Architects Are Using the Wrong AI Tools

Ask ChatGPT whether your four-storey building needs a second exit stair. You'll get an answer. It'll sound sure of itself. And there's a decent chance it's wrong for your city.

That's not because the model is bad. It's because it's reading from the wrong library.

What general AI is actually good at

Let me be fair first. Gemini, ChatGPT and Claude are great at visualization. You sketch a massing on trace, describe the mood, and thirty seconds later you have an image that's more detailed than most sketches. It's quick, it's iterative, and it doesn't need to be precise to be useful. A render can be a little wrong and still do its job, which is getting people excited about an idea.

Garry Miley wrote a sharp piece in AEC Magazine this year about where AI is landing in architecture. His read: adoption is clustering at the two ends of the workflow. On one end, massing and feasibility tools that crunch zoning, FAR and budgets. On the other, generative images. The "missing middle," where a design actually becomes architecture, is still mostly untouched.

I'd go one step further. The middle isn't missing because AI can't reason about buildings. It's missing because we keep handing it the wrong books.

Here's where it falls apart

Ask a general model a question that has a correct answer and it starts to collapse. Code compliance. Buildability. Local context. The stuff that decides whether a project gets a permit or gets sent back for a third round of comments.

Think about where the answer comes from. A general model learned from the internet: blog posts, forum threads, marketing pages, someone's half-remembered take on egress from 2014. That's the library it's searching. So when you ask about stair requirements in Montreal, you're getting an average of everything anyone has ever written about stairs, anywhere.

It's like asking a librarian for the building code and getting handed a stack of magazines.

The books you actually need

To get the right answers, you need the right books on the shelf. For any real project, that means:

• The national building code
• The provincial or state building code
• Fire and life safety regulations
• The accessibility code
• The zoning bylaw, and the data for the actual lot
• A site report
• Any heritage guidelines that apply

None of this is exotic. Every architect already works from these documents. The question is whether the AI is working from them too.

If they aren't loaded, AI just agrees with you

This is the part that worries me most. A model without the source documents doesn't push back. It won't tell you that setback doesn't work on this lot. It takes your assumption, wraps it in confident language and hands it right back.

I spent years as a computational designer and BIM specialist. The tools worth anything were the ones that told you no, early, before anyone had spent money on drawings. A clash detection that never finds a clash isn't a feature. It's a liability.

So use the general tools for what they're good at: mood, early massing, client conversations, first sketches. But the minute a question has a right answer, check what's on the shelf. If the code, the bylaw and the site aren't loaded, you're not getting an answer. You're getting an echo.

Right books, right answers.