logo
language iconarrow
menu icon

ISO/IEC 42001:2023

42001-06 AI system impact assessment

4 min read · V, IA, EA

All articles

For: V, IA, EA
Source: ISO/IEC 42001:2023(en), English. 3.24, 6.1.4, 8.4, A.5, B.5

The impact assessment is not the risk assessment. 6.1.2 assesses AI risks, including potential consequences for the organization, persons and societies. 6.1.4 is a process of its own. Its results shall be taken into account in the risk assessment. The note to 6.1.2 allows the impact assessment to be used when assessing the consequences. “Can”. It does not replace 6.1.4.

6.1.4 requires a process to be defined in order to assess the potential consequences for persons or groups of persons, or both, and for societies that can arise from the development, provision or use of AI systems. The impact assessment shall determine the potential consequences that the deployment, the intended use and the foreseeable misuse of an AI system can have for these persons and for societies. It shall take account of the specific technical and societal context in which the system is deployed, and of the applicable jurisdictions. The result shall be documented. As appropriate, the result can be made available to relevant interested parties, as the organization determines them. “Can”.

The note says that in some contexts, for example with safety-critical or privacy-critical systems, the organization can require that discipline-specific impact assessments, for example on safety, privacy or security, be carried out as part of the overall risk activities. “Can”.

8.4 requires the impact assessments under 6.1.4 to be performed at planned intervals or when significant changes are proposed to occur. The English wording is “proposed to occur”. It is not the same as in 8.2, where it is “proposed or occur”.

A.5 is the reference group in Table A.1. The objective: assess impacts of AI systems on persons or groups, or both, and on societies throughout the life cycle. Whether a row is necessary follows from 6.1.3. Where it is necessary, the control reads:

A.5.2: establish a process in order to assess the potential consequences for persons or groups, or both, and for societies that can arise from the AI system throughout its life cycle.

A.5.3: document the results and retain them for a defined period. The standard names no number of years.

A.5.4: assess and document the potential impacts on persons or groups of persons throughout the life cycle.

A.5.5: assess and document the potential societal impacts throughout the life cycle.

Annex B repeats these controls predominantly with “should” and gives implementation guidance. B.5.2 says the organization should assess, on the basis of purpose and use, whether a system affects the legal position or the life chances of persons, their physical or psychological well-being, universal human rights or societies. Procedures should include, but not be limited to, the circumstances, the elements, the body carrying them out, the use of the assessment and the persons and societies that may be affected. The elements can comprise identification, analysis, evaluation, treatment, and documentation, reporting and communication. “Can include”. That is not a second shall-list beside 6.1.4. The same guidance says the processes could vary according to the role, the domain and the discipline, for example security, privacy and safety. The organization should determine whether discipline-specific assessments in risk management take up AI considerations sufficiently. According to the note, ISO/IEC 23894 describes how an organization can carry out impact analyses as part of risk management. The standard is not available.

B.5.3: documentation can help to determine what should be communicated to users and other parties. Assessments should be retained and updated when needed. Retention periods can follow the organization’s periods or be informed by legal or other requirements. What should be documented is named by the guidance as points to be considered, including the intended use and reasonably foreseeable misuse, positive and negative impacts, foreseeable failures, relevant demographic groups, complexity, and the role of humans including oversight, employment and qualification. “Should consider”, not the wording of A.5.3.

B.5.4: when assessing, the organization should take account of its governance principles, the AI policy and the objectives. The specific protection needs of groups such as children, persons with impairments, older persons and workers should be taken into account. Areas that can be considered depending on purpose and use include, among others, fairness, accountability, transparency and explainability, security and privacy, safety and health, financial consequences, accessibility and human rights. “Can include”. Where necessary, the organization should consult specialists.

B.5.5: societal impacts can be beneficial or detrimental. Examples in the guidance are the environment, the economy, the state, health and safety, norms and culture. The organization should consider the impacts in connection with its environmental objectives, possible misuse and historical harm, and positive as well as negative outcomes. ISO/IEC TR 24368 is named as an overview of ethical and societal concerns and is not retold. The examples are guidance, not a mandatory matrix.

Documented information: The result of the impact assessment shall be documented (6.1.4). Retained documented information about the results of all impact assessments (8.4). Where A.5.3 is necessary, the results shall be documented and retained for a defined period.

Not in this text: a method for the impact assessment, an obligation to work through every point from B.5.4 and B.5.5, and a human-rights or environmental expert opinion. 6.1.4 and, to the extent necessary, A.5 are the requirements. The risk assessment remains 6.1.2.