An AI system’s explanation describes its output; by itself, it does not show what a person did with that output. To document human oversight, a record needs to identify the person’s action and what it concerned. Andreas Ehstand’s individual Internet-Draft proposes four terms for distinguishing those actions: observation, check, decision, and release.
What an AI explanation says about human review
A description of how an AI produced an answer and a record of someone examining that answer are different things. A system explanation can document information about the output. It does not, on its own, identify who reviewed the output or whether they checked anything.
Ehstand’s Internet-Draft on human oversight acts uses four categories to make the human action more precise. Each supports a different kind of claim:
| Oversight act | What the act establishes |
| Observation | The person was presented with the system’s output. It does not establish what they noticed. |
| Check | The person examined whether a named property held. The result concerns that property, not every other aspect of the output. |
| Decision | The person selected a course of action, such as using, overriding, interrupting, or referring the output. It does not establish that a check took place. |
| Release | The person permitted a specified action under authority they already held. Permission alone does not establish that the action or its inputs were correct. |
The labels matter because a generic “reviewed” entry can blur all four. Someone may have seen a result, checked one detail, chosen what to do, or authorized an action. Those are not interchangeable claims.
What a proposed oversight record identifies
In the draft’s terminology, a record of an oversight act identifies the overseer, the type of act, the object, and the time. If the act is a check, the record also identifies the property checked and the outcome. That extra detail limits what the record can support: a positive result for one named property does not verify unrelated properties.
For a release, recording the authority the person already held and the action they permitted helps show whether the release fell within that authority. The draft describes these as semantic details, not a required software format or log template. A record generated by a system but not attributable to the overseer also does not meet the draft’s definition of an oversight record.
What testing a review step can show
A “checked” entry records activity; it does not show how well the check detects errors. Ehstand’s draft defines error detectability in relation to a stated reference population and the people and conditions involved. A test result therefore needs that context to say what kinds of errors the step detects.
The share of cases a process flags is a different measure. A process that flags every case may also flag correct cases, so that figure alone cannot show how well it distinguishes errors. The draft proposes no test procedure, measurement method, or threshold.
In 2022, Ben Green’s paper reported a survey of 41 policies requiring human oversight of government algorithms and argued that oversight effectiveness needs empirical support. The paper’s abstract does not give a universal failure rate for human review; the survey is about policies, not a measured performance rate for review processes.
A terminology proposal, not a required standard
Ehstand’s document is an individual Internet-Draft dated September 26, 2026, with intended status “Informational.” It defines terminology; it does not specify a protocol, record format, procedure, measurement method, or legal interpretation. Its four categories give organizations a vocabulary for describing what a person did, while leaving the format and evaluation of any real oversight process to other work.
