The Loop Has Grown. The Human Has Not.

It is late afternoon. On my screen, four markdown files sit side by side — summaries that Claude wrote from transcripts that Plaud captured earlier this week. I have just asked Claude to read across the four and pull out what matters. The synthesis it gives me back is coherent, and almost ready to inject it into Confluence.

Each step was accurate enough on its own. Each step saved time. But standing in front of the composite, I cannot tell any more what has been compressed away, what has been quietly recombined, what is mine and what is merely mine to defend.

I catch myself about to paste it into Confluence. And I pause — not because I doubt the content, but because I am not certain I can say I have read it, in the way that reading used to mean. I have seen it. I have moved through it. Whether I have judged it is another question.

That pause is where I want to start.

The question underneath that pause is not new, but it has changed shape. For as long as people have delegated parts of their work to machines, there has been some version of this moment — the accountant with a calculator, the editor with a spellchecker, the pilot with an autopilot. What is new is the texture of what has been delegated, and the scale at which it has been delegated. The machine is no longer executing a task I have defined. It is producing material I am meant to evaluate, in volumes I did not set and at a speed that does not wait for me.

I notice this in my own work. But the shape of the problem does not belong to my field. A doctor reviewing an AI-generated differential diagnosis, a lawyer skimming a contract summarised by a model, an HR manager scanning a shortlist produced by a hiring system, a policy officer reading a briefing that multiple AI agents have already touched — all of them are being asked to do something the language of oversight does not quite capture. They are being positioned as the final human judgment in a chain that is already largely beyond their attention.

The loop, in other words, has grown. The human at the end of it has not.

None of this would matter if it were only my experience. It is not.

In March of this year, Harvard Business Review published research conducted by a Boston Consulting Group team (Julie Bedard and colleagues), surveying nearly 1,500 full-time workers across industries. They asked people how they were using AI and how it was affecting their cognition and mood. The results are straightforward enough, and they map almost exactly onto the experience I described.

Workers with high degrees of AI oversight in their role reported fourteen per cent more mental effort on the job, twelve per cent more mental fatigue, and nineteen per cent more information overload than those with low oversight demands. At the end of the study, the team asked a direct question: had participants experienced mental fatigue from AI use or oversight that went beyond their cognitive capacity? Fourteen per cent said yes. They gave it a name: AI brain fry.

A curious finding sits alongside this. Productivity self-reports rose as people used one, two, and three AI tools in parallel. At four or more, productivity began to fall. The ceiling of effective orchestration, at least in this sample, sits at three.

Not everything about AI use pointed in the same direction. Workers who used AI to take over repetitive, low-value tasks reported fifteen per cent lower burnout than those who did not. The distinction matters. Burnout is an emotional exhaustion, and AI can genuinely relieve some of it. What the study identifies as AI brain fry is something else — an acute cognitive strain from sustained oversight, which no amount of tool-replacement of toil addresses.

The authors put this plainly: "AI oversight cannot simply be layered on top of human oversight." There is a ceiling to how many agents a single person can meaningfully oversee.

What strikes me about how this problem gets discussed is how much of it is framed as a human failing. People need to be more vigilant. Teams need better AI literacy. Organisations need to train their staff. All of this may be true, and none of it addresses the shape of the thing.

What the research describes is not a failure of attention. It is a failure of structure. The number of tools, the rhythm between them, the moments where a decision has to be made and the moments where there is no time to make one, the point at which a human is expected to catch what several layers of synthesis have altered — these are structural features. They are ingredients of a system. Someone, or something, decided them. Most of the time, no one decided them on purpose.

The designer Dan Hill has a term for this. He calls it dark matter: the invisible substance of organisations — policies, cultures, governance, unspoken conventions — that quietly determines what is possible. His work on the subject has shaped how I think about this.

The loop I described earlier is mostly dark matter. The interface you see on a screen is only the visible tip of it. Underneath sit choices — often not choices at all, but defaults — about how many tools run in parallel, how outputs get chained, which steps happen while you are doing something else, what gets logged and what disappears. None of this is neutral. All of it shapes whether the human at the end of the chain can still function as a human at the end of the chain, or whether they have become something else: a signature attached to a process they did not quite witness.

When such a system breaks, this is who remains. Not the pipeline. Not the tool. The person who pressed send.

I am not a lawyer, and this is not legal advice. But it is worth noting that the legal discussion is moving in the same direction as the research.

