that works as well for them. Neither is silly. Here are two reasonable responses.

Neither response is the correct one for everybody. Elif's decision would depend on her learners, her ability to run consent and deletion properly, and the alternatives available. What makes a response defensible is not which principle it favours, but that it names the conflict, protects the people it puts at risk, says what it will check and by when, and has a named owner.

Precaution and experimentation

A related conflict runs through many AI decisions: how cautious to be when the effects are uncertain. A precautionary approach asks for evidence of safety before wide use. An experimentation approach allows careful, limited use so that evidence can be gathered in real conditions. Both are used by serious, caring people.

Both camps can be right in different situations. A workable middle is to scale the care to the stakes: for low-stakes, reversible uses, try and watch; for high-stakes, hard-to-reverse uses, require evidence, human oversight and a way to appeal before going wide. You can ask of any plan: what evidence would we need, what would a small trial look like, who is protected during it, and what would make us stop? These are the same questions as steps 5, 6 and 8 of the decision pattern.

Finding the primary source of a framework or standard

Articles, social posts and short summaries often describe frameworks wrongly or give out-of-date versions. The primary source is the document or page published by the body that created it. Follow this routine.

Template:

SOURCE CHECK
1. Name the thing exactly (e.g. "NIST AI Risk Management Framework", not "the NIST AI law").
2. Find the issuing body's own site by name: do not trust a link in a summary.
3. Record: title, version or edition, date, status (draft, in force, revised, withdrawn).
4. Read the part you are relying on; note what is a recommendation and what is a requirement.
5. Check for newer versions or amendments, and note the date you checked.
6. If you cannot reach the official text, say so and treat secondary summaries as "reported".

Where to start, by name. For EU rules: the European Commission's AI regulatory-framework page, the AI Act Service Desk, and EUR-Lex for the legal text. For the US framework: NIST's AI Risk Management Framework page. For principles: the OECD.AI website (OECD AI Principles) and UNESCO's website (Recommendation on the Ethics of Artificial Intelligence). For standards: the ISO catalogue, where you check the current edition and status yourself. Search for each body by name and prefer the official domain.

A short worked check. Someone tells you "NIST has released AI RMF 2.0". Following the routine: the issuing body's own page (NIST's AI Risk Management Framework page), as read on 2026-10-12, lists version 1.0 and says it is being revised, with no timeline. The April 2026 item is a concept note for a critical-infrastructure profile. So the claim does not match the primary source, and your note should say what you found and the date.

Where this shows up at work

Customers and partners increasingly ask whether a supplier "aligns with" a framework or standard. A person who can ask "which version, which parts, who checked, and as of what date?" turns a marketing phrase into a verifiable statement. In a small business, the same four NIST words can organise a one-page AI-use note. In any organisation, the habit of recording the source, version and date beside each rule you quote saves arguments later.

Common mistakes

• Treating a voluntary framework as a legal duty, or the reverse. Fix: for each item ask whether it is a principle, a standard, a framework or a law, and who can enforce it.
• Quoting dates from an old summary. The EU dates for high-risk rules moved. Fix: quote the source and the date you checked, and read the current official text.
• Treating a standard as proof that a model is safe. A management-system standard concerns an organisation's processes. Fix: ask what is in scope and what evidence exists.
• Taking a company's principles page as proof of practice. Fix: look for evidence such as retention periods, contact routes, owners and public reports.
• Declaring a winner in a principle conflict. Fix: name both principles, compare at least two responses, and record the reasons and the review date.

Try it yourself

1. Easy. Label each as a principle (P), standard or framework (S), or law (L), and say if it is binding: (a) the OECD AI Principles; (b) the EU AI Act; (c) the NIST AI Risk Management Framework; (d) ISO/IEC 42001.
2. Medium. Choose a school you know, real or imagined, that uses an AI tutoring tool, and apply Govern, Map, Measure and Manage to it. Write one sentence for each function that is specific enough to check. Then say which of the four your school is likely weakest at and why.
3. Hard. Invented scenario: Amara's council assistant also offers voice input, so residents who cannot type or read well can speak their questions. The voice is sent to an outside service to be turned into text. Some residents worry about being recorded. Identify the two principles in tension. Describe two reasonable responses, compare them in a table (what each favours, main risk, who is accountable, what would change your mind), and say which you would choose and why, in no more than eight sentences. Do not claim one response is the only right one.

Answers

Easy. (a) P, not binding. (b) L, binding within its scope. (c) S (a framework), voluntary. (d) S (a management-system standard), voluntary unless someone, such as a customer or a rule, asks you to follow it. Pass: all four labelled correctly with the binding status. Weak: calling OECD or NIST "laws", or ISO/IEC 42001 a certificate of safe AI.

Medium. A strong answer: Govern: the head teacher names Haruka as accountable and writes a rule on pupil data. Map: list pupils, parents and next year's teacher, and the risk of wrong hints. Measure: check 20 hints each month against a teacher's solutions. Manage: pause the tool if two wrong hints a week are found, and give parents a route to ask for correction. The answer then picks one weak function with a reason, such as "Measure, because no one has time to check hints". Pass: four checkable sentences and an honest weakest point. Weak: restating the four words without anything specific.

Hard. A strong answer names privacy and accessibility/inclusion. Response A: keep the voice service but add a notice, an opt-out with a typing or desk alternative, deletion of recordings after processing, and a contract limit on the supplier's use. It favours inclusion with privacy protected by notice and choice; the risk is that residents do not read the notice; Amara is accountable; evidence that few residents opt out or that the supplier keeps audio would change the choice. Response B: process speech on the council's own systems only, or offer phone calls with a staff member instead. It favours privacy; the risk is cost and slower service; accountable: Amara and the head of IT; evidence that waiting times harm residents would change it. A good choice gives reasons that refer to the residents, the alternatives and the stakes, and sets a review date. Pass: two principles correctly named, two real alternatives, a fair table, a reasoned choice, and a stated condition for changing it. Weak: one option presented as obviously right, or no accountable person.

Revisit Prompt Engineering for AI: Beginner on personal data in chats: both a voice service and a chat service raise the same question about what leaves your control.

Summary

• The NIST AI RMF 1.0 (2023-01-26) is a voluntary framework with four functions: Govern, Map, Measure, Manage. There is no "2.0"; a revision was announced with no new version as of 2026-10-12.
• The EU AI Act is a binding, risk-based law with four levels. Its dates have been amended, so quote the source and date, and read the current official text.
• ISO/IEC 42001 is described as a management-system standard for AI. It concerns an organisation's processes and is not proof that a model is safe.
• Principles say why, standards say how, laws say whether you may. A company's published principles are a promise to be checked against evidence.
• When principles conflict, or precaution and experimentation pull apart, name the conflict, compare at least two responses, and record the owner, the condition for changing your mind, and a review date.
• Always find the primary source and record its version, date and status.

Next: Where Bias Comes From: Data, Design, Proxies and Feedback Loops starts the next chapter by showing where in an AI system unfair differences between groups can enter.