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. 6.1.1, 6.1.2, 6.1.3, 8.2, 8.3, term 3.26. Annex A.1 as the statement that not every row of Table A.1 shall be used.
6.1.1 requires that, when planning, the issues from 4.1 and the requirements from 4.2 be considered and the risks and opportunities that shall be addressed be determined, in order to ensure that the system can achieve its intended results, to prevent or reduce undesired effects, and to achieve continual improvement.
The organization shall establish and maintain AI risk criteria that help to distinguish acceptable risks from unacceptable risks, to perform AI risk assessments, to perform AI risk treatment and to assess the impacts of AI risks.
It shall determine risks and opportunities according to the domain and application context of an AI system, according to the intended use and according to the external and internal context from 4.1. Note 2: more than one AI system can lie within the scope. The determination of the opportunities and uses is then made for each system or for groups of systems.
It shall plan actions to address these risks and opportunities, and plan how the actions are integrated into the processes of the system and implemented and how their effectiveness is evaluated.
Note 1 refers to ISO/IEC 38507 and ISO/IEC 23894 for considerations of what degree and what kind of risk an organization wants to pursue or retain. Note 3 refers to ISO/IEC 23894 for the implementation of risk management. Neither is available. Note 4: context and activities can influence the risk activities. Note 5: the way of defining risk can differ between sectors. 3.7 allows a broad view. Within the assessment the organization first adopts a view adapted to its context. That can include sector definitions, for example ISO/IEC Guide 51. “Can”.
6.1.2 requires an AI risk assessment process to be defined and established that:
a) is informed by the AI policy and the AI objectives and is consistent with them. The note says that, when assessing the consequences under d) 1), the organization can use an impact assessment under 6.1.4. “Can”. The obligation to take the results of the impact assessment into account in the risk assessment is in 6.1.4, see 42001-06;
b) is designed so that repeated assessments can produce consistent, valid and comparable results;
c) identifies risks that promote or prevent the achievement of the AI objectives;
d) analyses the AI risks in order to assess the potential consequences for the organization, for persons and for societies that would arise if the risks were to materialize; as applicable, to assess the realistic likelihood of occurrence; to determine the levels of risk;
e) evaluates the AI risks in order to compare the results of the analysis with the criteria from 6.1.1 and to prioritize the evaluated risks for treatment.
6.1.3 requires a treatment process to be defined, taking the results of the assessment into account, in order to:
a) select appropriate treatment options;
b) determine all controls necessary to implement the chosen options, and compare the controls with those in Annex A in order to verify that no necessary control has been omitted. Note 1: Annex A provides reference controls for organizational objectives and for risks related to the design and use of AI systems;
c) consider the controls from Annex A that are relevant to the implementation of the treatment options;
d) identify whether further controls beyond Annex A are necessary in order to implement all the treatment options;
e) consider the guidance in Annex B for the implementation of the controls determined under b) and c). Note 2: control objectives are implicitly contained in the controls chosen. The organization can select a suitable set from Annex A. The controls in Annex A are not exhaustive. Further objectives and controls can be necessary. Where other or additional controls are necessary, the organization can design them itself or take them from existing sources. AI risk management can be integrated into other management systems, if applicable;
f) produce a statement of applicability that contains the necessary controls from b), c) and d) and justifies inclusion and exclusion. A justification for the exclusion can include that the risk assessment does not regard the control as necessary, and that external applicable requirements do not require it or provide for exceptions. Note 3: the organization can provide documented justifications in order to exclude control objectives in general or for particular AI systems, whether the objectives are in Annex A or were established by the organization itself;
g) formulate an AI risk treatment plan.
The organization shall obtain the approval of the designated management for the treatment plan and for the acceptance of the residual AI risks. The sentence does not further specify who this management is. The necessary controls shall be aligned with the objectives from 6.2, be available as documented information, be communicated within the organization and be available to interested parties as appropriate.
A.1 repeats: not all objectives and controls in Table A.1 shall be used. The statement of applicability is the place for the justification. Note 2 to 3.26 requires, with “shall”, that all risks identified are documented and that all risks identified and the measures determined for their treatment are reflected in the statement.
8.2 is the carrying out, not the planning once again. The organization shall perform AI risk assessments in accordance with 6.1.2 at planned intervals or when significant changes are proposed or occur. The English wording is “proposed or occur”.
8.3 requires the treatment plan under 6.1.3 to be implemented and its effectiveness to be verified. If an assessment identifies new risks that require treatment, a treatment process under 6.1.3 shall be carried out for these risks. If treatment options of the plan are not effective, they shall be reviewed and revalidated in accordance with the process in 6.1.3, and the plan shall be updated.
Documented information: Retained documented information about the actions taken to identify and address AI risks and AI opportunities (6.1.1). Retained documented information about the assessment process (6.1.2). Retained documented information about the treatment process (6.1.3). The necessary controls shall be available as documented information (6.1.3). Retained documented information about the results of all AI risk assessments (8.2) and of all AI risk treatments (8.3). The statement of applicability is the documentation described in 6.1.3 f and 3.26.
Not in this text: a risk matrix, the content of ISO/IEC 23894, and the claim that every row of Table A.1 is, without further ado, a nonconformity. A.1 and 6.1.3 say which controls are necessary and that the exclusion is justified. The impact assessment is a process of its own, 42001-06.