Who Is Really in Control?

Much of the global conversation about advanced AI converges on one urgent question:

How do we keep increasingly capable AI under meaningful human control?

It is an essential question.

But it may not be the only one.

Because keeping a human somewhere in the decision process does not necessarily tell us how much meaningful human judgment remains.

And that leads to another question:

What does it actually mean for a human to remain in control?

THE REASSURING ANSWER: KEEP A HUMAN IN THE LOOP

One widely used principle is straightforward:

AI can analyze, predict or recommend—but a human remains involved.

The intuition is compelling. Keep the final decision in human hands, and Human Agency is preserved.

Yet evidence suggests the relationship is more complicated.

In a peer-reviewed 2024 experiment involving 292 participants, researchers Daniela Sele and Marina Chugunova examined how people used recommendations when they retained the ability to monitor and adjust them.

Allowing participants to monitor and adjust recommendations increased their preference for algorithmic recommendations by seven percentage points. Participants in the human-in-the-loop condition also reported greater confidence in the predictions they submitted.

But retaining the ability to intervene did not guarantee better oversight.

In this particular experimental task, adjustments made by human monitors reduced average decision accuracy. Participants were also less likely to intervene when recommendations contained larger errors; when they did intervene, they corrected those larger errors by less.

The study does not establish that human oversight is generally ineffective. The authors explicitly note that outcomes can depend on the task, the performance and reliability of automated support, the expertise of human monitors, and the information available to them.

But the findings reveal a distinction worth examining:

Human in the Loop ≠ Human in Control.

Not as a universal conclusion—but as a distinction worth examining.

LOOK AGAIN

Imagine an AI-assisted decision.

The human still signs off at the end.

Formally, the authority remains human.

But before that final approval, an AI system may have participated in determining what information is presented, how the problem is framed, which alternatives are surfaced, what appears relevant, or which action is recommended.

The important question may therefore be deeper than:

Who makes the final decision?

It may also be:

Who shaped the field from which that decision emerged?

That question becomes increasingly consequential as AI systems move beyond answering questions toward carrying out more complex sequences of work.

A 2026 position paper by Margaret Mitchell, Avijit Ghosh and Samir Passi argues that current approaches to AI-agent development can impede effective human oversight. The authors further argue that extended reliance on AI systems can degrade some of the cognitive capacities effective oversight requires, and they call for systems and organizational practices designed to support human critical judgment.

This is a position paper, not conclusive empirical evidence.

But the question it raises deserves attention.

WEIGH AGAIN

Perhaps Human Agency cannot be preserved merely by assigning a human the final click.

Meaningful agency may also depend on whether the human can:

• notice what matters,
• recognize what is influencing perception,
• question the framing,
• examine alternatives,
• give relevant factors appropriate weight,
• and ultimately,
• choose intentionally.

That changes the AI discussion.

The problem is no longer only:

How capable should AI become?

or:

How should AI be governed?

A third question appears:

What must humans remain capable of doing as AI becomes more capable around them?

THE CLARITY LENS

This is one of the questions behind the Clarity Project.

Clarity is being explored as a Human Capability for the Age of AI: the capacity to see more clearly what is shaping perception and judgment, distinguish what is relevant, give those influences appropriate weight, and make more intentional choices.

This is not proposed as a replacement for AI safety, regulation, technical safeguards or institutional accountability.

It is a different layer of the problem:

the human-capability layer.

If Human Agency is to remain meaningful in an increasingly AI-mediated world, protecting the formal right to decide may not be enough.

We may also need to strengthen the human capability required to exercise that right well.

That remains a proposition to investigate—not a conclusion to protect.

THE OPEN QUESTION

So perhaps the question is not simply:

Can humans remain in control of AI?

It is also:

As intelligence becomes more capable around us, what must become stronger within us?

See Again. Weigh Again.