ISO/IEC 23894:2023 Guidance on AI risk management

What it is

ISO/IEC 23894:2023, Information technology, Artificial intelligence, Guidance on risk management, was published in February 2023 by ISO and IEC through JTC 1/SC 42. It gives organisations guidance on managing risks that are specific to AI and on building AI risk management into their wider activities and functions.

It is written as an extension of ISO 31000:2018, the general risk management standard. Rather than inventing a new method, it walks through ISO 31000’s principles, framework and process and explains what changes when the subject is an AI system. The standard is paid.

Who it applies to

Any organisation that develops, produces, deploys or uses products, systems and services that use AI can apply it. It is voluntary and can be tailored to any organisation and its context. Its most common use is inside an ISO/IEC 42001 management system, where clause 6 requires an AI risk assessment and risk treatment, and as an AI extension to an existing enterprise risk management programme.

What it requires

As guidance, it recommends rather than mandates. Its content follows the three parts of ISO 31000.

Principles

The ISO 31000 principles apply, such as risk management being integrated, structured, customised, inclusive, dynamic and based on the best available information. The standard adds AI-specific notes to each one, for example that the best available information about an AI system can change quickly once it is in use, and that people affected by a system may need to be involved beyond the usual stakeholders.

Framework

The framework covers leadership and commitment, integration into organisational governance, design of the framework, implementation, evaluation and improvement. The guidance points to the particular need for AI expertise, clear accountability for AI systems, and coordination between technical, legal and business functions.

Process

The risk management process runs through communication and consultation, establishing scope, context and criteria, risk assessment (identification, analysis and evaluation), risk treatment, monitoring and review, and recording and reporting. For AI, the guidance stresses identifying risk sources across the whole lifecycle, monitoring systems after deployment because their behaviour and context can change, and recording decisions so they can be traced.

Annexes

The annexes are where most organisations spend their time.

• Annex A sets out common AI-related objectives that risk management should protect, including accountability, AI expertise, availability and quality of training and test data, environmental impact, fairness, maintainability, privacy, robustness, safety, security, and transparency and explainability.
• Annex B describes AI-specific risk sources, such as the complexity of the environment, lack of transparency and explainability, level of automation, risks related to machine learning, system hardware issues, system lifecycle issues and technology readiness.
• Annex C maps risk management activities onto the AI system lifecycle, from inception through design and development, verification and validation, deployment, operation and monitoring, to retirement.

Using it in practice

Most organisations do not run ISO/IEC 23894 as a separate process. They extend the risk process they already have. In practice that means adding the Annex A objectives to the organisation’s risk criteria, so an AI risk is rated against fairness, transparency and data quality as well as financial and operational effects. It means using the Annex B risk sources as prompts when identifying risks for each AI system, and using Annex C to decide which risk activities happen at which lifecycle stage, so a system is assessed before it is built or bought, again before deployment, and again when it changes. The results belong in the same risk register as other enterprise risks, with owners, treatments and review dates, so AI risk is reported to the same governance forums as everything else.

Key dates

ISO/IEC 23894 was published in February 2023 as its first edition and is current.

How it relates to other frameworks

ISO/IEC 23894 is the detailed risk companion to ISO/IEC 42001. Where 42001 says an organisation must assess and treat AI risk, 23894 explains how to do it in a way consistent with ISO 31000. ISO/IEC 42005 complements it by focusing on impacts to people and society rather than risk to the organisation.

Its content overlaps with the NIST AI Risk Management Framework, which covers similar ground in its Map, Measure and Manage functions with a more detailed catalogue of outcomes. Organisations often use 23894 to align AI risk with their enterprise risk process and the NIST AI RMF for the working detail. Both support the risk management system the EU AI Act requires of providers of high-risk AI systems, though neither creates a presumption of conformity with the Act.