The three established models
Human in the loop
The system waits for a person's approval. Nothing is issued without it. Recent academic work calls this position constitutive: the human contribution is necessary to the output itself.
Human on the loop
The system acts, a person watches and may intervene. The same work calls this position corrective: it sits outside the primary causal chain, and it operates synchronously, asynchronously or in anticipation.
Human in command
A person holds ultimate authority over the system, including the authority not to use it. The AI Act frames oversight in these terms in Article 14, alongside awareness of automation bias and the power to interrupt.
What the three have in common
They are local. Each describes one control point at a time, inside one decision. That is not a weakness — it is what they were built for, and they do it well.
The question none of them answers
A director does not ask who approves this output. They ask where the humans are across everything we now run, and what the whole set costs. That question is about the map, not the loop.
You can hold a compliant control point on every workflow and still not know what your use of AI costs you. The loop is answerable locally. The map is not answerable at all without someone whose position is outside the chains.
The field states the confusion itself
This is not our claim. The literature on human oversight records that interdisciplinary discourse is affected by semantic misalignment, and that “human in the loop” covers situations as different as a person labelling training data and a manager who may override a recommendation — configurations whose causal influence differs and which call for terminological differentiation.
Work on testing compliance with oversight requirements states the same tension from the other side: the choice lies between simple checklist approaches that may be ineffective, and empirical testing that is resource-intensive and context-dependent. The need is named. It is not filled.
What Human in the Map is, and what it does not claim
It does not claim a taxonomy. That work exists, it is recent, and it is done by researchers better equipped for it. It does not replace or improve the three established models.
It names the operation that applies those categories to one real organisation, on an agreed scope, at a given date. Not a discovery — the identification of a subject. A discovery is contested; a grouping is verified.
What came before
An earlier framing, the preconformity layer, described the same concern from the regulatory side. It is not deleted here. It gave the published research its frame, and removing it would leave that work without its context. It was set aside because it answered where an organisation stands before the regulation, where the question that matters is about who does what once the machine has taken part of the work.
Sources
- Baum & Laux — Constitutive vs. Corrective: A Causal Taxonomy of Human Runtime Involvement in AI Systems (2026)
- Langer, Lazar & Baum — On the Complexities of Testing for Compliance with Human Oversight Requirements in AI Regulation (2025)
- Article 14 — Human oversight, EU AI Act
- Designing meaningful human oversight in AI — AI and Ethics (2026)
Each source is cited for what it states, not for what it would support.