Skip to content
FORGENO ADS GAMESUnlock all $8

Responsible AI & AI Safety · Unit 7 of 8

Unit 7: Governance frameworks and AI law

The NIST AI RMF (Govern, Map, Measure, Manage), ISO/IEC 42001, the EU AI Act risk tiers and timeline, and US federal, state and city rules, including hiring-tool audit laws.

Opens the first lesson of this unit in Responsible AI & AI Safety. Every lesson is open from the start.

Lesson 1: The NIST AI Risk Management Framework

Goal: Use the four functions of the NIST AI RMF to organize an AI risk program.

A framework is voluntary guidance on how to organize work; a law is binding. The most widely used AI framework is the US National Institute of Standards and Technology's AI Risk Management Framework (AI RMF 1.0), released on January 26, 2023. It is voluntary, works for any sector and any size of organization, and is referenced by laws and contracts around the world. As of November 2026 it is still version 1.0; NIST says it is being revised under the July 2025 US AI Action Plan.

The four functions of the NIST AI RMF 1.0
FunctionThe question it answersTypical activities
GovernWho is responsible, and what are our rules?AI policy, roles and accountability, risk appetite, training, rules for suppliers, a culture where people can raise concerns
MapWhat is this system, and what could go wrong, for whom?Purpose and context, intended and foreseeable uses, affected people, legal requirements, benefits and costs, the system's category
MeasureHow big are the risks, and how do we know?Tests, evaluations and metrics (accuracy, error rates by group, privacy, security, robustness), red teaming, monitoring in production
ManageWhat do we do about them?Prioritize and treat risks, decide whether to deploy, plan incident response and recovery, retire systems that cannot be made safe

Govern is cross-cutting: it runs through the other three rather than coming first and stopping. Map, Measure and Manage then repeat across a system's life, from design to retirement. The functions break down into 19 categories and 72 subcategories, and the free companion Playbook suggests actions for each. Organizations write profiles: a "current" profile of what they do now and a "target" profile of where they want to be, which turns the framework into a plan.

Worked example: one deployment, four functions

  1. An invented hospital plans to deploy a vendor's sepsis-alert model that flags patients for urgent review.
  2. Govern: the chief medical officer is named accountable; the AI policy requires clinical validation before go-live; nurses can report bad alerts without blame.
  3. Map: affected people are patients and the nurses who respond; a missed alert can kill, and too many false alerts cause "alert fatigue"; FDA rules on clinical software are checked.
  4. Measure: the model is tested on 12 months of the hospital's own records: it catches 82% of sepsis cases, but only 70% for patients over 80, and sends about 9 false alerts per true one.
  5. Manage: go-live is approved for adult wards only, with a threshold tuned to cut false alerts, monthly monitoring by age group, and a plan to switch alerts off if performance drops.

Govern sets who decides, Map says what is at stake, Measure supplies the evidence, and Manage turns evidence into decisions.

For generative AI, NIST's Generative AI Profile (NIST AI 600-1, July 2024) lists twelve risks that are new or worse with generative models, including confabulation (hallucination), information integrity, information security, data privacy, harmful bias, intellectual property, dangerous or violent content, and environmental impacts, with suggested actions for each under the four functions.

NIST AI RMF
The voluntary AI Risk Management Framework from the US National Institute of Standards and Technology, version 1.0 of January 2023.
Govern (function)
The cross-cutting AI RMF function covering policies, roles, accountability and culture.
Map (function)
The AI RMF function that establishes a system's context, uses, affected people and risks.
Measure (function)
The AI RMF function that analyzes, tests and tracks risks with metrics and evaluations.
Manage (function)
The AI RMF function that prioritizes and treats risks and responds to incidents.
AI RMF profile
A description of an organization's current or target use of the framework, used to plan improvements.
Generative AI Profile
NIST AI 600-1 (July 2024): the AI RMF applied to generative AI, listing twelve risks and suggested actions.

Remember: Govern (who and what rules), Map (context and harms), Measure (evidence), Manage (decisions and response), repeated for the system's life.

Lesson 2: ISO/IEC 42001 and AI management systems

Goal: Explain what an AI management system is, what ISO/IEC 42001 certification shows and how it relates to other frameworks.

ISO/IEC 42001:2023, published in December 2023, is the first international standard for an AI management system (AIMS): the policies, roles, processes and records an organization uses to develop, provide or use AI responsibly. It works like ISO/IEC 27001 does for information security. Unlike the NIST framework, it states requirements ("the organization shall…"), so an organization can be audited and certified against it. As of November 2026 it is still the first edition, with no amendment.

