The manifesto left the two loops as a picture. A machine making on one side, a person feeling on the other, and us in the middle, keeping time. A picture is enough to start an argument. It isn't enough to build a book on.
The joint is where every later chapter puts its weight, so it is worth getting right. I'll start with the machine's loop, because it's the one everybody can already see.
The machine's loop
The machine's loop is a making loop. It takes something in, produces something out, takes the result back as the next thing in, and produces again. Round it goes. A draft becomes a prompt for a better draft. A summary becomes the input to a summary of summaries. The loop doesn't stop when the output is good enough. It stops when we stop it, or when it runs out of things to chew.
The important part is what the loop is made of. It runs on representations: tokens, patterns, the model's own sense of what tends to follow what. The tokens point only at one another, never at anything in the world, the way a dictionary defines every word with other words and never points out the window. The field has a name for it: the symbol grounding problem. It never leaves that space. It has no way to.
Inside the loop, a sentence about a grieving widow and a sentence about a parking fine are the same kind of object, weighed the same way, made of the same stuff. The loop is fluent and fast, entirely indoors.
This used to be fine, because the loop was slow and needed us to turn it. We wrote the thing, we judged the thing, we fed it back ourselves. The machine was a tool we picked up between decisions.
The loop now turns on its own. It can make, read its own output, decide it's lacking, and make again, all before we've finished reading the first version. We used to be the part that turned it.
Most of the worry about AI is aimed at the wrong thing. The worry says the machine makes mistakes. It does, but we've always made mistakes, and we have ways of catching those.
The deeper change is that the machine now closes its own loop. It no longer waits at the point where we used to step in. If we want to be in the work, we have to put ourselves there on purpose. Nobody leaves us a gap any more.
The human's loop
The other loop is older, and it has nothing to do with computers.
A person meets a made thing. It lands on them some way: clear or baffling, warm or cold, worth their trust or not quite. How it lands changes what they do next. What they do next changes what they need. And what they need is the next thing somebody has to make. That's a loop too. It has been turning since long before any of us, every time a person used a thing another person made and was helped or let down by it.
This loop doesn't run on tokens. It runs on experience, in the proper sense: what it actually cost the person to use the thing. Whether they understood the letter or read it three times and still phoned a friend. Whether the form let them say the true thing or forced them to round it off into a box that didn't fit. Whether they came away feeling dealt with or fobbed off.
None of that lives inside the machine's loop, because none of it is a property of the output. It's a property of the meeting between the output and a life.
That's the loop user‑centred design has always worked on. Whatever the craft, we've stood at the point where the made thing meets the person, and we've taken that meeting seriously as the thing that matters.
Here's the part the machine can't borrow. The machine makes the thing. It doesn't feel it land. We do.
We know from the inside what it is to be on the receiving end of something, to be confused by it or reassured by it, because we have been on the receiving end of ten thousand things ourselves. That knowing is not information the model is missing. It's a different kind of thing entirely, and no amount of the first kind adds up to the second.
The coupling
We are not in the machine's loop. We are not in the person's loop either. That one belongs to them. We are the coupling between the two.
That's the practitioner's job, and it is the whole of it. We carry what we learn from the human loop into the machine's loop, so the machine makes things shaped by how they'll land. And we carry what the machine makes back out to the human loop, watching how it actually lands, so the next turn is better.
A coupling, in the plain mechanical sense, is the part that joins two turning shafts so that force passes from one to the other. The thing about a good coupling is that it lets the two shafts stay two. They turn together, in time, but they remain two shafts. Weld them into one and you haven't improved the coupling. You've removed the reason there was a coupling at all.
Two directions, kept in time. That timing is the whole craft, and it's why the brand is called what it is. Loopwork is the work of keeping the loops in rhythm.
The temptation, always, is to fuse them. To fold the human into the machine's loop, let the model both make the thing and judge how it lands, and call that efficient. The reason not to isn't sentiment. It's that the human loop is only worth anything because it stands outside the making. You can only feel how a thing arrives if you are standing where it arrives, which is to say outside the machine that sent it.
Fuse the two and what you keep is the machine's loop alone, judging its own output by its own lights, with no one left standing where the person stands. The vantage point goes. And that is what we're guarding: somewhere to stand outside the machine, where you can still feel how the thing lands.
The human loop is four crafts
So far I've written "we" as if user‑centred design were one thing. It isn't, and the difference matters more now, not less. Four crafts work the human loop, each responsible for something the others can't cover: understanding the people, writing so they can act, connecting the whole path, and shaping how they and the system act together.
A fair question: aren't these four blending into one? In places they are, and that's no bad thing. More practitioners work T‑shaped now, one person carrying more of the loop than before. But the functions don't blend the way the titles do. Each still has to be done, by someone, on purpose. Stop performing any one of them and the machine fills the gap with material that passes for true.
