05｜Public Essay｜Human-in-the-loop Is Not the Same as Judgment-in-the-loop

Rebuilding Your Cognitive OS in the AI Era

SubscribeSign in

05｜Public Essay｜Human-in-the-loop Is Not the Same as Judgment-in-the-loop<br>Confirmation is not judgment, permission is not understanding, and approval is not responsibility; the real question in the age of AI is not whether humans appear in the workflow, but whether human ju

KunYuan<br>Jun 21, 2026

Share

I’m KunYuan.<br>Among this first group of entrance essays, if I could leave only one distinction, it might be this:<br>Human-in-the-loop is not the same as judgment-in-the-loop.<br>A human being can remain in the loop while human judgment has already left the process.<br>This may sound like a technical distinction, but I increasingly think it names one of the most overlooked issues of the AI era. Many products, systems, and agent frameworks have already designed humans into the workflow. Humans can confirm, review, authorize, approve, and say yes or no at the final step. On the surface, the human has not disappeared; the human is still in the loop.<br>But the real question is whether human judgment is still being formed inside the process.<br>This is not an abstract problem. It is already appearing in AI products, coding agents, automated workflows, content generation systems, organizational decision processes, and everyday AI use. Many systems emphasize human-in-the-loop because it sounds safer, and in many cases it is indeed necessary. Without human confirmation, the system cannot continue; without human authorization, an agent cannot access certain files, run certain commands, or call certain tools; without human approval, a result cannot be published, merged, sent, or deployed.<br>These designs are not meaningless. On the contrary, they are often created for safety, responsibility, and control. The problem is not that human-in-the-loop has no value, but that we can easily mistake it for being sufficient.<br>In many cases, human-in-the-loop only guarantees that a human has been placed somewhere inside the workflow. It does not automatically guarantee that human judgment is participating in the definition of the problem, the evaluation of reasons, the setting of permission boundaries, the rejection of errors, or the taking of responsibility.<br>One common form is confirmation.<br>AI generates a piece of text, a proposal, a summary, a piece of code, or an analysis. A human looks at it at the end and clicks approve, accept, publish, merge, send, or deploy. This action looks important because without it the workflow cannot be completed. But final confirmation is not the same as judgment.<br>Real judgment is not merely checking whether something looks acceptable, whether there are obvious mistakes, whether the format is normal, or whether it seems “good enough.” Real judgment means that I understand how this result was formed, I know where its reasons are, I know where it may be weak, I know why I am keeping it, and I know what I am willing to take responsibility for.<br>If those processes did not happen, confirmation is only confirmation; it is not judgment.<br>A second common form is permission.<br>This is especially visible in coding agents and automated agent systems. In the name of safety, the system repeatedly asks whether the human allows the next step: whether it may read files, modify code, run commands, access external tools, continue execution, or expand its permissions.<br>On the surface, these are safety mechanisms. The human still has the final authority to grant permission. The system has not bypassed the human; it has handed each step back to the human for approval.<br>But in real use, something else often happens. There are too many confirmations, too many permission prompts, too many interruptions to the flow of work, and the human begins to feel the friction. Very quickly, the more convenient path appears: always allow, approve all, trust this workspace, grant full access, continue automatically.<br>This is not necessarily because the user is irresponsible. Often, it is because the workflow, the pressure for efficiency, and the design of the product together train the user to behave this way. A person is placed in the loop for safety, but when they are repeatedly asked to perform low-quality, repetitive, fragmented acts of authorization, authorization soon becomes a default gesture. In the end, human-in-the-loop becomes a one-time ritual of permission.<br>The system asked. The human allowed it. The workflow looks compliant. But did the human truly understand what was opened, where the boundary was, what the risk was, and who would carry the consequence?<br>That is the problem.<br>A third form is review.<br>Many AI systems say: do not worry, a human will review it. AI only generates the draft, the suggestion, the analysis, or the execution plan; a human still checks it at the end.<br>This is certainly better than having no human involved at all. But the word “review” can also create a false sense of safety. In reality,...