AI governance is no longer a nice-to-have. Laws are now in force, customers ask hard questions, and boards want proof that AI is under control. The problem is that there are dozens of frameworks, standards and laws, and they all sound alike.

This guide walks through 30 of the most important ones in plain English. For each, you will learn what it is, who it is for, and why it matters. No legal background needed.

The 30 are grouped into six families:

  1. Management & Governance — how an organisation runs AI responsibly
  2. Risk, Trust & Lifecycle — how to spot and manage AI risk from idea to retirement
  3. Principles & Policy Frameworks — the values that shape AI rules worldwide
  4. IEEE Practice Standards — engineering methods for ethical system design
  5. Laws, Treaties & Codes — the binding and voluntary rules governments have set
  6. Security, Data & Assurance — protecting AI, its data, and proving it works

You do not need all 30. Most organisations start with one management standard, one risk framework, and the laws that apply where they operate. Read on, and you will know which ones fit you.

Part 1: Management & Governance

These five frameworks answer one big question: how does an organisation run AI responsibly, from the boardroom down to the project team?

1. ISO/IEC 42001 — AI Management System

ISO/IEC 42001 is the world's first certifiable management system standard for AI. Published in 2023, it does for AI what ISO 27001 does for information security. It gives you a repeatable system for governing every AI project, not just one.

An "AI management system" is simply a set of policies, roles, processes and records. It covers how you approve AI uses, assess risks and impacts, manage data, monitor performance, and handle suppliers. It follows the familiar Plan-Do-Check-Act cycle, so you keep improving over time.

The standard applies to anyone who builds, provides or uses AI. That includes a bank buying an AI tool, not just tech companies. Size does not matter either; a small firm can scale it down.

Its biggest draw is certification. An accredited auditor can check your system and issue a certificate. That certificate is fast becoming a trust signal in sales deals and procurement. It also gives you a strong head start on laws like the EU AI Act, since many requirements overlap.

Annex A lists 38 controls, such as AI policy, impact assessment and data quality. You pick the ones that fit your risks and explain why.

Who should care: compliance leads, CIOs, and any organisation that wants to prove its AI is well managed.

2. ISO/IEC 38507 — Governance of AI by the Board

ISO/IEC 38507 is written for the people at the very top: directors, board members and senior executives. Published in 2022, it explains what the governing body must do when the organisation uses AI.

The key message is simple. Using AI does not reduce the board's accountability. If an AI system makes a bad decision, the organisation and its leaders still own the outcome. The standard helps leaders oversee AI without needing to be technical experts.

It builds on ISO/IEC 38500, the long-standing standard for IT governance. It asks leaders to evaluate how AI is used, direct its use through policy and strategy, and monitor results. It also covers setting a risk appetite for AI, defining who is accountable, and making sure the right controls exist.

The standard highlights issues that are new with AI. These include decisions made with less human involvement, reliance on data quality, and the speed at which AI can scale mistakes. It also pushes leaders to consider the culture and values behind AI use.

Think of it as the board-level partner to ISO/IEC 42001. ISO 42001 tells managers how to run the system. ISO 38507 tells leaders how to steer and oversee it.

Who should care: board members, CEOs, company secretaries and governance advisers.

3. NIST AI RMF 1.0 — AI Risk Management Framework

The NIST AI Risk Management Framework is one of the most widely used AI governance tools in the world. It was released in January 2023 by the US National Institute of Standards and Technology. It is voluntary, free, and designed for any organisation, in any sector or country.

The framework is built around four functions:

  • Govern — build a culture, policies and roles for managing AI risk
  • Map — understand the context, purpose and possible impacts of each AI system
  • Measure — test and track risks using suitable methods and metrics
  • Manage — prioritise risks and act on them, including deciding not to deploy

Govern sits across the other three, because good risk work depends on leadership and culture.

The framework also defines what "trustworthy AI" looks like. Its seven traits are: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed.

NIST provides a companion Playbook with suggested actions for every category. This makes it practical, not just theory. Many organisations use the AI RMF as their day-to-day risk method, and ISO/IEC 42001 as the management wrapper around it.

Who should care: risk managers, AI product teams and anyone wanting a flexible, non-certifiable starting point.

4. NIST AI 600-1 — Generative AI Profile

NIST AI 600-1 applies the AI RMF to generative AI, the kind of AI that writes text, creates images, and powers chatbots. NIST published it in July 2024, after generative AI spread rapidly across workplaces.

A "profile" is a version of a framework tailored to one use. This one asks: what risks are unique to, or made worse by, generative AI? It identifies twelve of them. Examples include confabulation (confidently made-up answers), harmful or violent content, data privacy leaks, and information integrity problems such as deepfakes and misinformation.

