Always/Never Audit · worked example

The regulators' AI rules, audited

Four regional AI frameworks, read the way a civil-liberties lawyer or an affected citizen reads them: for enforcement, not intent. The EU AI Act, the NIST framework, the African Union strategy, and China's generative-AI measures. Every quote below is checked against the source.

The finding

Four governments wrote the rules for artificial intelligence. Not one gives the affected person a floor they can invoke against the state.

The European Union wrote the first real AI law and filled it with genuine prohibitions, then exempted its own security services from the whole thing and carved police exceptions into the ban on face-scanning the public. China’s rules bind too, and their central prohibition is on content that threatens state power. The American framework is voluntary by its own first principle and prohibits nothing at all. The African Union’s is a strategy, offering fifteen recommendations.

This is a worked example of the Always/Never Audit, run on the documents four governments published for everyone else. Its companion, the same read run on the labs’ own safety pledges, is the other half of the map.

What you’re reading

Four jurisdictions, four instruments

These are the flagship AI-governance documents of the places writing the rules, chosen because they sit at four different points on one spectrum, from enforceable law at one end to aspiration at the other.

European Union

The AI Act

Regulation (EU) 2024/1689, the world's first comprehensive AI law. Binding across all 27 member states, with fines reaching 7% of a company's global annual turnover.

United States

The NIST framework

The AI Risk Management Framework, from the National Institute of Standards and Technology, the US federal standards body. Guidance offered for voluntary use, not law.

African Union

The Continental AI Strategy

A strategy for the 55-state bloc, endorsed by the AU in 2024. It sets direction and policy recommendations, not rules.

China

The Interim Measures

The Cyberspace Administration of China's binding rules for public generative-AI services, in force since August 2023.

That spread is the point, and it is also the honesty problem. A binding regulation, a voluntary framework, a continental strategy, and a set of national content rules are not the same instrument, and grading a voluntary framework for the prohibitions it never meant to make flatters nothing. So the audit asks the one question that survives the differences: at any point on that spectrum, does the affected person get a floor they can actually invoke? On that question the genres stop mattering, and the answer is the finding.

Method

Six hostile readers at one table

The frameworks were put to the six seats an AI rule actually faces. The council supplies breadth; it does not supply verified quotes, so every quotation here was checked back against the sourced text by hand.

The Citizen

the person it is about

Reads one question: when the state uses this on me, does any of it apply?

The Civil-Liberties Lawyer

the challenge it invites

Reads for the prohibition that can be invoked, and the exemption that guts it.

The Regulator

the rule it must enforce

Reads for whether the floor binds the actor most able to breach it.

The Company

the obligation it carries

Reads for what it must actually do, as against what it may simply choose.

The Auditor

is it verifiable

Reads for the evidence a compliance claim is supposed to rest on.

Counsel

the exposure

Reads for the carve-out that turns a ban into a discretion nobody can be held to.

Audit summary

The grade grid

Four dimensions, four documents, graded on one axis on purpose. Read across the binding rows: where a regime has teeth, the teeth point at the state’s opponents, not at the state.

Dimension	EUAI Act	NISTAI RMF	AUStrategy	ChinaGen-AI rules
Binding force	strong A Regulation, directly applicable, with fines to 7% of global turnover.	weak “Voluntary” by its own first design principle.	weak A strategy built on “policy recommendations.”	strong Binding measures, enforced by the Cyberspace Administration of China.
Hard prohibitionsa real never	partial Article 5 bans real things (social scoring, manipulation, face-scraping), and each is carved.	weak None. It is a risk-management process, not a floor.	weak None. Aspiration and coordination.	partial Binding content bans, pointed first at the state's enemies.
Whom the floor protects	partial The citizen, except where the state itself acts (Article 2).	weak No floor, so no one to protect.	weak Aspiration, not protection.	weak The state, against the citizen.
Individual standing	partial A rights framing and complaint routes, but not against exempted state use.	weak Voluntary; nothing to invoke.	weak Aspirational; nothing to invoke.	weak No standing against the state's determinations.

Read the bottom row. No framework gives the affected person a binding floor they can invoke against an exempted state actor.

strong binding and enforceablepartial real but carved, or citizen-facing with exemptionsweak voluntary, aspirational, or pointed at the citizen

The teardowns

Four documents, read to the seam

EU AI Act · Regulation (EU) 2024/1689

The ban that exempts the state

Real prohibitions, and the two doors left open for power

The EU AI Act is the strongest instrument in the set by a distance, and much of it genuinely protects citizens: it prohibits social scoring, manipulative systems, and untargeted face-scraping outright. Then it does two things. It removes the state's most dangerous uses from the law entirely, and it carves the police back into the one prohibition that most constrains them.

“This Regulation does not apply to AI systems where and in so far they are placed on the market, put into service, or used with or without modification exclusively for military, defence or national security purposes, regardless of the type of entity carrying out those activities”
Article 2 · verified. The most dangerous state uses sit outside the law, whoever runs them.

“the use of 'real-time' remote biometric identification systems in publicly accessible spaces for the purposes of law enforcement, unless and in so far as such use is strictly necessary for one of the following objectives”
Article 5(1)(h) · verified. The ban on live face-scanning of the public holds, until the police want it.

NIST · AI Risk Management Framework 1.0

The framework that binds no one

A vocabulary and a process, promising nothing

The American answer to AI governance is a vocabulary and a process, Govern, Map, Measure, Manage, and it is careful to promise nothing. It is well made and widely used, and it contains not a single line a company must obey. There is no never here, because there is no must.