The standard follows the Plan-Do-Check-Act cycle. Its main clauses cover the organization's context and scope, leadership and an AI policy, planning (AI risk assessment, AI system impact assessment and objectives), support (resources, competence, awareness, documentation), operation, performance evaluation (monitoring, internal audit, management review) and improvement (nonconformities and corrective action). Annex A lists 38 reference controls in nine areas, from policies and internal organization through data for AI systems and system life cycle to third-party relationships. A Statement of Applicability records which controls the organization applies and why it excluded any.

Three tools that are often used together
NIST AI RMF 1.0ISO/IEC 42001:2023EU AI Act
What it isVoluntary frameworkCertifiable management-system standardBinding law (a regulation)
Who uses itAny organization, mostly in the USAny organization worldwideProviders and deployers in or serving the EU
What it producesProfiles and practicesAn audited management system and a certificateLegal duties, conformity assessments, fines
Checked bySelf-assessment, sometimes customersAccredited certification bodiesNational market surveillance authorities and the EU AI Office

Companion standards fill in detail: ISO/IEC 22989 (AI concepts and terms), ISO/IEC 23894:2023 (guidance on AI risk management), ISO/IEC 42005:2025 (published May 2025, guidance on AI system impact assessments) and ISO/IEC 42006:2025 (published July 2025, requirements for the bodies that audit and certify AI management systems). Certification normally involves a two-stage audit (documents first, then evidence that the system really runs), a certificate valid for three years, and yearly surveillance audits.

Worked example: scoping an AI management system

  1. An invented HR-software company, 120 staff, sells a candidate-matching feature and uses AI tools internally.
  2. Scope: "the design, development, provision and monitoring of the candidate-matching feature, and internal use of generative AI tools".
  3. Plan: an AI policy signed by the CEO; a named AIMS owner; an AI risk assessment and an impact assessment for candidate matching (who could be unfairly ranked, and how).
  4. Do: controls chosen in the Statement of Applicability, such as data-quality checks, documentation, human review for rejections and supplier reviews for the model vendor.
  5. Check: quarterly bias metrics, an internal audit, and a management review of results.
  6. Act: an internal audit finds that model updates skip the bias test, which is logged as a nonconformity; the fix makes the test a release gate.

Certification shows the management system exists and runs. It does not certify that every model is fair, so the system must keep checking that.

ISO/IEC 42001
The 2023 international standard specifying requirements for an AI management system, against which organizations can be certified.
AI management system (AIMS)
The policies, roles, processes and records an organization uses to manage its AI responsibly.
Statement of Applicability
A record of which reference controls an organization applies, and why any are excluded.
Plan-Do-Check-Act
The improvement cycle behind ISO management systems: plan, implement, check results, then act on them.
Nonconformity
A failure to meet a requirement of the standard or of the organization's own management system.
Certification body
An independent, accredited organization that audits management systems and issues certificates.
ISO/IEC 42005
The 2025 international guidance on assessing the impacts of AI systems on people and society.

Remember: ISO/IEC 42001 certifies that an AI management system exists and runs (plan, do, check, act). It does not certify each model.

Lesson 3: The EU AI Act: risk tiers and timeline

Goal: Classify an AI system under the EU AI Act and state which duties apply and when.

The EU AI Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024. It is a law, with fines, and it reaches beyond Europe: it applies to anyone placing AI systems on the EU market or putting them into service there, to deployers in the EU, and to providers and deployers elsewhere when the system's output is used in the EU. It sorts systems by risk: the higher the risk to health, safety and fundamental rights, the stricter the rules.

EU AI Act risk tiers, as amended, as of November 2026
TierWhat falls in itWhat the law says
Prohibited (Article 5)Harmful manipulation or exploitation of vulnerabilities; social scoring; predicting crime from profiling alone; untargeted scraping of faces for recognition databases; emotion recognition at work or school (except for medical or safety reasons); biometric categorization to infer race, religion, sexual orientation and similar traits; live remote biometric identification in public by police, with narrow exceptionsBanned since 2 February 2025. From 2 December 2026, AI systems that generate child sexual abuse material or non-consensual intimate images of real people are banned too
High-risk (Article 6, Annexes I and III)Safety components of regulated products (Annex I), and uses in Annex III: biometrics, critical infrastructure, education, employment and worker management, access to essential services (such as credit scoring and life and health insurance pricing), law enforcement, migration and border control, justice and democratic processesStrict duties for providers and deployers. Annex III duties apply from 2 December 2027; Annex I duties from 2 August 2028
Transparency duties (Article 50)Chatbots and other systems that interact with people; generators of synthetic content; emotion recognition and biometric categorization; deepfakesTell people they are dealing with AI; mark AI-generated content in a machine-readable way; label deepfakes. Applies from 2 August 2026 (generators already on the market have until 2 December 2026 for marking)
Minimal riskMost AI: spam filters, recommendations, games, inventory forecastsNo new duties beyond AI literacy; voluntary codes encouraged

