Most conversations about AI settle into one of two pictures. In the first, AI is an oracle: you ask, it answers, and the answer is supposed to be right. In the second, it is an autopilot: you hand over the task and come back when it is done.
Both pictures put AI in the wrong place, at least for the kind of work I find interesting. An oracle invites you to stop thinking. An autopilot invites you to stop looking. Neither is how good collaboration works between people, and I see little reason it should work differently with software.
The picture I find more useful is a collaborator: very fast, widely read, tireless, frequently wrong, and at its best when you argue with it.
Cheap drafts change the economics of ideas
The most obvious change is cost. A first draft, a rough prototype, ten different approaches to the same problem: things that used to take an afternoon now take minutes. When trying an idea is nearly free, you try more of them, including the odd ones you would once have dismissed as not worth the effort.
The bottleneck does not disappear, though. It moves. If producing options is cheap, choosing between them becomes the scarce skill. Judgement, taste and a clear sense of what the work is for matter more than before, not less. That is as much a creative skill as an analytical one, and it is the part I would be most reluctant to hand over.
Fluency is not the same as truth
Language models are remarkably fluent, which is exactly what makes them easy to over-trust. A confident paragraph about the wrong product reads just as smoothly as a correct one.
This is the problem at the centre of Viska, my experiments with conversational commerce. The interesting question there was never whether a model can talk about products. It is whether it can say things that are true about one particular store’s products and policies, grounded in that store’s own knowledge through retrieval, and ideally honest when that knowledge runs out. The same rule applies on a smaller scale to Ask this site: answer from what is published here, link the sources, and say so when the answer isn’t there.
Fluency is cheap. Being right about something specific is expensive.
Designing conditions instead of outcomes
A less obvious creative use of AI is not asking it to produce something, but setting up conditions and watching what it does.
In Tingstad, the idea is not for a model to write the story of a city. It is to give 24 inhabitants their own backgrounds, values and ambitions, a shared world with workplaces, an economy and democratic decisions, and then let the history be whatever their choices make it. The author’s role shifts from writing plots to designing a world in which plots can happen. It feels less like using a tool and more like gardening.
The Living Atlas pushes the same idea in another direction: an autonomous research agent that follows its own questions across a growing map of knowledge. Whether an AI can be curious in any meaningful sense is an open question. Whether it can surface connections a person would not have thought to look for is a more practical one, and a more interesting one to build.
Both projects share a useful constraint. If the system is allowed to surprise you, you need a record of what actually happened, or the surprise is just noise. Tingstad’s stories are meant to be rooted in the city’s documented history; the Atlas is meant to show how its knowledge changed over time. Autonomy without a trail is hard to learn from.
What stays mine
A good human collaborator improves the work and leaves it recognisably yours. With AI, the first part gets easier every year. The second takes deliberate effort: deciding what the work is for, checking what comes back, and holding on to the parts I would be sorry to give away.
Often those turn out to be the questions rather than the answers. The answers can come from almost anywhere, as long as someone can check them.