Social AI Expedition coffee 4 · Den Bosch

"Is it okay to record this?"

It starts, like every time, with a check-in. How do you feel, how did you prepare, and what do you want to experience in this hour. Only when everyone has said "I'm here" does the conversation really begin. Lars goes first: he ran out the door before his children this morning, without having to hurry them to school for once, and feels curious and grateful. Sameer says his mornings are filled with work for the startup, and that he's curious about the conversations that came before this one. Marijn biked to the station and is still half with a client meeting later that afternoon. Joost notices this moment is closer to home than the previous three, and that he invited the people he wanted to be here. This, he says, is him showing up.

Quite soon a pattern emerges: the four of them stand at different distances from the technology. Lars places himself furthest away. "As my children tell me, I'm a boomer," he says, and yet he sits in the middle of these conversations. His work, he says, is bridging: translating what is technically possible into the ordinary, everyday dynamics people work in. Marijn, with his design and software background, knows what's possible. Sameer builds the tool itself. And Joost sits somewhere in between, building daily, but with his attention on the social side.

It's exactly from that bridging position that Lars points at a small, decisive detail. How you introduce something determines whether people join. "If you ask in a meeting, do you consent to recording, you go in a completely different street with people who are not used to AI." You get a discussion, and it almost always ends with no. Not because of the technology, but because of the question.

"How do we understand each other?"

When the question comes up what drives him, Sameer tells where he comes from. Twenty-four years old, from Hyderabad, landed at Dembrane knowing nothing about this field. He describes a rainstorm in his hometown where the water rose hip-high, a basic infrastructure problem almost nobody talks about. A country of 1.4 billion people, and the question of how you solve education, healthcare and the most basic needs. That's where his drive comes from.

The bridge he builds to AI is surprisingly human. The biggest challenge we face as a species, he says, is understanding each other. I think something, but I can't share it well, or the other person can't quite hear it. That's where the core problem starts. And a tool that can help people in a personal conversation understand each other's perspective better is, for him, exactly where the value sits.

Someone at the table adds an image from The Hitchhiker's Guide to the Galaxy: the Babel fish, a little fish you put in your ear, after which you understand every language in the galaxy. A playful metaphor, and immediately its own limit. In India alone there are thirty-six languages and many more dialects. The question that follows is as sober as it is big: who's going to pay for this, and who has the imagination to believe something like this can work in a democracy of that scale?

"There's an emotional value attached."

Then Marijn deliberately plays the skeptic, recalling a conversation from more than a year ago about whether an AI might facilitate better than the facilitator. Maybe it can. But there's a catch: it's very easy to wave away anything an AI says. The tool can give the best answer, better than any of us, and you still just shrug.

Why? When you say something to me, there's an emotional value attached. That makes me put in the effort to listen, and I won't cut you off halfway. With an AI that weight is missing; the best idea is dismissed as easily as the worst. And there's a second observation that belongs to it: an AI rarely pushes back. A human can say, you know, I'm not sure that's the right way to look at this, let's try another angle. An AI takes your starting point all too willingly and dutifully builds on it.

Lars recognizes the pattern from his own profession: the relationship between management and employees. Management makes an analysis and a plan, and the plan is usually right. Then the communicating begins, the selling of it, and the employees wait for the coffee to arrive. In the restroom someone says "same shit all over again", and everyone waits for it to blow over. Not because the analysis is wrong, but because the meaning is missing that makes you accept something from someone else. So his question isn't whether AI can give the right answer, but: how do we add meaning to what some other source is giving us?

"It's just trying to be as helpful as it can."

How come these models are so helpful, sometimes almost fawning? Sameer explains it from his craft. Under the hood it's next token prediction: the model keeps predicting the next word. During training, people said which version they liked better, and everything kept stacking on top of that agreement. So at the end of the day, it's mostly trying to be as helpful as it possibly can.

Helpfulness is good for the masses, but not automatically good for work. Around the table the observation lands that an overly agreeable AI becomes less useful for exactly that reason, and that a critical countervoice is needed. How well you can judge something also depends on what you already know. Ask an AI to build a website and as an expert you'll find thirty mistakes; ask about something you know nothing about, and you'll swallow it whole.

There's a surprising upside, and it enters the conversation through a kitchen stove. Lars tells about his wife, who used an AI to find a new one. First she got the five best options, then, asking further, an equally long list of reasons not to buy them. I don't know what to buy anymore, she said. So the old-fashioned work started again: searching, reading and checking for yourself. And that's how she became the stove expert, and they bought one they're happy with. The critical eye you develop by using AI, it's said at the table, trickles down into the rest of your life, even without AI. A kind of media literacy, learned anew.