Other risks on the list include intellectual property issues, information security, harmful bias, and dangerous capabilities related to chemical or biological threats. It also flags environmental impact, value chain risks from third-party components, and over-reliance by humans on AI outputs.

For each risk, the profile offers more than 200 suggested actions. These map back to the Govern, Map, Measure and Manage functions. Examples include red-teaming models, labelling AI content, tracking where training data came from, and having plans for incidents.

If your organisation uses tools like chatbots, copilots or image generators, this profile is the most practical public guide to their specific risks.

Who should care: teams deploying or buying generative AI, and anyone writing a GenAI use policy.

5. ISO/IEC 42005 — AI System Impact Assessment

ISO/IEC 42005 explains how to assess the impact of an AI system on people and society. It was published in 2025. While risk management asks "what could go wrong for us?", impact assessment asks "who could this system affect, and how?"

The standard guides you through the full process. First, describe the system: its purpose, its users, its data and how it works. Then identify the people and groups it might affect, including those who never use it directly. Next, consider the possible benefits and harms, both intended and unintended.

These harms can involve fairness, privacy, safety, human rights, jobs, or the environment. The standard also asks about misuse and foreseeable failures. It encourages you to document everything clearly and to repeat the assessment when the system changes.

It also explains when to carry out an assessment and how to fit it into existing processes. It pairs neatly with ISO/IEC 42001, which requires impact assessments but does not explain in detail how to do them.

Impact assessments are also becoming a legal expectation. The EU AI Act requires fundamental rights impact assessments for some high-risk uses, and South Korea's AI Basic Act encourages them for high-impact AI. ISO/IEC 42005 gives you a recognised method to meet those expectations.

Who should care: product owners, ethics and privacy teams, and anyone deploying AI that affects people.

Part 2: Risk, Trust & Lifecycle

This group covers the technical foundations: how to manage AI risk, what makes AI trustworthy, how AI systems are built and maintained, and the shared vocabulary that ties it all together.

6. ISO/IEC 23894 — AI Risk Management Guidance

ISO/IEC 23894, published in 2023, shows how to manage risks that are specific to AI. It builds on ISO 31000, the general risk management standard used by thousands of organisations worldwide.

That link is its main strength. If your company already has a risk process, you do not need to start again. ISO/IEC 23894 shows you how to extend that process to cover AI.

The standard follows the familiar risk cycle. You set the context, identify risks, analyse and evaluate them, treat them, and then monitor and review. At each step, it explains what is different about AI.

So what is different? AI systems learn from data, so poor or biased data creates risk. They can behave unpredictably in new situations. Their decisions can be hard to explain. They can also change over time as data or models drift.

The standard includes helpful annexes. One lists common AI risk sources, such as lack of transparency, complexity, and level of automation. Another shows how risk management fits into each stage of the AI lifecycle.

ISO/IEC 23894 is guidance, not a certifiable standard. It works well alongside ISO/IEC 42001, which requires AI risk assessment, and the NIST AI RMF, which shares many of the same ideas.

Who should care: enterprise risk managers, internal auditors, and teams who want AI risk to fit their existing risk register.

7. ISO/IEC TR 24028 — Trustworthiness in AI

ISO/IEC TR 24028, published in 2020, is a technical report that explores what makes AI trustworthy. "TR" means it is an overview and discussion paper, not a set of requirements.

It was one of the earliest ISO documents on AI. Much of its thinking now runs through later standards, which makes it a useful background read.

The report looks at the main qualities people expect from trustworthy AI. These include reliability, resilience, robustness, transparency, explainability, controllability, safety, security and privacy. It also looks at bias and fairness.

It then explores the threats to trust. These include hidden bias in data, unpredictable behaviour, attacks on AI systems, and a lack of clear accountability. It also discusses the challenge of explaining how complex models reach their decisions.

Alongside the problems, it surveys possible responses. Examples include better testing, transparency measures, human oversight, and careful data handling. It also points out where more standards work was needed, which helped shape later documents.

The report does not tell you exactly what to do. Instead, it gives you a shared way of thinking and talking about trust. That helps when technical teams, managers and regulators need to discuss the same issues.

Who should care: people new to AI governance, policy writers, and anyone building a trustworthy-AI policy from first principles.

8. ISO/IEC 5338 — AI System Life Cycle Processes

ISO/IEC 5338, published in 2023, describes the processes for building and running an AI system, from first idea to final retirement. It extends the well-known software and systems life cycle standards to cover AI.

Why is a separate AI life cycle needed? Because AI is not built like normal software. Its behaviour comes partly from data, not just code. Models must be trained, tested and retrained. Performance can drift after release as the real world changes.

The standard keeps the usual life cycle stages, such as planning, design, building, testing, deployment, operation and retirement. Then it adds or adjusts processes for AI. Key additions include knowledge acquisition, data management, model building and continuous validation.

