Is “human-in-the-loop” a coping mechanism because we are scared of the future?

You want to know the truth? I’m in a loop right now.

Our studio has a literal one. An AI makes edits to one of our sites from a Slack thread. Nothing goes live until a named person replies “approve.” Some days that person is me.

So I’m not here to tell you the loop is a lie. I use AI every day. I’m here to ask a harder question: when I type “approve,” what am I actually doing?

The most comforting phrase in tech

“Man in the loop” was control-engineering language by the 1960s. NASA was using it in 1967.¹ The version we use now, with a human in, on, or out of the loop, came out of the debate over autonomous weapons.²

Today it’s on every AI slide. Somewhere between the architecture diagram and the pricing, there’s a rounded rectangle that says a human reviews everything.

I get why. People are nervous. Half of Americans say they’re more concerned than excited about AI, up from 37% four years earlier.³ Only 27% trust businesses to use it responsibly.⁴ When trust is that low, “a person checks it” is the cheapest reassurance you can buy.

Fair point: the loop makes sense today

Models still make things up, and they do it with confidence. In 2023 a New York lawyer was fined $5,000 for filing a brief with six cases that didn’t exist.⁵ One researcher now tracks these. His list had 2,292 cases on it this week.⁵

When a wrong answer is public, a human checkpoint isn’t a coping mechanism. It’s the job.

But look at that list again. Every one of those filings had a human in the loop. A licensed one. They just didn’t check.

What the research says about the human

This is the part that changed how I think about it. Having a person in the loop and having oversight are not the same thing.

Psychologists call it automation bias. When a system is usually right, people stop checking it. It hits experts as much as beginners, and training doesn’t fix it.⁶ Europe’s AI Act, the law that asks for human oversight, names automation bias as a risk in the same article.⁷ The law that requires the loop warns you the loop is leaky. Its start date also slipped 16 months.⁸

And the law never says “human in the loop.” It says systems have to be built so people can effectively oversee them. Colorado’s AI law was replaced before it ever took effect. The new version offers human review on request, to the extent it’s commercially reasonable.⁹ That’s a long way from a person checking everything.

The examples aren’t fun. In the Dutch childcare benefits scandal, civil servants reviewed every case the algorithm flagged. They weren’t told why it was flagged. Thousands of families were wrongly accused of fraud and the government resigned.¹⁰ ProPublica reported that doctors at one US insurer spent about 1.2 seconds on each claim they denied. The insurer disputes that.¹¹

And when it goes wrong, the human takes the fall. After a self-driving test car killed a pedestrian in Arizona, the safety operator was convicted. The company wasn’t charged.¹² One researcher has a name for that role: the moral crumple zone.¹³ You’re not there to steer. You’re there to absorb the impact.

Then there’s the big one. In 2024, a team at MIT pulled together 106 experiments that compared people, AI, and people working with AI.¹⁴

On average, the combination did worse than the stronger of the two working alone. It was worst on decisions, which is exactly what “approve or reject” is. It helped when people were making something, and when the human was the one who was actually better at the task.

Read that as a designer and it lands hard. The loop fails where the human is a checkpoint. It works where the human is the craft.

It works when the role is designed

There are real wins. In a Swedish trial with 106,000 women, AI-supported breast screening led to 12% fewer cancers missed between screenings, with 44% less reading work for radiologists.¹⁵

That wasn’t a person bolted onto a machine as a safety label. Somebody designed it. The AI sorts. The radiologists read what needs reading. Each side does what it’s good at. That’s all design.

The famous consulting study says the same thing from another angle. Harvard and BCG gave 758 consultants an AI model. On tasks the model was suited for, they finished 12% more work, 25% faster, at over 40% higher quality. On a task just outside what it could do, they were 19 points less likely to get it right.¹⁶ Same people, same tool. Knowing which side of that line you’re standing on is the skill.

What I see from inside a studio

Different toolset, same job. That’s how I describe it to our team. But the tools have changed who is afraid of what.

We had a brand client pause a project and, after a few rounds of logos, suggest that maybe AI could just make the logo. Don’t even get me started. The tools are open to everybody. Clients keep paying us anyway, because the people who use AI mostly don’t know how to use AI. Knowing what to ask for, and knowing when the answer is wrong, turns out to be the work.

