If AI Decides When Humans Should Intervene, What Should Humans Decide?

한국어 원문 읽기 →

Human Directing Research Note 5
The shift in the human role → delegation → validation criteria → exception judgment and intervention → revisiting judgment criteria and the scope of delegation

Human Directing Research Notes 5 — Revisiting judgment criteria and the scope of delegation

The central question: If AI detects exceptions and risks and decides when to call a human, where does human judgment move?

The first four notes explored how human roles and forms of intervention change as AI advances. Our focus moved from giving detailed instructions to deciding what to delegate and how far, then to setting criteria for accepting results and intervening in important exceptions.

Yet the fourth note ended with another question.

What happens if AI also becomes good at identifying exceptions, assessing risks, and asking people to make a decision when needed? Is it still sufficient to describe the human role as “the person who steps in at important moments”?

This note follows that question.

Calling a human already involves judgment

Suppose an AI says during a task, “This part needs your review.”

The request may look like a simple notification, but several judgments precede it. The system must distinguish what counts as an exception, how much uncertainty is acceptable, how consequential an outcome must be before it stops, and what can continue without intervention.

People might specify all these conditions. Alternatively, they might let AI propose the conditions or interpret them in context.

An important issue follows.

If AI decides when to call a human, that decision also shapes the range of problems the human gets to see.

A person might assess the exceptions AI presents well while missing problems AI did not classify as exceptions and therefore did not present. Responding correctly when called does not, by itself, establish that the overall judgment structure was appropriate.

From reviewing results to reviewing criteria

The third note asked, “What results should we accept as sufficient?” The fourth asked, “Under what conditions should a human intervene?”

Now we need to question the criteria themselves.

Do our completion criteria fit the actual purpose?
Are important problems being identified as exceptions?
Does unnecessary review repeatedly interrupt the work?
Is a scope of delegation that was appropriate initially still appropriate now?

For example, checking spelling and sentence structure is useful when writing a document. But if the purpose is to keep a research record, those checks are insufficient. We also need to examine whether the writing connects with the existing research questions, distinguishes observations from hypotheses, and avoids presenting untested claims as conclusions.

Assessing whether prose reads smoothly and assessing whether the research direction is appropriate answer different questions.

Checking whether a result meets the criteria and checking whether those criteria are appropriate need to be distinguished.

Actual observation: Preparing the fifth note also required a check on direction

While preparing this note, I asked AI to review the earlier articles and choose a topic for the fifth. I then emphasized an additional condition:

“The article must stay within the topics, questions, and concepts we have been exploring in Human Directing.”

What I did directly in that exchange was not edit individual sentences or prescribe a detailed structure. I clarified the criteria that should guide topic selection and writing.

Even a topic that follows naturally from the previous article may be unsuitable if it departs from the questions we have been investigating. Drifting into a guide to using AI, or rushing to conclude that some capability must remain human, would also depart from the research’s purpose.

However, this case does not establish that only humans can judge research direction. Had AI adequately reviewed the research context and applied the same conditions from the outset, my additional intervention might have been unnecessary.

What the case establishes is narrower.

In this task, a person restated the research scope, and that condition was presented as a criterion for selecting the next topic. We need further observation to understand when such intervention is necessary and when it becomes unnecessary.

What if AI can also propose the purpose?

We might conclude, “Ultimately, humans just need to set the purpose.” But that is not enough to close the inquiry.

We can also imagine AI inferring a purpose from a user’s previous choices and present circumstances, then suggesting a more suitable goal. When a person’s initial purposes are unclear or conflict with one another, AI’s questions and suggestions might help clarify them.

Purpose, then, cannot be explained only as something a person defines completely at the beginning and hands over for AI to execute. It may also be adjusted through conversation and work.

The question to examine becomes:

On what grounds should we accept the purposes and criteria AI proposes, and under what conditions should we change them again?

For example, AI might suggest a direction intended to increase an article’s page views. That proposal may be reasonable, but how much priority to give the accuracy of the research record and continuity of the inquiry requires separate consideration.

We need to distinguish AI’s ability to produce good proposals from the decision to adopt a proposal as the criterion for this activity.

The scope of delegated judgment must also be revisable

In Human Directing, delegation is difficult to understand as a decision made once and then settled.

If AI repeatedly produces sufficient results under the same conditions, human review can be reduced. Conversely, if it misses important conditions in a new situation, the scope of delegation may need to narrow or the review criteria may need to change.

We therefore need to consider two questions together.

“Can we delegate this task to AI this time?”

“Can we continue delegating it in the same way?”

Approving one result and continuing to delegate the same kind of judgment are different decisions. The first concerns an individual result; the second concerns how work will operate in the future.

In the current research hypothesis, this process of readjusting the scope of delegation is an important subject of observation. Rather than asking whether more human intervention is better, we need to examine whether experience provides grounds for increasing or reducing it.

Responsibility cannot be explained by a final approval alone

Even if AI performs much of the judgment, the fact that a person gave final approval does not adequately explain the structure of responsibility.

The chosen purpose, the information provided, the range of actions permitted, and the checks required all influence the outcome.

To examine responsibility, we therefore need to look beyond the final approver and consider how purpose-setting, delegation, validation, and revision took place.

In this research, responsibility is less a stage attached to the end of the task than a question of who can decide and revise what throughout the process of judgment and action.

In particular, if a person has approval authority but cannot understand the criteria or change the outcome, we need to ask whether that approval represented substantive judgment.

Counterexamples that could weaken this hypothesis

The discussion so far leads to a hypothesis: the human role may move from checking individual results to revisiting purposes, criteria, and the scope of delegation.

But we cannot conclude that this is a domain that will permanently remain human.

What if AI becomes better at identifying conflicting purposes, detecting gaps in validation criteria, and appropriately adjusting the scope of delegation? What if human revisions instead introduce unnecessary delays or errors?

If such cases recur, the current hypothesis must change.

Conversely, if important omissions in purpose or flawed criteria repeatedly become visible only after human intervention, we need to specify the conditions under which that intervention helped.

To distinguish these possibilities, recording successful outcomes alone is insufficient. We also need records of the criteria people changed, their reasons, the outcomes after those changes, and cases in which the results were sufficient without human intervention.

The usefulness of Human Directing depends on whether it can explain which judgments and interventions helped in actual collaboration, rather than merely explain that humans still have a role.

Fact, hypothesis, and interpretation

Observed fact: While preparing this article, I restated the condition that AI should preserve the existing research scope. The earlier notes also treated the scope of delegation, completion criteria, and conditions for exceptions and intervention as central questions.

Current hypothesis: As AI takes on more exception detection and decisions about when to intervene, the human role may move from handling individual exceptions to revisiting the appropriateness of purposes, judgment criteria, and the scope of delegation.

Interpretation: Human Directing can be described not only as the ability to give AI direction once, but also as a process of continually revising direction and the structure of judgment through experience. Which parts people should perform and which can be delegated to AI require continued testing.

The next research question

This note explored the hypothesis that, when AI also decides when to intervene, the object of human review may expand beyond results to include judgment criteria and the structure of delegation.

Yet a condition underlying this hypothesis remains unresolved.

If people are to review and revise AI’s judgment criteria, how can they retain the ability to understand those judgments and compare alternative possibilities?

Can we delegate more judgment while remaining able to intervene meaningfully when needed? The next note will continue with this question.

Human Directing is a provisional concept for tracking how human roles, judgment, and forms of intervention change as AI advances. This article is an open research note that revises hypotheses through actual use and counterexamples, rather than a presentation of an established theory.