Social AI Expedition coffee 2 · Utrecht
Two facilitators, one microphone, and the question of what AI does to the human work.
Joost and Lars were sitting in a cabin at the Huis van Actief Burgerschap in Utrecht for a conversation about Social AI. Before they recorded, they had already done a quick check-in. Joost switched Dembrane on. Lars's first question once the device was running: why not record the check-in too? What followed are six places where Social AI shows itself most sharply: a check-in, a rupture, a mirror, a paradox, a campfire and an echo.
"Why don't we record the check-in?"
It's the first question Lars asks the moment the recording starts. They had just checked in without a microphone. Lars notes: shouldn't that actually have been on tape too? And without anyone noticing, that question opens everything that comes after. What shifts when the tool is listening in from moment one?
The check-in is the moment where you put on the table what you want to experience. In the past you carried that intention yourself. At the end you'd look at whether you had experienced it, and if not, there was apparently a good reason for that. You were the owner of what you had wanted.
But when you know it's being recorded, then you can, Lars says, "switch on the big rambler and just chatter away": let go, just talk, see what comes out. It's not a joke about the device, it's an inner stance. And that's exactly where the first shift sits, the one they keep circling for the whole hour.
Two places where the intention can land
Ownership is what you carry yourself. The presence of a recording offers to take over that carrying, and with it shifts, often unnoticed, who is still responsible for the experience.
A session a while back.
Joost tells about an earlier session. A large gathering. Dembrane in every conversation. Forty people, four breakouts. At the end the facilitator asks for a report-back. Someone gets angry.
"You introduced the AI as the one that would take the minutes. So we didn't appoint a chair. Didn't appoint a minute-taker. And now you're suddenly changing the way of working like this."
It's a small incident, but it shows something big. The procedure these people have known their whole professional lives, appoint someone, keep time, give a summary, had fallen away without anyone saying so. The AI would handle it.
What "we're recording" can all mean
Every announcement (a microphone, a tool, a procedure) opens its own cloud of meaning in each head. A good facilitator hears those clouds even before they're spoken. The AI doesn't hear them.
Turning the intuitive into instructions.
Joost is working on something he provisionally calls the twelve principles of Social AI. He notices that the exercise of explaining to a machine what a good facilitator does forces him to articulate that craft more precisely than was ever needed in front of a person.
How do you teach a Large Language Model that not every tension needs to be resolved? That "there is no agreement" is also a valid outcome? That an individual shouldn't just be put in the spotlight? That an AI can't damage a bond of trust, because it doesn't have one, but that such a bond does very much exist between the people taking part?
From intuition to instruction
Every prompt is an attempt to articulate craft knowledge. Lars: "You're really writing a manual on good facilitation, and you hand it to a tool." That is, among other things, also a form of passing on the craft.
A concept that came out of a series of steering-group sessions.
Joost reads something out loud. The concept emerged from a series of AI interactions on transcripts of the steering-group meetings in Doesburg. It has been given a name.
The bottom-up paradox
The felt tension that every form of steering, including the steering to work bottom-up, is by definition top-down.
The more structure you bring in as a facilitator to organize participation, the more the group tends to fill in that structure as if it were an assignment.
The paradox is inherent and unsolvable. You can only navigate it consciously.
Lars recognizes it immediately. "Every bottom-up process starts with someone making a top-down decision that something bottom-up is going to happen." And then a second paradox arises between them. One they had run into before, in a two-day session with Floor: does change begin with someone who holds a strong opinion, or does it begin with someone who listens?
Where does change begin?
Neither of them solves the paradox. They let it stand, and walk around it together for a while. This is perhaps Social AI at its best: a machine-distilled concept that comes out of the conversation stronger, through two people, than it went in.
Do you program the program with the humanity?
From under the tabletop comes the real concern. Lars names it: the AI is good at directedness. Joost puts it in one word: efficiency. And then what?
Lars calls it his fear of unmasking: the fear that a machine will give back, afterwards, how much irrationality, repetition and bullshit there is in your own hour. Joost finds it scary too. But underneath that fear lies something bigger, and it's Lars who brings it to the surface.
Work isn't only about earning your bread. It's nice to sit around a campfire and talk with each other. It's nice to throw a log on the fire and hold a bite for the other person. If there's an abundance of apple trees around it, work is mostly a social activity that we, as social beings, find pleasant.
Task-directedness vs. relationship-directedness
The campfire isn't a metaphor for inefficiency. It's a metaphor for what work also is, besides being useful: being a social being. The question isn't whether AI makes us more productive. The question is whether it lets us keep the campfire.
"Do you feel like we're recording a podcast?"
At a certain point Joost looks up. He notices he's suddenly talking more to an audience than with Lars. The microphone has brought a third position into the room: not just me and you, but also the listener afterwards. Just before the end, Joost presses a button on Dembrane.
How Dembrane generates a closing question
The prompt is small, the question can be big. Joost notes that Gemini 2.5 is by now almost a year old, and that a newer model (Claude Opus 4.7) makes a world of difference in how well it can reflect back what was really said. The quality of the model is, at this moment, the quality of the echo.
Joost answers first: from the feeling of constantly having to prove himself. AI as his autonomous signboard, casting his opinion and vision into a workable format so others can do something with it. Lars answers at greater length. First pragmatically: from making reports and endless run-sheets. But then he hesitates, because in the craftsmanlike fiddling with it there's also quality. What it does do is deliver a half-finished product he can carry forward, no more blank page. And then, deeper:
Lars: "I experienced a free space. We wandered freely through a forest of philosophical and very practical questions. In the end, with the help of the question the tool offered up, we reached a beautiful synthesis."
Joost: "You and I have both become facilitators in our own way. And today that really touched a meeting point with the very latest technology. How the superhuman and the intuitive of the craft come back. Thoughtfulness, that's the word I keep returning to. That word alone shows what the human has to offer in this."
Was this helpful?