Defining risk criteria

The organization states how much risk, and of what kinds, it is prepared or not prepared to take, and fixes criteria for judging how significant a risk is (ISO 31000:2018, 6.3.4), considering: the kinds of uncertainty, tangible or intangible, that can affect outcomes and objectives; how consequences and likelihood are defined and measured; timing factors; whether measurement is consistent; how a level of risk is arrived at; how several risks together or in sequence are handled; and the organization's capacity. The step before 6.3.4 supplies its input, so run both together.

What it asks

The organization states how much risk, and of what kinds, it is prepared or not prepared to take, and fixes criteria for judging how significant a risk is (ISO 31000:2018, 6.3.4), considering: the kinds of uncertainty, tangible or intangible, that can affect outcomes and objectives; how consequences and likelihood are defined and measured; timing factors; whether measurement is consistent; how a level of risk is arrived at; how several risks together or in sequence are handled; and the organization's capacity. For AI the criteria also reflect the Annex A objectives and effects on individuals and society, not only on the organization.

Where generative AI changes the answer

Risk criteria written for classic IT risk do not name hallucination, prompt injection, data leakage through prompts, or copyright in generated output.

Who it binds

The AI risk owner, the assessment team and the business owner of each generative AI use. For 6.3.4 the named role is the owner of documented AI risk criteria covering consequence and likelihood scales and the Annex A objectives.

Evidence an auditor normally samples

• Documented AI risk criteria covering consequence and likelihood scales and the Annex A objectives

The test you can run

Compare documented AI risk criteria covering consequence and likelihood scales and the Annex A objectives with the use as it runs today; the gap between the two is the finding to record.

Common failures

• Criteria that measure financial impact only, with no criterion for fairness, safety or privacy

Before the draft publishes

For 6.3.4, run the step once, by hand, on one live generative AI use, and keep the working paper. A method that has never been run on a real use is a template. The first artefact to produce is documented AI risk criteria covering consequence and likelihood scales and the Annex A objectives.

Where this sits against the draft

6.3.4 carries the risk management process part of the draft's published scope, which names risk sources, the stakeholders exposed to them, risk analysis, risk treatment and controls. When the draft publishes, compare its wording on 6.3.4 (defining risk criteria) with this record and note the change in the register.