It also places strong focus on what happens after go-live. Teams need to monitor models, detect drift, and decide when to retrain or retire them. Many AI failures happen here, long after launch.

The standard helps engineering teams build good governance into everyday work. Risk checks, documentation and approvals become part of the process, rather than a last-minute review.

It connects well with ISO/IEC 42001, which asks for a defined AI life cycle, and with ISO/IEC 23053, which describes how ML systems are structured.

Who should care: engineering leads, MLOps teams, and project managers running AI delivery.

9. ISO/IEC 23053 — Framework for AI Systems Using Machine Learning

ISO/IEC 23053, published in 2022, explains how AI systems that use machine learning are put together. It gives a common map of the parts of an ML system and how they connect.

Machine learning is the most common type of AI today. Instead of following hand-written rules, the system learns patterns from data. That simple idea involves many moving parts, and this standard names and describes them.

It covers the main building blocks. These include data collection and preparation, model training, the trained model itself, evaluation, and how the model is used to make predictions. It also explains different learning approaches, such as supervised, unsupervised and reinforcement learning.

The standard describes the ML pipeline step by step. It shows where data enters, where models are trained and tested, and where results come out. This helps people see where risks and controls belong.

It is not a rulebook. You will not find requirements or controls here. Instead, it is a reference model that gives teams, auditors and managers a shared picture.

That shared picture is valuable. When an auditor asks about your "training data" or "model evaluation", everyone knows what is meant. It pairs naturally with ISO/IEC 22989, which defines the vocabulary.

Who should care: auditors and risk teams who need to understand ML systems, and technical teams documenting their architecture.

10. ISO/IEC 22989 — AI Concepts and Terminology

ISO/IEC 22989, published in 2022, is the dictionary of AI standards. It defines the key terms and concepts used across the whole family of ISO AI standards.

This may sound basic, but it solves a real problem. People use words like "AI system", "model", "bias" or "explainability" in very different ways. Without shared meanings, policies, contracts and audits become confusing.

The standard defines a wide range of terms. These include AI agent, machine learning, neural network, training data, and many more. It also explains key ideas such as transparency, explainability, robustness, controllability and trustworthiness.

It goes further than a word list. It describes the AI ecosystem: the different roles involved, like AI provider, AI producer, AI customer and AI partner. It also outlines the AI system life cycle and the main types of AI technology.

Other standards depend on it. ISO/IEC 42001, 23894 and 5338 all use the terms it defines. So if you plan to adopt any ISO AI standard, this is a smart first read.

It is also useful beyond ISO. Teams writing internal AI policies can borrow its definitions so that everyone, from lawyers to engineers, means the same thing.

Who should care: anyone starting an AI governance programme, policy writers, legal teams and trainers.

Part 3: Principles & Policy Frameworks

Before there were AI laws, there were AI principles. These five frameworks set out the values that governments and organisations agree AI should respect. Many later laws are built on them.

11. OECD AI Principles

The OECD AI Principles were the first intergovernmental standard on AI. They were adopted in May 2019 and updated in May 2024 to reflect the rise of generative AI. Dozens of countries have signed up, including all OECD members and several non-members.

The principles have two parts. The first sets out five values for trustworthy AI:

  • Inclusive growth, sustainable development and well-being
  • Respect for human rights and democratic values, including fairness and privacy
  • Transparency and explainability
  • Robustness, security and safety
  • Accountability

The second part gives five recommendations to governments. These include investing in AI research, building an inclusive AI ecosystem, creating flexible policy, preparing workers for change, and cooperating across borders.

The principles are short and high-level. They do not tell you how to build a system. Their power is in their reach. They shaped the G20 AI Principles, and their definition of an "AI system" was adapted into the EU AI Act.

The OECD also runs OECD.AI, a free online observatory. It tracks national AI policies, offers a catalogue of tools for trustworthy AI, and monitors AI incidents around the world.

For organisations, the principles are a strong anchor for an AI ethics policy. If your values line up with the OECD, they will broadly line up with most national rules too.

Who should care: policy teams, ethics leads, and anyone drafting an organisation-wide AI policy.

12. UNESCO Recommendation on the Ethics of AI

The UNESCO Recommendation on the Ethics of Artificial Intelligence is the first global standard on AI ethics. It was adopted in November 2021 by all 193 UNESCO member states. That makes it the most widely agreed AI ethics text in the world.

Its focus is on human rights and human dignity. It argues that AI should benefit all of humanity, not just a few countries or companies.

The Recommendation rests on four core values: human rights and dignity; living in peaceful and just societies; diversity and inclusion; and a healthy environment. It then adds ten principles, including proportionality, safety, privacy, transparency, accountability, and human oversight.