An Annex III system is not high-risk if it poses no significant risk, for example if it only performs a narrow procedural task or improves the result of work a person has already done, but a system that profiles people is always high-risk. General-purpose AI models (the large models behind chatbots) have their own duties, applying since 2 August 2025: technical documentation, information for downstream providers, a copyright policy and a summary of training content, with extra evaluation, incident-reporting and cybersecurity duties for models with systemic risk. A voluntary General-Purpose AI Code of Practice (July 2025) explains one way to comply.

The dates moved in 2026. The Digital Omnibus on AI (Regulation (EU) 2026/1744), in force since 27 July 2026, pushed the high-risk dates back to December 2027 and August 2028, added the new prohibition, softened the AI literacy duty to "take measures to support" literacy, and added relief for smaller companies. Fines reach €35 million or 7% of worldwide annual turnover for prohibited practices, €15 million or 3% for most other breaches, and €7.5 million or 1% for supplying incorrect information to authorities.

Worked example: classifying four systems

  1. A tool that ranks job applicants for a recruiter in Germany: employment is in Annex III, and ranking people profiles them, so it is high-risk. The provider needs risk management, data governance, documentation, logging, human oversight design, a conformity assessment and EU database registration; the employer must follow the instructions, assign trained overseers and inform workers.
  2. A camera system that reads call-centre workers' facial expressions to score their mood: emotion recognition in the workplace, not for medical or safety reasons: prohibited.
  3. A website chatbot that answers questions about opening hours: transparency duty: it must tell people they are talking to an AI.
  4. An AI that predicts how many loaves a bakery should bake: minimal risk.

Ask in order: is it prohibited? Is it in Annex I or III (and not exempt)? Does it interact with people or make synthetic content? If none apply, it is minimal risk.

Hands-on lab in the room: EU AI Act risk-tier classifier. Classify invented systems into the four tiers and see which article decides each case. This is a simplified teaching tool, not legal advice.

EU AI Act
Regulation (EU) 2024/1689, the European Union's risk-based law on AI systems and general-purpose AI models.
Prohibited practice
An AI use banned outright by Article 5 of the EU AI Act, such as social scoring.
High-risk AI system
Under the EU AI Act, an AI system in a regulated product or an Annex III area, subject to strict duties.
Annex III
The EU AI Act's list of high-risk use areas, such as employment, education, credit and law enforcement.
Conformity assessment
The process of showing that a high-risk system meets the AI Act's requirements before it goes on the market.
General-purpose AI model
A model trained on broad data that can perform a wide range of tasks and be built into many systems.
Digital Omnibus on AI
Regulation (EU) 2026/1744, which amended the AI Act in 2026, moving high-risk dates and adding a prohibition.
Fundamental rights impact assessment (FRIA)
An assessment some deployers of high-risk AI must do before use, including public bodies and those scoring credit or pricing life and health insurance.

Remember: Prohibited, high-risk, transparency, minimal: check them in that order, and check the date each duty applies.

Lesson 4: US federal and state AI rules

Goal: Explain how existing US law, federal policy and state and local AI laws apply to an AI system, including hiring-tool audit laws.

As of November 2026 the United States has no comprehensive federal AI law. That does not mean AI is unregulated. Existing laws apply to AI the same way they apply to people and paper forms: anti-discrimination laws in employment (Title VII, the Age Discrimination in Employment Act, the Americans with Disabilities Act), credit (the Equal Credit Opportunity Act and the Fair Credit Reporting Act) and housing; the FTC Act's ban on unfair and deceptive practices; and sector rules such as the FDA's for medical software and HIPAA for health data.

Reported enforcement shows this. In August 2023 the EEOC settled its first AI-related hiring case, against iTutorGroup, whose software had automatically rejected women aged 55 or older and men aged 60 or older, for $365,000. In December 2023 the FTC banned Rite Aid from using facial recognition for surveillance for five years, after the company's system wrongly flagged customers, disproportionately women and people of color, as suspected shoplifters. In May 2025 a federal court allowed the age-discrimination case Mobley v. Workday to proceed as a nationwide collective action over AI screening tools.

