Applying AI to Architecture's Unanswered Questions

In this Post:

• What Are Historical LLMs?

• Why This Matters: The Hindsight Problem

• The Questions Architecture Can’t Ask (Yet)

• What Does ‘Architect’ Mean, Anyway?

• Asking the Past About Its Futures

• What This Means for Practice

I haven’t published many posts lately, because I’ve been rolling much of my recent writing into larger projects, hopefully to be shared soon. As I’ve been working on these, I’ve tried to ignore the cool shiny things that float across my newsfeed, but this one was too good to pass up. It got me genuinely excited about new, previously unimaginable applications of AI:

A team at the University of Zurich trained a family of AI models (the “Ranke-4B” family) exclusively on texts published before specific dates - 1913, 1929, 1933, 1939, 1946. Ask their 1913 model about Adolf Hitler, and it confuses him with an obscure philosophy professor from Giessen. Query it about the “gravest dangers to peace,” and it warns about Balkan tensions and Austro-German ambitions, exactly as an educated European reader might have in early 1913.

According to the project documentation, these aren’t parlor tricks. They’re what the researchers call “aggregate witnesses to the textual culture of their era” - compressed representations of massive historical corpora that let us query the past in ways that weren’t previously possible.

The architectural implications aren’t immediately obvious. But they’re there, and they’re worth exploring. Imagine:

Or, at least that’s how we imagined it must have seemed. We can reconstruct almost everything about Woolworth that institutions love to preserve - design intent, structural bravado, financial mythology, the “firsts,” the press kit version of history.

What we can’t reconstruct, at scale, is the first-contact experience: what ordinary observers thought they were looking at, what metaphors they reached for, what bothered them, what thrilled them, what they took for granted without even noticing. Now, potentially, with Historical Large Language Models (HLLMs), we can.

Large Language Models like ChatGPT or Claude are trained on vast amounts of the text we have now - internet pages, books, articles, social media posts. These corpora of texts necessarily subsume everything that has come before, back through history. Through them, LLMs learn patterns in how language works, what concepts connect to what, what arguments people make about what topics. The process is inherently biased to the present and recent past, because that’s when most of our written material has been produced.

To eliminate such bias, Historical LLMs (HLLMs) apply LLM training methods to time-locked training data. The Ranke-4B family of models was trained from scratch on tens of millions of pages of historical text with strict cutoff dates. The 1913 model sees only texts published through 1913. The 1929 model stops at 1929. And so on.

This creates something genuinely different from asking GPT-5 to “pretend you’re from 1913.” Modern LLMs know how the story ends. They know about World War I, the Great Depression, World War II, the Cold War, climate change, the internet. That knowledge shapes every response, even when explicitly instructed to ignore it. You can’t truly believe the sun revolves around Earth once you understand heliocentrism. You can’t read 1913 anxieties about the Balkans with genuine uncertainty once you know what August 1914 brought.

HLLMs don’t pretend ignorance. They embody it. World War I literally doesn’t exist in Ranke-4B-1913’s textual universe. When it discusses European peace prospects, it’s genuinely uncertain in ways that post-1918 observers - human or AI - cannot be.

We design buildings that must function for fifty, seventy, one hundred years. Every design decision encodes assumptions. About climate, about materials, about how people will move through space. About what problems matter. About what solutions are possible.

Those assumptions have a date stamp, even when we can’t see it.

The modernists assumed abundant cheap energy and universal automobile ownership. Postmodernists assumed stable climate and gradual change. Contemporary sustainable design assumes various futures - some optimistic about technology, others pessimistic about human behavior.

But we’re all designing from ‘after’ climate change became undeniable, after 9/11, after 2008, after COVID. Those events shape what we think is possible, what we value, and that encodes our designs.

This is the hindsight problem. Once you know how it turned out, you can’t think from inside the uncertainty anymore. You can try - scenario planning, stress-testing assumptions, exploring alternative futures - but you’re always contaminated by knowledge of what actually happened.

HLLMs offer something different: a way to see how things looked from within a moment of genuine uncertainty, before hindsight simplified the picture into obvious narratives.

The 1913 model doesn’t know about antibiotics, nuclear weapons, computers, or spaceflight. It can’t conceive of climate change or extinction risks. Its technological optimism is uncomplicated by Hiroshima or Bhopal or Chernobyl.

That limitation is also its value. It shows us what people could see from where they stood, with the conceptual vocabulary and empirical knowledge they actually had.