What sets it apart is its practical policy reach. It covers eleven areas of action, such as gender equality, education, culture, health, the environment and the economy. It pays special attention to developing countries and to the risk of AI widening global inequality.

UNESCO has also built tools to put it into practice. The Readiness Assessment Methodology helps countries check how prepared they are for ethical AI. The Ethical Impact Assessment helps evaluate specific AI systems, especially in public procurement.

For businesses, it offers a useful lens on social impact, inclusion and sustainability, which risk-focused frameworks sometimes overlook.

Who should care: governments, public bodies, NGOs, and companies with strong ESG or human rights commitments.

13. Singapore Model AI Governance Framework

Singapore's Model AI Governance Framework is one of the most practical AI governance guides ever published. It was first released in 2019, with a second edition in 2020. It was created by Singapore's data protection regulator and its media and technology authority.

Its aim is simple: turn high-level ethics into steps a business can actually take. It rests on two guiding principles. AI decisions should be explainable, transparent and fair. And AI systems should be human-centric.

The framework covers four practical areas:

  • Internal governance — clear roles, responsibilities and risk controls
  • Human involvement — deciding how much human oversight each AI decision needs
  • Operations management — handling data, models and testing responsibly
  • Stakeholder communication — being open with customers and users

Its approach to human involvement is especially useful. It helps you choose between human-in-the-loop, human-over-the-loop, and fully automated decisions, based on the severity and likelihood of harm.

Singapore has kept building on it. A companion guide includes real company case studies. In 2024, it released a separate Model AI Governance Framework for Generative AI. It also created AI Verify, an open-source testing toolkit.

The framework is voluntary and sector-neutral. That makes it a great template for any organisation, anywhere, that wants practical guidance rather than legal requirements.

Who should care: businesses starting their AI governance journey, especially in Asia, and anyone wanting worked examples.

14. UK AI Management Essentials (AIME)

AI Management Essentials, or AIME, is a free self-assessment tool from the UK government. It was published in draft for public consultation in November 2024 by the Department for Science, Innovation and Technology. The government published its response to that consultation in February 2026.

AIME is designed for organisations that want to check their AI management basics without hiring consultants. It is aimed especially at startups and small to medium-sized businesses. Larger firms can also use it for single teams or divisions.

The tool is a simple questionnaire, mostly with yes or no answers. It covers three broad areas: internal processes, managing risks, and communication. Questions ask about things like AI policies, record-keeping, fairness, impact assessment, data management, and how you handle issues raised by users.

AIME does not create new rules. Instead, it distils key ideas from three major sources: ISO/IEC 42001, the NIST AI RMF, and the EU AI Act. So by working through it, you also take a first step toward those bigger frameworks.

It is important to know what AIME is not. It is not a certification and it is voluntary. It also looks at your organisation's processes, not at whether a specific AI product is safe.

The UK has explored using AIME in public sector procurement. For suppliers selling AI to government, being familiar with it could become an advantage.

Who should care: small businesses, startups, and UK suppliers looking for a quick, low-cost health check.

15. EU Ethics Guidelines for Trustworthy AI

The EU Ethics Guidelines for Trustworthy AI were published in April 2019 by an independent High-Level Expert Group set up by the European Commission. They became the ethical foundation for the EU AI Act.

The guidelines say trustworthy AI should be lawful, ethical and robust. They then set out seven key requirements:

  1. Human agency and oversight — AI should support human decisions, with proper oversight
  2. Technical robustness and safety — systems should be reliable, secure and have fallback plans
  3. Privacy and data governance — protect personal data and ensure data quality
  4. Transparency — make systems traceable and explainable, and tell people when they interact with AI
  5. Diversity, non-discrimination and fairness — avoid unfair bias and ensure accessibility
  6. Societal and environmental well-being — consider sustainability and impact on society
  7. Accountability — enable auditing and offer ways to fix harm

In 2020, the expert group released a practical companion called ALTAI, the Assessment List for Trustworthy AI. It turns the seven requirements into a checklist that teams can work through, and it is available as a free online tool.

The guidelines are voluntary. But many of their ideas, such as human oversight and transparency, now appear as legal duties in the EU AI Act. Reading them helps you understand why the Act asks for what it does.

Who should care: anyone doing business in Europe, ethics committees, and teams building a values-based AI policy.

Part 4: IEEE Practice Standards

The IEEE is the world's largest technical professional organisation. Its 7000 series turns ethics into engineering practice. These five standards give designers and developers concrete methods, not just principles.

16. IEEE 7000 — Addressing Ethical Concerns in System Design

IEEE 7000, published in 2021, is a standard process for building ethics into a system from the very start. Its full title is the Model Process for Addressing Ethical Concerns during System Design.

