I’m going to say something that might sound reckless: In 2026, “Human-in-the-Loop” may become the new “Red Flag Act.”

In 1865, the UK required a man to walk 60 yards ahead of every automobile, waving a red flag to warn pedestrians. It wasn’t about safety, it was about comfort. It made the new technology legible to our intuition. And it may have held back the British car industry back for 30 years.

Are we, right now, waving a red flag in front of AI?

The insistence that every decision be explainable in human language, that every output wait for human approval - it feels responsible. But is it actually making us safer? Or are we just paying a “Comfort Tax”?

In a lot of agent conversations lately, ‘Human-in-the-Loop’ has started to sound non-negotiable. The rationale is that AI needs our collective thought, ethics, and life experience to guide it. It feels safe. It feels responsible. But there is a part of me that asks a somewhat uncomfortable question:

Are we so sure that human experience is the ultimate benchmark for intelligence?

Isn’t it incredibly presumptuous of us to believe that the human way to solve a problem is the only way - or even the best way.

Now don’t get me wrong, I still want a human pilot to be in charge of a flight I’m on, and I hate talking to a customer service robot. I’m also not referring to the “automation” tasks that AI Agents are doing and will be doing. I’m referring to say the KYC process for a bank. We still need agents to do OCR, process rules etc, but what if we have to re-think the concept of authentication itself ?

Take KYC. Strip away the forms and PDFs and it is just this: how do you prove you are you, reliably, consistently. For centuries, the gold standard here was the “wet signature”; that weird, messy scribble on a piece of paper. It was chaotic, but it was effective because it was behavioral. It captured the unique muscle memory of your hand, and tied that to your identity. It was a physical non-fungible token.

When we moved to the digital age, we couldn’t digitize that nuance (not very well), so we downgraded to the Password. We replaced a unique behavioral trait with a static secret. We forced humans to memorize strings of text and numbers - a task our brains are terrible at. Resulting in an entire industry of password managers and the friction of Multi-Factor Authentication (MFA).

We are stuck in this loop because we are trying to solve identity in a way that humans can understand.

But if we let an AI solve identity from first principles? It wouldn’t ask for a password. It would maybe ask for a quick fingerprint/retina scan and combine it with some behavioural aspects - cadence of your typing, the specific jitter of your mouse cursor, the way you hold your phone. What if an AI discovered identity signals that make no intuitive sense to us - but correlate strongly with fraud or authenticity? At what point do we trust a signal we can’t emotionally understand?

Consider the history of Aviation. For decades, regulators enforced the “60-Minute Rule.” The logic was human and intuitive: flying over the ocean with only two engines is dangerous. If one fails, you are in trouble. Therefore, twin-engine jets were banned from flying more than 60 minutes away from an airport. This forced airlines to fly inefficient, curved routes that hugged coastlines. It wasted millions of gallons of fuel and added hours to global travel.

The data disagreed. Engineers knew that modern turbines were so reliable that the “safety in numbers” intuition was mathematically false. But we ignored the data because we preferred the feeling of safety. We paid a “Comfort Tax” of wasted time and fuel for 30 years until we finally trusted the math and allowed direct trans-oceanic flights.

We did the same thing with cars—the Red Flag Act I mentioned earlier held Britain back for three decades.

If we want to actually move forward, we have to ask: Are we becoming the bottleneck in the loop?

The Era of “Move 37”

There’s a counter-history to this. A series of moments where we let the AI off the leash and instead of chaos, we got something better.

  • In 2006, NASA gave an evolutionary algorithm a simple job: design an antenna for a spacecraft. No templates. Just physics and constraints. The result looked like a bent paperclip; an ugly, asymmetric thing that no human engineer would sign off on. It also worked better than anything we’d ever designed.
  • A decade later, AlphaGo played Move 37 against Lee Sedol. Commentators thought it was a bug. It wasn’t. It was a move no human would play because no human had ever conceived of the position it was setting up. A glimpse of a game that had been played for 2,500 years, seen from an angle we’d never considered.
  • And in 2022, DeepMind’s AlphaTensor found a faster way to multiply matrices; a problem computer scientists assumed was solved fifty years ago.

In all these cases, a “Human-in-the-Loop” would have been a liability. We would have “corrected” the antenna to look symmetric. We might have “fixed” Move 37 thinking it was a hallucination. We would have slowed progress down to the speed of human understanding.

I’m not arguing for blind trust. I’m arguing that “I can explain it in plain English” is a bad proxy for “it’s safe and correct.” Some of the most useful signals in complex systems are real, measurable, and still hard to narrate. If we force every decision into a story that feels satisfying, we will systematically discard the advantages that don’t fit the story.

This doesn’t mean we surrender control. It means we change our role.

We need to move from Human-in-the-Loop (where we are the bottleneck, micromanaging every step) to Human-on-the-Loop.

Think of it less like an architect drawing a static plan, and more like a surfer on a massive wave. You don’t control the ocean; you can’t dictate the shape of every swell. But if you have the skill, you can harness that momentum to go places you couldn’t reach by swimming.

You define the destination, you set the safety boundaries, and then you let the physics take over.

Will we fall? Absolutely. There will be hallucinations, errors, and moments where things go spectacularly wrong. But getting back up, recalibrating, and finding the balance is part of the process.

The mistake in 2026 would be treating “Human-in-the-loop” as a moral principle instead of an engineering decision. Put humans where judgment is real: defining goals, setting constraints, designing fail-safes, controlling outcomes, and deciding what happens when the system is wrong. But stop forcing every good decision to be explainable in human language? That is how you end up flying a powerful aircraft along the coastline, not because it is safer, but because it feels safer.