The lifecycle and where privacy risk enters
The abstract asks for privacy risks to be identified throughout the AI system lifecycle, not once at launch. The lifecycle is where data first appears, where it is copied into training sets, where a model is reused for a second purpose and where an output reaches a person.
Read plainly, the public abstract gives the draft three moves: find the privacy risk wherever it appears in the life of the AI system, weigh what it could do to people, then decide how to deal with it. That is the whole shape of the draft in one sentence: find, weigh, treat, across the lifecycle.
The lifecycle stages are the ones ISO/IEC 22989:2022 names: inception, design and development, deployment, operation and monitoring, and retirement. Each stage changes the privacy picture. A model can be privacy-safe at training and leak in operation through prompts, logs or a shared tenancy. A supplier can change a model under a version you did not assess.
The practical work is a register that follows the system, not the project. For each AI system, write down which personal data enters, at which stage, for which purpose and under whose decision. That register is the spine of the free check, the kit and the watch.
What an auditor or a privacy reviewer is likely to ask for
- Scope The AI systems in use and the personal data each one touches, with the role the organization plays in that data.
- Assessment The impact assessment for each system, the groups considered, the harms named and the decision taken.
- Control The data and provenance record, the supplier terms, the retention setting and the test that the control works.
- Review The monitoring thresholds, the log retention, the incident route and the date of the last review.
These are the records this site expects a reader to hold for the area. They are our reading of the public record and the current controls, not 27091 draft text.
Where it commonly fails today
- The AI systems are not listed, so the assessment covers the model and misses the personal data around it.
- The assessment stops at accuracy and never reaches the person the output decides about.
- The data provenance is a folder of files with no record of where each dataset came from or on what basis.
- A supplier changes the model or the retention period and the assessment is not repeated.
- The logs hold prompts and outputs far longer than anyone decided they should.
What to do now, while it is still a draft
Do not rebuild the system around a draft. Do the work that will be right under either instrument: list the systems and the data, settle the role, run the impact assessment, record the provenance, fix the supplier terms, set the retention and test the incident route. If the published guidance differs, the process survives the wording change.
Source for this public record page: the ISO page 56582 and its stage feed, with the current control set where a requirement is stated. The draft text is not public and is not reproduced. Every source.