The idea is called value-based engineering. Most projects begin with a list of features. IEEE 7000 asks teams to begin with values instead. What do the people affected by this system care about? Privacy? Fairness? Autonomy? Safety?

The process works in clear steps. First, define the system's concept and context of use. Next, identify all stakeholders, including people who are affected but never use the system. Then explore the values at stake, drawing on different ethical viewpoints.

Those values are then ranked and turned into specific "ethical value requirements". These requirements are traced into the system design, just like any other technical requirement. This means ethics is not just discussed in a meeting. It is written into the specification and can be checked later.

The standard also stresses transparency. Teams record the decisions they made and the trade-offs they accepted, so others can understand and review them.

IEEE 7000 works for any technology, not just AI. But it is especially valuable for AI, where design choices can shape outcomes for millions of people.

Who should care: product managers, system architects, and design teams who want ethics built in, not bolted on.

17. IEEE 7001 — Transparency of Autonomous Systems

IEEE 7001, published in 2021, sets out how to make autonomous systems transparent in a way that can be measured and tested. It applies to systems that act with some independence, such as robots, self-driving vehicles, and AI decision tools.

The standard's key insight is that transparency means different things to different people. A user wants to know what the system is doing and why. An accident investigator needs detailed logs. A safety certifier wants evidence about design and testing.

So IEEE 7001 defines several stakeholder groups. These include users, the general public and bystanders, safety certification bodies, incident investigators, and lawyers or expert witnesses. For each group, it sets out levels of transparency, from minimal to very detailed.

This turns a vague goal into something practical. A team can decide which transparency level each group needs, design the system to meet it, and then test whether it does. Examples include explaining actions in plain words to users, or keeping a reliable record of events that investigators can use, much like an aircraft's black box.

The standard also helps organisations assess existing systems. You can check which transparency levels a system meets today and where the gaps are.

As laws increasingly require explainability, IEEE 7001 offers a clear way to show what "transparent enough" means for each audience.

Who should care: robotics and autonomous system developers, safety engineers, and teams facing explainability requirements.

18. IEEE 7002 — Data Privacy Process

IEEE 7002, published in 2022, describes how to build privacy into the engineering process for systems and software that handle personal data. Its focus is on how products are built, not just on legal compliance after the fact.

Privacy laws such as GDPR say what organisations must achieve. But they rarely say how engineers should design a system to get there. IEEE 7002 fills that gap with a practical process.

The standard guides teams to identify where personal data enters the system, how it flows, where it is stored, and who can access it. It then asks teams to assess privacy risks and set privacy requirements early in development.

A central element is the privacy impact assessment. This helps teams spot privacy harms before a product launches, when fixes are cheaper and easier. The standard also encourages ongoing checks as the system changes over time.

It promotes well-known privacy ideas, such as collecting only the data you need, limiting how long you keep it, and giving people meaningful control. It also links privacy to the wider system life cycle, so it is not forgotten after launch.

For AI, this matters a great deal. AI systems often rely on large amounts of personal data for training and operation. IEEE 7002 helps make sure that data is handled responsibly from day one.

Who should care: privacy engineers, data protection officers, and development teams working with personal data.

19. IEEE 7003 — Algorithmic Bias Considerations

IEEE 7003 sets out how to identify and reduce unwanted bias in algorithmic systems, including AI. After years of development, it was published at the end of 2024 and released in early 2025.

Bias in AI is one of the most discussed risks. An AI system might treat people unfairly because of their gender, age, ethnicity, disability or other traits. This can happen even when nobody intended it, often because of biased data or design choices.

IEEE 7003 gives a structured way to deal with this across the whole life cycle. It asks teams to define who the system will affect and what "unfair" would mean in that context. It then guides them to look for bias in data, models and outcomes.

A key idea is the bias profile. This is a living record of the bias risks you have identified, the decisions you made, and the steps you took to address them. It helps teams stay accountable and gives auditors clear evidence.

The standard also recognises that some differences in outcomes may be justified, while others are harmful. It asks teams to think carefully and document their reasoning, rather than chase a single fairness number.

Bias rules now appear in several laws, including the EU AI Act's data governance requirements. IEEE 7003 offers a recognised method to show you took bias seriously.

Who should care: data scientists, fairness and ethics teams, HR technology buyers, and anyone using AI to make decisions about people.

20. IEEE 7010 — Well-Being Impact Assessment

IEEE 7010, published in 2020, helps organisations assess how AI and autonomous systems affect human well-being. It goes beyond avoiding harm. It asks whether a system actually makes people's lives better.

Most technology is judged by narrow measures, such as clicks, revenue or speed. IEEE 7010 argues this is not enough. A product can hit its business targets while harming users' mental health, relationships or communities.

To fix this, the standard draws on established well-being research and indicators. It looks at many areas of life. Examples include life satisfaction, health, psychological well-being, community, education, work, economy, environment and governance.