Federal policy has shifted. In January 2025 a new administration revoked the 2023 executive order on AI (EO 14110) and issued EO 14179, focused on removing barriers to AI development; America's AI Action Plan followed in July 2025. Office of Management and Budget memo M-25-21 (April 2025) requires federal agencies to apply minimum risk-management practices to "high-impact" AI. Executive Order 14365 (December 11, 2025) seeks a single national framework and created a Justice Department AI Litigation Task Force to challenge state AI laws it considers burdensome. An executive order cannot itself cancel state laws: only Congress or the courts can, and as of autumn 2026 Congress had passed no preemption law (a ten-year moratorium on state AI laws was stripped from the 2025 budget bill by a 99–1 Senate vote).

A sample of US state and local AI rules. Check each law's current text before relying on it.
PlaceLawStatus (as of November 2026)Main duties
New York CityLocal Law 144 (automated employment decision tools)Enforced since July 5, 2023An independent bias audit within one year before use, a published summary of impact ratios, and notice to candidates at least 10 business days before use
IllinoisAI Video Interview Act; HB 3773 amending the Human Rights Act2020; January 1, 2026Consent for AI analysis of video interviews; notice when AI is used in employment decisions; no discriminatory AI and no ZIP codes as a proxy for protected classes
CaliforniaCivil Rights Council rules on automated-decision systems; privacy agency (CPPA) rules on automated decision-making technology; SB 53October 1, 2025; compliance from January 1, 2027; January 1, 2026Discrimination by automated systems is unlawful under state employment law; pre-use notice, opt-out and risk assessment for significant decisions; large frontier developers publish safety frameworks, and frontier developers report critical safety incidents
ColoradoSB 26-189 (replaced the 2024 Colorado AI Act)Signed May 14, 2026; effective January 1, 2027Notice before automated decision-making technology materially influences a consequential decision; a plain-language explanation within 30 days of an adverse outcome; correction and meaningful human review where commercially reasonable
TexasResponsible AI Governance Act (TRAIGA)January 1, 2026Bans certain intentionally harmful uses; disclosure duties for government agencies and health care; a regulatory sandbox
New York StateRAISE Act (frontier models)Effective January 1, 2027Large frontier developers publish a frontier AI framework and report critical safety incidents within 72 hours

Worked example: one hiring tool, several jurisdictions

  1. An invented employer with offices in New York City, Chicago and Denver uses a vendor's résumé-ranking tool.
  2. Everywhere: federal anti-discrimination law applies, so the employer tests selection rates by sex, race and age (Unit 2) and keeps a human decision-maker; using a vendor's tool does not move liability to the vendor.
  3. New York City: an independent bias audit no more than a year old, its summary on the careers site, and notice to candidates 10 business days before use.
  4. Illinois: notice that AI is used in hiring decisions; check that no feature, such as ZIP code, acts as a proxy for a protected class.
  5. Colorado, from January 1, 2027: notice before use; within 30 days of a rejection influenced by the tool, a plain-language explanation and a way to request human review.
  6. The simplest compliant design meets the strictest rule everywhere: one audit, one notice, one explanation route, one human review process.

Start from existing law, add each place's specific duties, then design one process that meets them all.

Other countries take different routes, and this changes fast: China's rules requiring labels on AI-generated content took effect on September 1, 2025; South Korea's AI Basic Act took effect on January 22, 2026; the United Kingdom has so far relied mainly on its existing regulators applying shared principles. A professional keeps a dated list of the rules that apply to each system and checks it on a schedule.

Automated employment decision tool (AEDT)
Software that substantially assists or replaces human decisions about hiring or promotion; regulated by NYC Local Law 144.
Disparate impact
Discrimination through a neutral-looking practice that falls more heavily on a protected group without adequate justification.
Preemption
When a higher level of law, such as federal law, overrides lower-level laws, such as state laws.
Executive order
A directive from the US President to federal agencies. It cannot by itself repeal laws passed by Congress or the states.
Automated decision-making technology (ADMT)
Technology that makes or substantially influences decisions about people, a term used in California and Colorado rules.
Consequential decision
A decision with a significant effect on a person's access to employment, housing, credit, insurance, health care, education or essential services.

Remember: No federal AI law does not mean no rules: existing law applies everywhere, then add each state's and city's specific duties.

Practise it in Responsible AI & AI Safety

Each lesson ends with three questions that check you can use the idea, not just recognise it. Wrong answers explain why and come back in Review. The unit ends with a 7-question quiz; pass it with 80% or more to master the unit. Hands-on lab: EU AI Act risk-tier classifier.

Every lesson is free: no ads, no lives and no paywall.