A few weeks ago, in one of those happy convergences of life, I was reading an Italian essay on Italo Calvino and artificial intelligence (I wrote about it here), and there’s a very interesting set of passages that have to do with our contemporary obsession for answers and, in contrast, the importance of being able to formulate the right questions. As Turing observed, even if his famous test if often misinterpreted, the turning point happens when a machine is able to give us believable answers, and we’ve been obsessing over answers since Google. We’ve been asking questions, taking for granted that the answer could come from an algorithm that draws from our own data.
“La domanda è la chiave di apprendimento del mondo e del sé.”
(Il Visconte Cibernetico)
This resonated with me and with a conversation we had with my team a while back on how many of the applications around Artificial Intelligence, at least in the design and engineering segment of the construction industry, seem to be concentrated on asking questions to your model.
Plug a chatbot into your information modelling tool of choice, ask it questions, get answers, ship the engineering package. Query resolved. Case closed. But there’s a caveat, and it’s as big as that bug Claude encoded in your Dynamo script.
The society of the query
What we’ve inherited from Google, and the post-9/11 days of the internet, is a habit of mind: we ask, and we assume the answer is already sitting somewhere in a dataset, waiting to be retrieved. LLMs haven’t invented this habit: they capitalised upon the urge of getting quick answers, while fulfilling Turing’s old promise that a machine could answer believably. Believably is the operative word. It doesn’t mean correctly. It means the answer sounds like something a competent person would say, which is a much lower bar than we tend to realise.
Bertrand Russell sharpens the point: we still read the Greek philosophers not because their answers were right — most of them, scientifically, were a bunch of misinformed bullshit — but because their questions were on the spot. Plato left us a method for finding out we didn’t know what we thought we knew. That’s a very different inheritance.
Now bring that back to a design office. AEC has spent the last decade building what’s possibly the single most impressive data-hungry apparatus outside of finance: parametric models, clash detection, 4D scheduling, digital twins that are supposed to know everything about a building before it’s built. The BIM promise was never really about drawings, in this sense: it was about having a queryable, trustworthy source of truth. And now every software vendor at every trade show is racing to bolt a chat window onto that source of truth, so you can turn your model into an oracle and ask it whether you’ll be back from the construction site war.
“Ibis redibis non morieris in bello”
Convenient? Absolutely.
Revolutionary? Let’s not get carried away.
Two models, one blind spot
Here’s my problem with the current wave of people asking their models stuff, and it has nothing to do with the AI being unimpressive. After a few faltering steps of the first models that were rushed into the market, it’s genuinely good at answering. My problem is that almost none of the energy in the room goes into asking whether the question — or the model being questioned — deserves the confidence we’re about to place in it.
An Information Model is not a building, as much as we’d like it to be just as accurate as the reality. It is a virtualisation assembled by dozens of clumsy hands, on a timeline, under commercial pressure, with information targets that were negotiated and then rarely measured, and with more silent assumptions baked into it than most contracts admit. Ask it a question, and it will answer — an AI layer on top will make that answer sound articulate, contextual, even wise — but the authority of the answer borrows entirely from the quality of the model underneath. If the model is wrong, incomplete, or simply out of date because someone forgot to push a sync three weeks ago, the chatbot won’t tell you that. It will tell you, fluently, the wrong linear metres of skirting board.
This is Turing’s prophecy landing exactly on our desks: a machine that answers credibly is not the same thing as a machine that answers correctly, and the gap between those two doesn’t close on its own. It closes — if it closes at all — because a human questioned the source with the right doubts first, and asked the right question later. The first question, therefore, is as simple as rain: is this model good enough to be asked anything at all?
That is not a question most current AI-in-BIM tool is particularly interested in helping you ask, because it’s bad for the demo.
Question everything, and question it well
In our field, asking well means something specific and, frankly, unglamorous: it means knowing enough about model authoring, information requirements, and the actual provenance of your data to know when a fluent answer should make you suspicious rather than relieved. It means training junior staff not on how to phrase a prompt, but on how a model gets built, where its gaps typically hide, and why “the software said so” was never an acceptable sentence, not before AI, and certainly not after. We need to train people, and ourselves, to go hunting for those gaps, for those errors, for those things that might have gone wrong.
Basically, if we want to use AI to cut on manual labour (which might be fine by me), we all need to be little BIM Coordinators at heart.
This also means being honest, as an industry, about what we’re actually automating. We are not automating judgment. We are automating retrieval and articulation. Those are useful things to automate. They are not the same thing as competence, and no amount of interface polish will make them so.
The question behind the question
I don’t think the answer here is to distrust our models, digital or otherwise: that would be the intellectual equivalent of throwing out the drawings because one dimension was wrong (or, as we say in Italian, tossing the baby away with the dirty water after their bath). The answer is to stay in the business of asking questions about the model before we start asking questions to the model.
No Comments