The process has clear steps. First, understand the system and the people it affects. Then pick the most relevant well-being indicators. Measure them before and after the system is introduced. Finally, use the results to improve the design.

This gives teams real data on their social impact, not just good intentions. It also helps catch side effects early, such as an app that boosts engagement but increases anxiety.

IEEE 7010 links closely to wider goals like sustainable development and ESG reporting. It is a natural fit for organisations that want to show their AI benefits society, not just the bottom line.

Who should care: product and UX teams, ESG and sustainability leads, and companies building consumer-facing AI.

Part 5: Laws, Treaties & Codes

This is where AI governance gets real teeth. Some of these are binding laws with heavy fines. Others are voluntary codes that regulators and customers increasingly expect you to follow.

21. EU AI Act

The EU AI Act is the world's first comprehensive AI law. It entered into force on 1 August 2024 and applies in stages. It covers any organisation that places AI on the EU market or whose AI output is used in the EU, wherever that organisation is based.

The Act takes a risk-based approach. The higher the risk, the stricter the rules:

  • Unacceptable risk — banned outright, such as social scoring and manipulative AI. These bans have applied since February 2025.
  • High risk — AI used in areas like hiring, education, credit, healthcare and law enforcement. These systems need risk management, quality data, documentation, human oversight and conformity checks.
  • Limited risk — transparency duties, such as telling people they are talking to a chatbot or labelling deepfakes.
  • Minimal risk — most AI, like spam filters, with no new obligations.

General-purpose AI models, such as large language models, have their own rules. These have applied since August 2025.

Timelines have shifted. In 2026, the EU agreed a "Digital Omnibus" package that delays high-risk obligations. Stand-alone high-risk systems now face full rules from 2 December 2027, and AI built into regulated products from 2 August 2028. The risk categories and bans themselves are unchanged.

Penalties are serious, reaching up to EUR 35 million or 7% of global annual turnover for the worst breaches.

Who should care: anyone selling or using AI in Europe, plus their suppliers and legal teams.

22. EU GPAI Code of Practice

The EU General-Purpose AI Code of Practice is a voluntary guide that helps AI model providers comply with the EU AI Act. It was published by the European Commission in July 2025, after input from around a thousand experts, companies and civil society groups.

"General-purpose AI" means models that can do many different tasks, such as the large language models behind popular chatbots. The AI Act gives providers of these models specific duties. The Code explains, in practical terms, how to meet them.

It has three chapters:

  1. Transparency — a model documentation form that providers fill in and share with regulators and with companies that build on the model
  2. Copyright — policies to respect EU copyright law, including honouring website opt-outs from data mining
  3. Safety and security — for the most powerful models with "systemic risk", steps to assess and reduce serious risks, report incidents and protect against cyber threats

Signing the Code is voluntary. However, providers who follow it can use it to show they meet their legal obligations. This gives them more legal certainty and a smoother relationship with the EU AI Office.

Many of the largest AI developers have signed it. For businesses that buy or build on these models, the Code also matters. It shapes what information you can expect from your AI suppliers.

Who should care: AI model developers, companies building products on top of foundation models, and procurement teams.

23. Council of Europe AI Convention

The Council of Europe Framework Convention on Artificial Intelligence is the first legally binding international treaty on AI. Its full name refers to human rights, democracy and the rule of law. It was adopted in May 2024 and opened for signature in September 2024.

The Council of Europe is separate from the European Union. It is Europe's leading human rights organisation, with 46 member states. But this treaty is open to countries worldwide. Early signatories included the EU, the UK, the United States and several non-European countries.

The treaty requires countries to make sure AI activities respect human rights, democracy and the rule of law. It applies across the full AI life cycle. Its core principles include human dignity, transparency and oversight, accountability, equality and non-discrimination, privacy, reliability and safe innovation.

It also requires practical safeguards. People must have access to effective remedies when AI harms their rights. Countries must carry out risk and impact assessments. They must also protect democratic processes from AI-driven manipulation.

The treaty covers public bodies directly. For private companies, each country can choose how to apply it, either by applying it fully or through other measures. It enters into force once five countries, including three Council of Europe members, have ratified it.

For organisations, its effect will mainly come through national laws that countries pass to implement it.

Who should care: governments, public sector bodies, human rights lawyers, and companies supplying AI to the public sector.

24. G7 Hiroshima Code of Conduct

The G7 Hiroshima Process International Code of Conduct is a voluntary code for organisations developing the most advanced AI systems. It was agreed by G7 leaders in October 2023, under Japan's presidency, as part of the Hiroshima AI Process.

It was a fast political response to the rise of generative AI. The goal was to agree common expectations among the world's largest democratic economies, while full laws were still being written.

