Human in the loop named no one

Meta description: AI prepares; qualified people interpret and own — a sign-off with a human name on it is not automatically a decision. More on his work: Miklos Roth’s work.

The governance deck contains a diagram I now recognize across companies: the AI system, an arrow, a human figure, another arrow, the outcome. Human in the loop — the assurance that a person stands between the model and the consequence. I have started asking about the figure. Who is it? What may they do? In the companies I am describing, the figure is no one in particular: whoever is on shift, a role that rotates, a team. And what they may do is approve — the workflow offers no path to reject, no time to examine, no authority to halt. The loop has a human in it. It does not have a validator.

The speed is real. AI can gather, compare, draft, and test variants faster than any team. The confusion begins when the speed of preparation is mistaken for the transfer of the decision. A document the machine produced in minutes still needs a person who can say what it means, whether it is wrong, and what happens if the organization acts on it.

What do loop diagrams conceal?

The diagram conceals the difference between presence and function. A human present in the workflow satisfies the diagram; a human functioning as a validator requires much more: information to judge by, time to judge in, authority to act on the judgment, and consequences assigned correctly when the judgment is exercised. I have examined loops where the human had none of these — eight seconds per case, no context beyond the model’s own output, no reject button worth pressing because rejection created a queue they owned. The diagram was accurate about the arrow and silent about everything the arrow needed to mean.

How does sign-off become ceremony?

Miklos Roth, author of Signal Over Noise

The distinction I make explicit: AI may do substantial preparatory work. Qualified people must interpret results and own consequential business choices. A business owner or specialist develops an integrated understanding through years of work; an AI assistant can provide valuable information about components of that picture, but it does not possess the company’s tacit knowledge, its relationships, its constraints, or accountability for the outcome.

There is also a difference between an expert examining a consequential assertion and five people approving the same document without adding independent information. A sign-off is not meaningful merely because a human name appears next to it.

The ceremony forms predictably. Volume arrives — the model produces at machine pace — and the human’s attention becomes the bottleneck. The organization responds rationally: streamline the review, batch the approvals, measure the reviewer’s throughput. Each rational step thins the examination. Within months, the sign-off is a formality the metrics reward for speed — and the formality is now load-bearing, because downstream processes trust it. I have watched organizations discover, after an incident, that the validation step everyone relied on had been ceremonial for a year. Nobody decided to make it ceremonial. Throughput did.

Why does refreshing the matrix keep the gap?

The false explanation is assignment: the RACI needs a refresh, the roles need clearer owners, the loop needs a named validator. I have watched the matrix get refreshed — a name in the validator cell, announced with relief — and the gap persist, because the name received the accountability without the function. The named validator still lacks time, information, and a working reject path. Refreshing the matrix answers the auditor’s question — who validates? — with a name. The name cannot validate. The gap is now harder to raise, because it appears to be closed: there is, after all, a name.

What pattern marks role confusion at the decision?

The marker I use is the stress test: I ask what happens when the human in the loop disagrees with the model, persistently, on a busy day. In functioning systems, the answer is a path — escalation, investigation, authority to halt. In confused systems, the answer is social: the person would raise it with their manager, probably, if it kept happening, though quotas are tight. When the rejection path runs through personal courage rather than designed process, the loop is decorative. The system has decided in advance that the model will be accepted, and the human’s role is to absorb the responsibility for that foregone conclusion.

The finished material passes through human review, and quite often that review reveals changes that are necessary before the result is usable. I regard this as part of the operating process rather than an unfortunate add-on. Generative systems can sound confident when they are wrong. The review is where the organization’s own knowledge re-enters the work — where someone can recognize the plausible sentence that is not true of this company, this customer, this quarter.

At our agency, AI supports ideation and execution. On one occasion we sent a client a document of roughly sixty pages after reviewing the first thirty, which looked strong. The second half was poor enough that the client rejected it. The generation had worked; the validation had stopped halfway through. The human name was on the document. The human judgment was not in all of it.

What question separates deciding from validating?

Governance conversations treat “human oversight” as one thing. The question that clarifies it splits the work apart: who decides what the system should do, who validates that an output is fit, who answers when the outcome harms someone. Three verbs, three functions, and in the confused organizations I visit, one underpaid person doing a fourth thing — clicking — while the three functions go unassigned. The split question is where the diagram’s figure starts becoming a role.

The question that separates the three roles: for this output, who prepared it, who validated it against the business’s own knowledge, and who decided — and are those three acts actually happening, or did the middle one get skipped because the first one looked finished?

PR excerpt (80–120 words, Miklos’s own words): AI may do substantial preparatory work. Qualified people must interpret results and own consequential business choices. There is a difference between an expert examining a consequential assertion and five people approving the same document without adding independent information — a sign-off is not meaningful merely because a human name appears next to it. The review is where the organization’s own knowledge re-enters the work: where someone can recognize the plausible sentence that is not true of this company, this customer, this quarter.