Rethinking the Human Loop: Is the final human decision on AI an illusion?

“In the end, the human must decide.”

I think that’s nonsense.

This platitude has been around for quite some time now, and it always surfaces whenever AI is discussed. It refers to the “Human-in-the-Loop”, meaning the AI is welcome to prepare everything, but ultimately, the human checks and decides. It serves as a collective reassurance: Our jobs are changing, but don’t worry, we will still be needed. Just as architects, rather than clerks.

This is quite similar to, or perhaps even the successor of, the prompt engineer. Do you remember that? For a while, there was an idea that better prompts would lead to better results, and that this was an art that had to be carried out by specific corporate roles. Namely, the prompt engineer. Today we know: models manage to optimize prompts quite well on their own, at least since the introduction of reasoning capabilities.

If you look closely, the human-in-the-loop is already crumbling today. The developer who clicks “Accept & Commit” because they don’t really read their sixth AI pull request of the morning. The manager who skims the contract clause and approves it because asking questions would halt the whole process. The four-eyes principle, where the second pair of eyes has long since started blindly trusting the first.

Trust, once established with humans or technology, is an invitation to stop checking. Attention is expensive, and we are frugal with what is expensive. The human in the loop is often already no longer a reviewer today. These are merely workflow rituals.

Why shouldn’t the review process also be manageable by specialized reviewer agents? As soon as you are able to translate an abstract gut feeling into a concrete check logic, an AI can do it too. Only matters of taste are difficult to delegate to AI agents.

The uncomfortable consequence: The next generation of review processes will run in reverse. The human will not control the AI. The AI will control the human. It will proofread the lawyer’s draft before it goes to the client. It will check the sales quote against the policy. It will flag the inconsistency in the spreadsheet that the controller missed at half past five. Simply because the AI stays awake when we get tired.

Of course, there are objections. The loudest being: liability. But liability has never been equivalent to checking. Executives have long signed reports they do not understand in detail; they are liable anyway. The law regulates who pays the bill, not who does the work. Moreover, the law is changeable, and AI liability issues are still to be clarified.

The more subtle objection comes from AI safety: Scalable Oversight. How do we control systems that become smarter than us? A serious question, but its answer is not the human controller at the end of the chain. It is the human architect at the beginning, who sets the guardrails before the system starts running.

That is the actual shift. The human no longer decides at the end, but at the beginning—namely, on goals, boundaries, and values. In between, the system works, and the system will check itself, including those who operate it.

Anyone who finds this dystopian should pause for a moment: the human as the final instance of scrutiny is not a law of nature, but a habit. And habits rarely outlive reality for long.