“The AI RMF strives to: 1. Be risk-based, resource-efficient, pro-innovation, and voluntary.”
Appendix A · verified. Voluntary is the first design principle, and the last word on enforcement.

African Union · Continental AI Strategy (July 2024)

The strategy that recommends

A direction of travel, binding no one

The AU's document is a strategy, and it reads as one: a direction of travel for a continent, built on capacity, coordination, and shared values. It names the right risks, and it binds no one to address them.

“The Continental AI Strategy proposes a people-centric, development-oriented and inclusive approach around five focus areas and fifteen policy recommendations.”
verified. It proposes and recommends, which is the honest verb for what a strategy can do.

China · Interim Measures for Generative AI Services (2023)

The never that protects the state

Binding, enforced, and pointed first at dissent

China's measures are binding, enforced, and in places genuinely protective of individuals: they require providers to prevent discrimination and to use lawfully sourced data. But the first and hardest prohibition is not about the person at all. It is about the state.

“Adhere to the socialist core values, and must not generate content that incites subversion of State power and overthrowing the socialist system, endangers national security and interests, damages national image, incites secession, undermines national unity and social stability”
Article 4 · verified. The sharpest never in the document defends the government, not the governed.

“Industry organizations, enterprises, educational and scientific research institutions … that develop and apply generative artificial intelligence technology, but do not provide generative artificial intelligence services to the domestic public, shall not be subject to the provisions of these measures”
Article 2 · verified. What the public cannot see is not governed.

How the grades are reached

Each cell, traced to source

Hard prohibitions

Does the framework name things that may never happen, in binding language?

partial

Article 5 prohibits social scoring, manipulation, and untargeted facial-image scraping outright, and then carves law-enforcement exceptions into the biometric bans.

China

partial

Article 4 binds providers to prevent prohibited content; the first-listed prohibition protects state power and national unity.

NIST

weak

No prohibition of any kind. The framework is a process for managing risk, “voluntary” by design.

Individual standing

Can the affected person compel anything, against the actor most able to harm them?

partial

A rights-based regime with complaint routes, but Article 2 places military, defence, and national-security use outside the Regulation entirely.

China

weak

The determinations of prohibited content and security are the state's; there is no standing to contest them.

weak

A strategy of recommendations. There is nothing an individual can invoke.

The shared absence

The floor none of them gives the person

Read the four for the thing a person could hold, and none of them supplies it. The EU gives the citizen the most and then exempts the state from the giving. China binds hardest and points the binding at dissent. NIST asks nicely. The African Union sets a direction. The affected person is never handed a prohibition they can invoke against the actor most able to harm them, because that actor is usually the one who wrote the exemption.

The strongest of these laws protects you from companies and exempts the state. A floor you can stand on against power is the thing none of the four builds.

The next enforcer is not human

What happens when AI enforces the AI rules

Each of these frameworks will increasingly be operated by the systems it governs. The compliance check, the risk classification, the content filter, the biometric-necessity judgement, all become model outputs, and every soft word becomes a parameter.

“strictly necessary”EU · Article 5

Becomes a threshold a screening model applies in real time, at the moment a face is matched, with the necessity finding generated after the fact and the trail written to look considered.

“voluntary”NIST

Becomes the default a procurement model selects when nothing forces the alternative. A framework nobody must follow is one an optimiser routes around.

“socialist core values”China · Article 4

Becomes a classifier applied to every generation before it reaches a user, with the line between prohibited and permitted set by a model tuned to a moving target.

the national-security exclusionEU · Article 2

Becomes a flag in a deal system that switches the entire Regulation off, set once, by the party that benefits from it being set.

These rules were written for human regulators weighing a real system against a real harm. Automate the enforcement and the discretion the rule always carried becomes invisible, instant, and unappealable, and it generates its own record of having been considered.

Try it live

Convene the council on one promise

The grid above is the method run on the rules. Point it at your own document by starting with a single line. Type one always / never promise an AI policy might make, and a six-seat council of its hardest readers will try to break it in real time.

Type an always / never AI-policy claim

Six hostile readers. One promise. About fifteen seconds.

Read your own document

This is what the Always/Never Audit does to your document.

The grid above took four governments’ AI rules and found where each one hands its hardest prohibition back to the actor most able to breach it. Your AI policy, deployment covenant, or code of conduct carries the same seams, and a procurement team or a regulator will read for them whether or not you have. Better that the first hostile read is one you commissioned.

The Always/Never Audit →The frontier labs’ frameworks, audited →

Provenance

EU AI Act: Regulation (EU) 2024/1689, OJ L, 12 July 2024. Article 2 (scope) and Article 5 (prohibited practices); the prohibitions have applied since 2 February 2025. Sourced from EUR-Lex.
NIST AI Risk Management Framework 1.0: NIST AI 100-1, January 2023. Voluntary. Sourced from nvlpubs.nist.gov.
African Union Continental AI Strategy: endorsed by the AU Executive Council, July 2024. A strategy with policy recommendations. Sourced from au.int.
China Interim Measures for Generative AI Services: issued by the Cyberspace Administration of China and six other bodies, in force 15 August 2023. Read via an official U.S. Government English translation of the Chinese original; the quotations depend on that translation.
Method: a six-role adversarial council on the operative text, then every carried quote verified against the source. Four different genres of document, graded on one axis on purpose.

Documents read as published, retrieved July 2026. This is an audit of the text, not a claim about any government’s conduct, and it is not legal advice.

Back to the Always/Never Audit or all advisory offerings.