The European AI Act, in Article 14, already acknowledges automation bias by name — the tendency of humans to over-rely on AI outputs — and requires that operators of high-risk systems remain aware of it. The awareness requirement is striking, because the underlying research suggests that awareness alone does not prevent the bias. Legal scholars are beginning to pick at this gap. In a recent piece for Harvard Journal of Law and Technology, Nanda Min Htin argues that requiring humans to oversee systems they cannot cognitively oversee turns the operator into what others have called a liability sponge — the place where blame settles when the system as a whole fails.

I find this useful less for its legal mechanics than for what it confirms. Two very different fields — organisational research and tort law — are arriving at a similar observation from opposite directions. The person at the end of the loop is carrying more than they can carry, and the systems around them have not been designed with that in mind.

For anyone who wants to go deeper, Min Htin's article is worth reading in full.

If the problem is structural, the response has to be structural too. Not better oversight, but better design of the thing that needs overseeing — and design of the human position within it. This is where I think the most interesting work sits right now, and where the least of it is actually being done.

A few shifts seem worth naming.

From oversight to interaction. The moments where a human is placed in the loop are not organisational arrangements — they are design moments. Where a review happens, how it is framed, what is shown and what is hidden, how much time the interface quietly assumes you have — all of this is shaped by someone. If no one has shaped it on purpose, it has been shaped by default, which is almost always worse.

From efficiency to cadence. Systems at the moment are built to minimise the gap between outputs. The assumption is that faster is better, that the waiting human is a bottleneck to be reduced. But judgement requires a different rhythm than production. The pause I described at the start is not wasted time. It is the time during which judgement actually happens. A system that does not make room for it is not efficient. It has simply moved the cost somewhere less visible.

From stacking to restraint. The research I cited earlier suggests that productivity gains reverse after three concurrent AI tools. That is a design signal, not a personal discipline problem. It tells us something about the shape of orchestration that works, and the shape that does not. Treating this as an individual failing — "you should be able to handle more" — misses where the limit actually lives.

From checking output to reading reasoning. If a human is expected to judge something, they need something to judge. Not a final answer, but a trace — what was considered, what was discarded, what the system is uncertain about, where the gaps are. Without this, oversight collapses into approval, and approval is not the same thing.

None of this is a toolkit. It is a set of questions about where the real design work lives when AI is in the picture. It lives in the invisible layer, in the structure of the loop, in the cadence of attention, in what the human is actually given to work with. These are not HR problems or IT problems. They are not legal problems either, though law will eventually follow them. They are design problems, in the fullest sense of the word — and the people best positioned to address them are rarely the ones being asked.

The pause I described at the start of this post was, for a moment, mine alone. But I have since learned that it is not. Researchers are measuring it. Lawyers are writing about it. Workers from a dozen fields describe versions of it. What began as a private moment of disorientation turns out to be one of the more common experiences of this particular stretch of technological history.

The question is not whether AI will play a role in work that requires judgement. It will. The question is what conditions make that work still worth doing — for the person doing it, and for everyone downstream of their decisions. Those conditions have to be designed. By someone. On purpose.

That is the work I find myself increasingly interested in.

I am interested in how this is playing out inside organisations that are actually using AI at scale. If you are in one, and something here landed, find me on LinkedIn. I read everything.

What I read

The research on AI brain fry: Julie Bedard, Matthew Kropp, Megan Hsu, Olivia T. Karaman, Jason Hawes, and Gabriella Rosen Kellerman, "AI Brain Fry," Harvard Business Review, 5 March 2026. Study conducted by Boston Consulting Group among 1,488 full-time US workers, January 2026.

On automation bias and its limits: Raja Parasuraman and Dietrich H. Manzey, "Complacency and Bias in Human Use of Automation: An Attentional Integration," Human Factors, 2010 — the foundational work showing that automation bias cannot be overcome by awareness or training alone.

On the moral crumple zone: Madeleine Clare Elish, "Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction," Engaging Science, Technology, and Society, 2019 — introduces the concept of the human operator as the component that absorbs moral and legal responsibility when automated systems fail.

On the liability question: Nanda Min Htin, "Redefining the Standard of Human Oversight for AI Negligence," Harvard Journal of Law & Technology, 9 February 2026.

On the regulatory frame: Regulation (EU) 2024/1689 (the AI Act), particularly Article 14 on human oversight.

On dark matter in design: Dan Hill, Dark Matter and Trojan Horses: A Strategic Design Vocabulary, Strelka Press, 2012.