The Code sets out eleven actions. Key ones include:

  • Identify and reduce risks across the AI life cycle, including before release
  • Monitor for misuse and vulnerabilities after release
  • Publicly report on model capabilities, limitations and appropriate use
  • Share information on incidents responsibly with others
  • Put in place strong AI governance and risk policies
  • Invest in security, including protecting model weights
  • Develop ways to identify AI-generated content, such as watermarking
  • Prioritise research on safety and on AI for global challenges

In 2025, the OECD launched a reporting framework so companies can publicly show how they follow the Code. Several major AI developers have submitted reports.

The Code is not law. But it carries weight. It reflects what governments expect from frontier AI developers, and its ideas appear in later rules, including the EU's GPAI Code of Practice.

Who should care: frontier AI developers, policy teams, and buyers wanting a benchmark for responsible AI vendors.

25. South Korea AI Basic Act

South Korea's AI Basic Act is one of the world's first comprehensive national AI laws. It was passed in December 2024 and took effect on 22 January 2026, together with its enforcement decree. Its full name refers to developing AI and building a foundation of trust.

The law has two goals: to grow the AI industry and to make AI safe and trustworthy. So it mixes support measures, like funding, talent and infrastructure, with real obligations.

It focuses on three areas:

  • High-impact AI — AI in critical sectors such as healthcare, energy, transportation, hiring and biometric analysis. Providers must manage risks, explain results where feasible, ensure human oversight and keep records.
  • Generative AI — users must be told when a product uses generative AI, and AI-generated content such as deepfakes must be clearly labelled.
  • High-performance AI — the most powerful models, trained with very large amounts of computing power, face extra safety duties.

The law also reaches foreign companies. Large overseas AI providers serving Korean users must appoint a local representative.

Unlike the EU, Korea brought all its rules into force at once. However, the government promised flexible enforcement and a grace period before fines are imposed. Fines are also much lower than under the EU AI Act.

Who should care: any company offering AI products or services in South Korea, and teams tracking AI laws in Asia.

Part 6: Security, Data & Assurance

Good governance means little if your AI can be hacked, runs on poor data, or cannot be independently checked. These five frameworks cover certification, data quality and AI security.

26. ISO/IEC 42006 — Requirements for Certification Bodies

ISO/IEC 42006, published in July 2025, sets the rules for the organisations that audit and certify AI management systems against ISO/IEC 42001. In short, it audits the auditors.

Why does this matter? A certificate is only as trustworthy as the body that issues it. Before this standard, certification bodies relied on general auditing rules. Those rules were not written with AI in mind.

ISO/IEC 42006 builds on ISO/IEC 17021-1, the general standard for management system certification bodies. It then adds AI-specific requirements. These cover how audits are planned, how long they should take, and what evidence auditors must examine.

A major focus is auditor competence. Auditors must understand AI concepts, AI risks, impact assessment, data management and the AI life cycle. They must be able to judge whether an organisation's AI controls actually work, not just whether documents exist.

The standard also helps make certification consistent. A certificate from one accredited body should mean roughly the same thing as one from another.

You will probably never use ISO/IEC 42006 directly unless you work for a certification body. But it matters when choosing one. If you are pursuing ISO/IEC 42001 certification, ask whether your auditor is accredited and follows ISO/IEC 42006. It helps you trust that your certificate will be respected by customers and regulators.

Who should care: certification bodies, accreditation bodies, and organisations choosing an ISO/IEC 42001 auditor.

27. ISO/IEC 5259 — Data Quality for Analytics and Machine Learning

ISO/IEC 5259 is a multi-part standard on data quality for analytics and machine learning. Its parts were published between 2024 and 2025. It answers a simple question: how do you know your data is good enough for AI?

The saying "garbage in, garbage out" is especially true for AI. A model trained on incomplete, outdated or biased data will make poor or unfair decisions. Yet many organisations have no formal way to measure data quality.

The series has five parts:

  1. Overview, terminology and examples — shared language for data quality
  2. Data quality measures — characteristics such as accuracy, completeness, consistency, timeliness and representativeness, with ways to measure them
  3. Data quality management requirements — what an organisation needs in place to manage data quality
  4. Data quality process framework — how to plan, assess and improve data quality across the data life cycle
  5. Data quality governance framework — how leadership oversees and directs data quality

The standard covers the full data life cycle, from collection and labelling to use and eventual disposal. It also deals with issues specific to ML, such as whether training data truly represents the people the system will serve.

Data quality underpins almost every other framework in this list. The EU AI Act requires high-risk systems to use relevant, representative and error-checked data. ISO/IEC 5259 gives you a structured way to show that.

Who should care: data engineers, data governance teams, data scientists and chief data officers.