The conversation then puts on a bigger lens. What if generative AI isn't the better search engine, but the new internet? Just like everything once got an internet connection, whether it needed one or not (someone's stove is connected to the internet, nobody knows why), everything is going to get AI. The more interesting question is what becomes possible then: interfaces that adapt to you instead of the other way around, a house that figures out for itself when power is cheapest, an answer that doesn't arrive as a wall of text but as exactly the overview you were looking for. And in the middle of those dreams falls the sentence that returns in the check-out: let it widen your consciousness, rather than take over your decisions. Even if it's just one percent.

"These are the AI's memories of me."

After Sameer leaves, the conversation continues in Dutch, and gets more personal. A recurring discomfort lands on the table, one everyone recognizes: an AI stores things about you in its memory, and sometimes builds entire worlds on top of facts that aren't true. One example from Marijn: a memory that stored his girlfriend's name wrong from an insurance document, and months later cheerfully brought that name back in a completely different conversation. A small mistake you spot easily; there are also ones you don't.

Joost puts a distinction underneath it that looks small and is big. The memories an AI stores are not his; they are the AI's memories about him. His own reflex to keep that sharp: have the system note, when it stores something, that it's an interpretation, from the AI's perspective, and not Joost's truth. Then later, when such a memory becomes relevant again, it can be used with some thoughtfulness instead of taken at face value.

That same morning he had a large model, the freshly released Fable 5, take a look at his entire personal AI infrastructure, the collection of documents and routes his AI (he calls it Finn) gets handed every conversation. What the model said back struck him as remarkably insightful. You read this entire context almost like a collection of scars. Something happened, and now you're trying to prevent it from happening again. With the result that Finn is mostly busy not doing things wrong, instead of doing things well. It touched what Sameer had said earlier: we're trying to mimic our own consciousness, and then it matters whether you say "don't do this" or "think in this direction".

"Are we secretly afraid the technology mirrors us better than we dare to look at ourselves?"

A recurring thread through the whole conversation is the value of impartiality. People tell Lars it's nice to have a neutral facilitator, and he always answers: that's not me. He does his best to be less partial, but he always brings himself along. The interesting thing is that an AI sometimes is experienced as neutral, and that this opens something. Someone stuck in their own convictions pushes a human objection away instantly, an emotional reflex, but will sometimes let a different perspective from a machine sit on the scale for a moment.

Joost tells how he once tested that. On a winter sports trip with his father and two of his father's friends, one of them leaned towards firmly right-wing and conspiracy-shaped ideas, among them that covid was a planned virus. Joost put that same claim to Claude in an incognito window, and the model calmly built a substantiated rebuttal: we hear this more often, but by now there's demonstrable evidence it doesn't hold, look here and here. He names the limitation himself: he didn't read the rebuttal out loud (too long), he only said he'd checked it, and the man never read it himself. And it wasn't facilitator Joost thinking, let me show this man a fairy tale. It was just Joost, with a genuine question. Two different versions, he adds himself.

The striking thing is: it's the Echo itself that asks this question most persistently. Already at Sameer's departure, halfway through the morning, the first Echo closed with the question whether the technology might mirror us better than we dare to look at ourselves. At the end they press the button once more, and again the system puts its finger on the same spot: do we still dare to share our vulnerable doubts when we don't fully control the outcome? In the check-out Lars looks back on that ever-returning mirror question, and lands on the distinction that lingers: interesting how the system keeps surfacing the fear of the mirror. "I don't feel fear. I feel desire, if anything."

"AI is a magnifying glass for what makes us human."

Just before the end, they check out the way they began: by naming what they take with them. Sameer had already done so halfway, before he had to leave. His check-out was a call to keep having these conversations as a community, "so we don't become this AI accelerationism mess" where everything has to be AI. Marijn takes home a question he deliberately leaves open: in a world of giants, what is the role and the responsibility you can genuinely claim? Not training your own model, that's not worth starting. But choosing your own frames, and getting as much done as possible within them.

Lars takes home an intention that sums up the whole conversation: dealing more consciously with his own history, letting the moderate middle be heard out loud because systems amplify whatever you feed them, and letting the mirror be a mirror. Even one percent more consciousness, we're always better off with that.

For Joost an image returns from a little over a year ago, when he gave a presentation about AI to healthcare workers. His theory then: AI is a magnifying glass for what makes us human. The image had faded a bit, he says, but today's conversation confirms it once again.

And then he adds the shadow side that makes the image more honest. A magnifying glass doesn't only magnify the good. It magnifies our intention and our awareness, and at the same time our urge to create and our urge to be of value. The more you work with AI, the faster you get over the hill of novelty, hit a peak, and land in a valley where the shine comes off. The Valley of Despair, except this time it moves, because the technology itself keeps developing so fast that a new gap keeps opening up. That costs real energy. And still, the hour ends, it was mostly fun: not really a check-out, more of a check-through.