Explanations of ISO/IEC 42001:2023
Source is ISO/IEC 42001:2023(en). Title: Information technology — Artificial intelligence — Management system. It was prepared by the joint…
Read articleISO/IEC 42001:2023
For: V, IA
Source: ISO/IEC 42001:2023(en), English. A.6, B.6. A generic life cycle model is named by B.6.1.3 in ISO/IEC 22989. That edition is not available. The stages in this article are the controls in A.6, not a model taken over from 22989.
The objective A.6.1 is to ensure that the organization determines and documents objectives and implements processes for responsible design and responsible development. The objective A.6.2 is to define criteria and requirements for each stage of the life cycle. Each of the following controls applies in its wording, to the extent it is necessary under 6.1.3. Annex B repeats them with “should”, except where the box says otherwise, and provides examples. The examples are not copied out as a checklist.
A.6.1.2: determine and document objectives that guide the responsible development of AI systems, take those objectives into account and integrate measures to achieve them into the development life cycle. The English sentence reads “take those objectives into account and integrate measures to achieve them in the development life cycle”. B.6.1.2 says that such objectives should influence design and development. The example of fairness, which then appears in the requirement, the data, the training and the testing, is an example from the guidance. Annex C can be useful when determining the objectives. It is informative.
A.6.1.3: define and document the specific processes for responsible design and responsible development. B.6.1.3 says that responsible development should consider, among other things: life cycle stages, where the organization can define its own stages; testing requirements; requirements for human oversight, especially where the system can affect natural persons; in which stages impact assessments should be carried out; expectations of the training data; the expertise needed; release criteria; approvals; change control; usability and controllability; involvement of interested parties. “Should include consideration”. Human oversight is guidance here, not a row of its own in Table A.1.
A.6.2.2: specify and document requirements for new AI systems or for substantial improvements of existing systems. B.6.2.2: the organization should document the justification and the objectives and specify requirements across the whole life cycle. They should be considered again when the system cannot work as intended or when new information suggests a change. ISO/IEC 5338 and ISO 9241-210 are named and are not retold.
A.6.2.3: document the design and development on the basis of the organizational objectives, the documented requirements and the specification criteria. B.6.2.3 names design decisions that can be necessary, including the learning approach, the algorithm, training and data quality, evaluation, hardware and software, security threats including Data Poisoning, Model Stealing or Model Inversion, interfaces, human interaction, interoperability. Documentation of the stage should be retained, and a final architecture documentation should be available. “Should”. The attack examples are guidance.
A.6.2.4: define and document verification and validation measures and specify criteria for their use. B.6.2.4 names possible measures and evaluation criteria, including error margins, operational factors, interpretability of the outputs for persons who decide on the basis of the outputs or who are subject to decisions, and the handling of a failure to meet a performance objective. The system should be evaluated against the documented criteria. Where it does not meet them, especially with respect to the objectives of responsible development and use, the organization should reconsider the intended use, the performance requirements and the handling of impacts, or handle the shortcomings. ISO/IEC TR 24029-1 is named for the robustness of neural networks and is not retold.
A.6.2.5: document a deployment plan and ensure that appropriate requirements are fulfilled before deployment. B.6.2.5: development and deployment can lie in different environments. Components can be deployed separately. The organization should have requirements before release and deployment, occasionally called release criteria. The plan should include the perspectives and impacts of relevant interested parties.
A.6.2.6: define and document the elements necessary for ongoing operation. The second sentence of the control says that this should include at least system and performance monitoring, repairs, updates and support. “Should” stands here in the text of the control itself, after the “shall” of the first sentence. The requirement is to define and document the necessary elements. The minimum is formulated in the same sentence as “should” and is not reinterpreted here as a second “shall”. B.6.2.6 explains monitoring, continual learning, drift, repair, updating, support, use outside the design, and AI-specific attacks on information security. Performance criteria should fit the task. As an example, an organization can consider the F1 score according to ISO/IEC TS 4213 to be appropriate. That is an example from the guidance, not the specification of a metric. ISO/IEC 25059 can be used in order to define performance criteria. “Can”.
A.6.2.7: determine what technical documentation is necessary for each relevant category of interested parties, for example users, partners and supervisory authorities, and provide it to them in an appropriate form. B.6.2.7 lists possible elements, including the purpose, instructions for use, technical assumptions, limits, possibilities for monitoring, and elements across the life cycle stages according to ISO/IEC 22989. The organization should document technical information for responsible operation, including a plan for dealing with failures. Documentation should be up to date, accurate and approved by appropriate management. Where it is part of the user documentation, the controls from Table A.1 should be included. “Should”.
A.6.2.8: determine in which stages of the life cycle event logs should be kept, but at least while the AI system is in use. The “shall” is directed at determining the stages. The minimum “when the AI system is in use” stands in the same sentence. B.6.2.8 says that the organization should ensure logging for systems that it deploys. Logs should be retained for as long as the intended use and the retention rules of the organization require. Legal requirements for retention can apply. Some systems, for example biometric identification, can have additional logging obligations depending on the jurisdiction. Organizations should know these requirements. That is guidance and names no particular law.
Not in this text: a life cycle model from ISO/IEC 22989, a checklist of every dash item in B.6, an F1 target value, and a requirement of its own, human oversight, that does not stand as a row in Table A.1. It appears in B.6.1.3 and in 42001-17 as guidance.