28. NIST AI 100-2 — Adversarial Machine Learning

NIST AI 100-2 is a guide to attacks on AI systems and how to defend against them. Its full title is Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations. NIST updates it regularly, with a major edition released in 2025.

"Adversarial machine learning" means deliberately tricking or attacking AI. These attacks are different from normal cyberattacks, because they target how AI learns and makes decisions.

The report sorts attacks into clear categories. For traditional predictive AI, the main ones are:

  • Evasion — tweaking inputs to fool a model, such as stickers that make a stop sign look like a speed limit sign
  • Poisoning — corrupting training data so the model learns the wrong behaviour
  • Privacy attacks — extracting sensitive training data or stealing the model itself

For generative AI, it adds attacks such as prompt injection, where hidden instructions hijack a chatbot or AI agent, and misuse, where safety guardrails are bypassed.

For each type, the report explains how attacks work, what attackers need to know, and which defences exist. It is honest about limits too. Many defences are still imperfect, and there are trade-offs between security, accuracy and cost.

Its biggest value is a shared vocabulary. Security teams, AI developers and auditors can describe threats in the same way. It is free and is often used alongside MITRE ATLAS and the OWASP Top 10.

Who should care: security teams, AI engineers, red teamers and risk managers.

29. OWASP Top 10 for LLM Applications

The OWASP Top 10 for Large Language Model Applications lists the most critical security risks in apps built on LLMs, such as chatbots, copilots and AI agents. It is produced by OWASP, a respected non-profit known for its famous web security Top 10. The current version was released for 2025.

It was created because LLM apps bring new security problems that traditional checklists miss. The 2025 list is:

  1. Prompt injection — malicious inputs that override the model's instructions
  2. Sensitive information disclosure — the model revealing private or confidential data
  3. Supply chain — risks from third-party models, datasets and plugins
  4. Data and model poisoning — tampered training or fine-tuning data
  5. Improper output handling — trusting model output without checking it
  6. Excessive agency — giving AI too much power to act on its own
  7. System prompt leakage — exposing hidden instructions or secrets
  8. Vector and embedding weaknesses — flaws in retrieval systems (RAG)
  9. Misinformation — false but convincing outputs
  10. Unbounded consumption — runaway usage that drives up costs or causes outages

Each risk comes with examples, attack scenarios and practical prevention steps. That makes it very useful for developers and testers.

The list is free, community-driven and updated as threats change. It now sits within the wider OWASP GenAI Security Project, which also offers guides on AI agents and red-teaming.

Who should care: developers building LLM apps, application security teams, and anyone buying an AI chatbot or agent.

30. MITRE ATLAS — Adversary Tactics Against AI

MITRE ATLAS is a free knowledge base of real-world tactics and techniques that attackers use against AI systems. ATLAS stands for Adversarial Threat Landscape for Artificial-Intelligence Systems. It is run by MITRE, a US non-profit.

It is modelled on MITRE ATT&CK, the framework security teams around the world use to describe cyberattacks. If your security team knows ATT&CK, they will find ATLAS familiar.

ATLAS is organised as a matrix. The columns are tactics, which describe the attacker's goal at each stage. Examples include reconnaissance, gaining initial access, getting access to the AI model, persistence, evading defences, stealing data, and causing impact. Under each tactic are specific techniques, such as poisoning training data, crafting adversarial inputs, or injecting prompts into an LLM.

What makes ATLAS especially valuable is its real-world grounding. It includes case studies of actual attacks and red-team exercises against AI systems. This shows that these threats are not just theory.

It also lists mitigations that link back to techniques. Teams can use it for threat modelling, planning red-team exercises, and checking whether their defences cover known attacks.

Where NIST AI 100-2 explains the science of AI attacks and OWASP focuses on app development, ATLAS shows how real attackers chain techniques together. Used together, the three give a well-rounded view of AI security.

Who should care: security operations teams, threat intelligence analysts, red teamers and CISOs.

Where Should You Start?

Thirty frameworks can feel like a lot. The good news is that they overlap heavily, and you can build up in layers:

  1. Know your laws. Check whether the EU AI Act, South Korea's AI Basic Act or other national rules apply to you.
  2. Pick a management backbone. ISO/IEC 42001 if you want certification, or the NIST AI RMF for a flexible start. Small firms can begin with the UK's AIME checklist.
  3. Add depth where your risks are. Use NIST AI 600-1 for generative AI, ISO/IEC 42005 for impact assessments, and IEEE 7003 for bias.
  4. Secure it. Use the OWASP Top 10, NIST AI 100-2 and MITRE ATLAS to protect your AI systems.
  5. Anchor it in values. Use the OECD, UNESCO and EU principles to shape your AI policy.

AI governance is not a one-time project. Laws, standards and threats keep changing. Pick a starting point, build steadily, and review your approach at least once a year.

Sources