The machine's counter‑offer is to cover all four at once, from no standpoint in particular. There's a name for that dream of a single all‑seeing view: the view from nowhere. It is the one place a person can never be met from. They are met from somewhere, and where you stand decides what you can see.
User research is responsible for what's actually true of the people: what they do rather than what they say, what they need rather than what we assumed. The machine offers a tempting stand‑in: synthetic users, simulated panels, a model that will play a claimant on demand and never need a room booked or a form signed. What it can't do is surprise you with something true, because it can only return a blend of what's already written down. Research earns its keep on the day a real person says the thing no one thought to ask.
Content design is responsible for the language: the words a worried person can act on, that say the true thing plainly and don't make them feel small for asking. The machine has mastered producing sentences. It hasn't touched the harder reading: what a sentence does to the person who has to act on it. A fluent line that's subtly wrong about what someone is owed travels further than a clumsy one before anyone catches it.
Service design is responsible for the whole path a person takes: every channel and handover they cross to get what they came for, over days or months. AI tends to arrive bolted onto a single step, a clever box on one screen. Service design asks whether the step needed to exist at all, and whether automating it only pushed the difficulty somewhere the person can no longer see. The machine optimises the part. Someone has to be answerable for the whole.
Interaction design is responsible for how the person and the system act together, turn by turn. That used to mean drawing screens. It increasingly means shaping agentic behaviour with fewer fixed interfaces: a system that generates its own responses and takes its own actions, and has to stay legible and answerable while it does. It's improvised, moment to moment, and someone has to stay answerable for whether it still feels like a thing that understood you.
None of the four is more the loop than the others. Drop one and the loop gets a blind spot, and a blind spot is exactly what the machine covers over with something that looks handled. One practitioner doing all four is welcome. The danger is doing none of them on purpose, fooled by a loop the machine made look done. That's how the vantage goes.
The loop under pressure
The machine's loop is filling fast, and it's filling well. That's not a complaint. A lot of what used to clog our days was making loop work that didn't need a human: the first draft, the tidy-up, the hundredth variation, the summary of the summary. Handing that over is mostly good. The trouble is what happens to the human loop while the machine's loop races ahead.
It gets thinned. When the machine can produce a plausible answer in seconds, the pull is to skip the slow part, the checking against real people, because the answer already looks done. The human loop still turns, but with less and less actually flowing through it.
It gets sidelined. The making becomes so quick and so cheap that the schedule fills with making, and the time that used to go on understanding how things land gets squeezed to the end, where it becomes a tick-box if it survives at all. Nobody decides to cut it. It just keeps losing the argument to whatever can be shipped today.
And it gets faked, which is the worst of the three, because it looks like the real thing. You can stand a single agreeable person next to the machine, or a synthetic stand-in, and call that keeping a human in the loop. The loop turns, the box is ticked, and the people it was meant to serve are nowhere near it. That's a loop of one, closed against the very people it claims to represent. It earns its own chapter later; one beat here is enough.
None of this is new in kind. Lisanne Bainbridge set the pattern out in 1983, the ironies of automation: hand the machine the easy parts and the person is left the hard residue, then asked to stay sharp enough to catch what the machine can't, on ever less practice. What's new is only the speed, and that the loop now closes itself before anyone thinks to look.
This is what the pressure does by default, to good practitioners, when one loop can suddenly outrun the other. Nobody had to fail for it to happen.
What one real conversation can do
On a piece of assurance work, on an AI‑enabled service in a regulated public setting, a team had a model of who the users were and what they'd struggle with. It was a good model. It was built on real data, sharpened by sensible people, and the machine could elaborate it endlessly: personas, edge cases, sample dialogues, whole imagined transcripts of people using the thing.
By every indoor measure, the human loop looked well served. We knew our users. We had pages on them.
Then we sat with one group of actual users, the kind the service would lean on hardest, and within an hour the riskiest assumption in the whole build came apart. The thing they showed us was the kind of thing that doesn't get written down, and so was never in the data the model learned from, and so could never come back out of it.
The machine had filled that gap with material that read as true and wasn't. It took a real person, in the room, telling us something we hadn't thought to ask.
That's the human loop doing the one thing the machine's loop can't: meeting a real person and coming back changed.
The machine can generate endless material about the user. It cannot be surprised by one.
The lesson I took from it is small and practical. When the machine has made the human loop look well served, that is the moment to go and check it against one real person. The fuller the file looks, the more that check is worth.
Close
Two loops, then. One the machine's, one the person's, kept turning in time by people whose whole value is that they stand outside the machine and feel what it makes. That's the model. It won't get more complicated, but the years ahead look likely to lean on it hard.
The chapters ahead take the pressure one piece at a time. What we stop noticing once the machine drafts everything. What judgement is for when the answer always arrives first. Whose reading of the user is allowed to count. What the work itself becomes, and who we become while doing it. All of it sits on this one joint: two loops, close, never merged.