Another client took a screenshot of a prototype, asked an assistant to critique it, and sent us the result. It was a dissertation. No context about their users, their business, or what we’d already decided together. That’s a human in the loop in the most literal sense. A person passed it along. Nobody judged it.

And a software client worried a texture in their identity would read as “built with AI” in six months. That one’s interesting. The fear wasn’t the tool. It was looking like nobody chose it. The answer is the same as it has always been: intention. Put it where it means something.

Designers feel all of this. In Figma’s 2025 survey, 78% said AI makes them more efficient. Only 32% said they can rely on what it gives them.¹⁷ Faster, and not trusted. That gap is where the human lives.

It’s also why I tell our designers that skill caps out at some point and the questions take over. You have to get extremely inquisitive. We’re strategists, not pixel pushers.

Sensible today, a crutch tomorrow

The loop itself is already moving. Agents work longer on their own. One research group estimates the length of task a model can finish alone has been doubling about every four months, with wide error bars.¹⁸ Anthropic looked at how people supervise its coding agent. New users let it run without sign-off in about 20% of sessions. Experienced users, more than 40%. But the experienced ones interrupt more often, not less.¹⁹

That’s the shift. From approving every step to watching and steering. From in the loop to on it.

So is it a coping mechanism? Sometimes, yes. When “a human reviews every output” is the pitch, you’ve quietly said the human adds nothing but permission. That’s a comfort blanket.

When the human brings taste, context and accountability, the loop isn’t a safety feature. It’s the product.

How we use it without hiding behind it

• Name the person and the decision. Not “a human.” Who, and what they’re answering for.
• Put people where they’re the stronger party. Making, taste, context, consent. Not skimming the hundredth output.
• Give the reviewer the why. And the time, and the authority to say no. A reviewer without those is the crumple zone.
• Design the handoff. Decide what the model does alone and what always comes to a person. Write it down.
• Sell judgment, not supervision. Nobody hires a studio for its approval button.

A test we use. Take the human out for one cycle. If nothing changes, they were never in it. If the work gets faster but worse in ways a client would notice, you’ve found the job.

My job is to get us from 0 to 1 faster. AI helps with that every single day. It hasn’t once told me whether the 1 was any good.

The ruleDon’t sell the loop. Sell the judgment. The industry is shifting. Keep up.

Sources

1. NASA contractor report, “Man in the loop,” 1967. ntrs.nasa.gov
2. Human Rights Watch, “Losing Humanity: The Case against Killer Robots,” 2012. hrw.org
3. Pew Research Center, “How Americans View AI and Its Impact on People and Society,” 2025. pewresearch.org
4. Gallup, AI attitudes survey, May 2026. news.gallup.com
5. Damien Charlotin, AI Hallucination Cases database (includes Mata v. Avianca, 2023), viewed 12 November 2026. damiencharlotin.com
6. Parasuraman and Manzey, “Complacency and Bias in Human Use of Automation,” Human Factors, 2010. pubmed.ncbi.nlm.nih.gov
7. EU Artificial Intelligence Act, Article 14: Human Oversight. artificialintelligenceact.eu
8. National Law Review, “EU Digital Omnibus on AI enters into force,” 2026. natlawreview.com
9. Constangy, on Colorado’s repeal and replacement of its 2024 AI law, 2026. constangy.com
10. Amnesty International, “Xenophobic machines: the Dutch child benefit scandal,” 2021. amnesty.org
11. ProPublica, “How Cigna Saves Millions by Having Its Doctors Reject Claims Without Reading Them,” 2023. propublica.org
12. National Transportation Safety Board, Tempe collision investigation, 2019. ntsb.gov
13. Madeleine Clare Elish, “Moral Crumple Zones,” 2019. papers.ssrn.com
14. Vaccaro, Almaatouq and Malone, “When combinations of humans and AI are useful,” Nature Human Behaviour, 2024. pmc.ncbi.nlm.nih.gov
15. Lund University on the MASAI trial, The Lancet, 2026. lunduniversity.lu.se
16. Dell’Acqua et al., “Navigating the Jagged Technological Frontier,” Harvard Business School, 2023. papers.ssrn.com
17. Figma, 2025 AI Report. figma.com
18. METR, task time horizons. metr.org
19. Anthropic, “Measuring agent autonomy,” February 2026. anthropic.com