Architecture has plenty of history. We know who designed what, when, and often why - at least according to architects’ manifestos and critics’ interpretations.

What we lack is comprehensive reception history. The archival record is sparse and biased toward those who were holding microphones: critics’ reviews, architects’ explanations, institutional documentation. We have remarkably little texture about day-to-day public perception, casual language, mundane judgments.

What did educated Americans in 1913 think about city design? About suburban development? About housing reform? About the relationship between built form and social outcomes?

We can read period texts, of course. But HLLMs let us query those texts in ways that weren’t previously possible - asking questions the original authors never considered, exploring conceptual connections they never made explicit, testing hypotheses about implicit assumptions.

According to the research team’s documentation, you could theoretically ask an HLLM to simulate reactions to a specific building from different social positions, probing what aspects people noticed or

Share

Here’s something that got me really excited: using HLLMs to settle something we’re always arguing about: the cultural valence of the word ‘architect’, and how its changed over time.

Language is a dataset of reputational weather. You can track not just how often “architect” appears in texts, but what it appears ‘near’. Architect + genius. Architect + expensive. Architect + irrelevant. Architect + visionary. Architect + bureaucrat. Architect + luxury.

The proximity patterns reveal something about public perception that’s different from counting positive versus negative reviews. They show what associations were available, what metaphors people reached for, what cultural position the architecture profession occupied in different eras.

The profession constantly debates its “public value.” But public value is partly a story problem - and stories leave linguistic fingerprints across decades of text.

This wouldn’t tell you what to do Monday morning. But it would give you an unusually powerful mirror: what the culture thinks you’re for, and how that’s changed.

There’s another capability that feels almost like science fiction: you could use HLLMs to ask past eras to imagine their futures - meaning, our present.

The researchers suggest you could simulate a conversation with a 1950s voice and ask: What future do you expect? What future do you fear? What would you want cities to become?

The point isn’t prediction accuracy. People in 1950 probably weren’t any better at forecasting 2025 than we are at forecasting 2100. The point is recovering the ‘horizon of expectation’ - the futures people could see from where they stood, with the knowledge and conceptual tools they had.

We remember the eerily accurate predictions and the hilariously wrong ones - flying cars, videophones, nuclear everything. We forget the ordinary expectations, the quiet assumptions, the default trajectories people considered inevitable.

For architecture, this matters because we’re constantly making implicit bets about how long our assumptions will hold. Contemporary sustainable design assumes certain things about future technology, future behavior, future values. Some of those assumptions will prove correct. Many won’t.

Understanding how previous generations thought about their own uncertain futures might teach us something about epistemic humility. Not because they were wiser, but because we can see their blind spots in ways they couldn’t.

Historical LLMs reveal how smart, educated, thoughtful people could fail to see what was coming, not because they were stupid, but because they were human.

That’s useful. It should make us more cautious about our own predictions. More willing to imagine we’re wrong. More prepared for surprises.

Are there any practical takeaways? Yes, a few:

1. Be suspicious of confident predictions about AI’s impact on architecture (yes, including mine). The people making confident predictions in 1913 about urban futures were mostly wrong. We’re not smarter than they were - we just have the benefit of hindsight about their moment, not ours.

2. Focus on adaptability over optimization. Buildings designed for one specific future fail when that future doesn’t arrive. Buildings designed with multiple possible futures in mind - with flexibility, convertibility, expandability - tend to age better.
   This principle applies to professional practice itself, as well. The AI capabilities available in 2026 won’t be the ones available in 2030 or 2035. Designing practice around today’s tool limitations means rebuilding when those limits shift. Designing for adaptability means preparing for capabilities you can’t fully specify – or maybe even imagine – just yet.

What will seem obvious to future observers that we can’t currently see? What are we taking for granted that will seem absurd in hindsight?

Historical LLMs are a reminder that this uncertainty is a fundamental feature of human experience and architectural practice, not a problem to be solved.

HLLMs are being developed primarily for behavioral science research, not architectural applications. But the concepts they embody - time-locked perspectives, epistemic humility, awareness of hindsight contamination - are directly relevant to designing human-centered environments in an AI-native world.

The researchers building these models understand something architects already know: the past isn’t just context. It’s a laboratory for testing assumptions and practicing humility about our own limited perspective.

Architecture has unanswered questions not because we lack diligence, but because some questions required capabilities we didn’t have. How do you interview thousands of historical observers simultaneously? How do you track subtle linguistic drift across a century of texts? How do you simulate conversations with perspectives that no longer exist?

You couldn’t. Now, potentially, you can.