Adoption is racing ahead of accuracy. The data shows where the failures originate — and a structured methodology shows how to close the gap.
If you have spent any time on architecture social media in the past two years, you already know what an AI-generated building looks like before you have even seen one. The proportions are exquisite. The lighting is impossible. The cantilever defies gravity in a way no structural engineer would sign off on. And somewhere in the back of your mind, a quiet voice is asking: but does this thing actually work?
The voice is right. Image generation has moved from novelty to daily use across the profession, and according to the AIA’s 2024 Firm Survey, AI is now embedded in day-to-day workflows at sixty-one percent of large architecture firms. Adoption has been swift. The harder problem has been quieter: the renderings are beautiful, and increasingly unaccountable to the discipline that produced renderings in the first place. The gap between what AI can output and what architecture actually requires has been growing, and most of the conversation has not caught up to it yet.
Where It Started
I started thinking about this gap before there was a vocabulary for it. In May of 2022, well before generative AI became the phrase in every consultant’s pitch deck, I was developing the concept for 533 Kirkham, a mixed-use multi-family residential project, and I was already using the AI image tools available at the time. The project worked. The visualizations did not read as composites or accidents; they read as architecture. But the reason they worked had nothing to do with the tools.
What I had developed, almost without naming it, was a discipline of upstream reasoning: a framework for deciding what had to be true about a project before any image was generated. The relationship between architecture and landscape. The cultural intent. The hierarchy of program. The order of spatial sequence. The logic of materials in their environment. The image came last; the thinking came first. I would later formalize this as the Pre-Generative Reasoning Framework. On Kirkham, it was simply the way I worked, and it was the only reason the AI outputs landed.
Quiet Adoption
The framework did not stay private long. Over the following years, it was integrated into the day-to-day visualization workflows at Solomon Cordwell Buenz, Jensen Architects, and DJL Architecture — three firms whose project portfolios span campus and university buildings, multi-family residential developments, data centers, and other large-scale work. Some of those projects were public; many remained confidential. What they shared was the same upstream discipline: image generation governed by architectural reasoning rather than the other way around.
What surprised me was how little selling the framework needed. Architects who had been frustrated by AI’s inconsistency recognized the problem the moment they saw the methodology articulated. The order of thinking was familiar; the profession has been doing it for centuries. What was new was the requirement to make the reasoning explicit enough that it could survive being handed to a model.
The Gap, Mapped
Outside the firms where the framework had taken hold, the rest of the profession was running the experiment in reverse — and the data was starting to come back.
The anatomy of AI architectural visualization failure, drawn from AIA, McKinsey, Stanford HAI and Autodesk surveys alongside peer-reviewed studies of diffusion-model behavior. Failures cluster across seven categories; almost all originate upstream in Phase 01 and surface only in Phase 02
The headline numbers from the AIA’s 2024 Firm Survey describe a profession that has flipped on AI faster than almost anyone predicted. Roughly one in three firms of every size now uses AI in day-to-day work. The figure climbs to forty-two percent at midsize firms and sixty-one percent at large ones. McKinsey’s research on the architecture, engineering, and construction sector adds the global frame: more than forty percent of leading firms worldwide now incorporate AI into their design processes, and generative tools have already cut preliminary design time by thirty to fifty percent. Autodesk’s 2024 State of Design and Make survey, drawing on more than five thousand industry leaders across eighteen countries, found that seventy-six percent of professionals now trust AI for their industry. Trust, in other words, has outrun verification.
That is where the second half of the data starts to bite. Stanford HAI’s AI Index Report tracks hallucination rates across leading models and finds that even the strongest exhibit fifteen to twenty percent error on domain-specific tasks — a band that climbs to thirty-five to fifty five percent when the topic moves into specialized or niche territory, which is exactly where most architectural prompts live. Peer-reviewed work on diffusion models documents the specific failure shapes. Pix2Pix and HouseDiffusion architectures generate floor plans with incorrect room counts, broken connectivity, and unconstrained boundaries. Newer studies on locality bias confirm what every working visualization team has noticed in practice: the model can render one wing of a building beautifully, then forget how it connects to the other wing entirely.
The cultural picture is no better. A 2024 paper in Nature’s Humanities and Social Sciences Communications, examining architectural representation across ChatGPT, Midjourney, and Google Maps, found that generative tools systematically default to Western frames. A separate MDPI study on Islamic architectural heritage in Midjourney noted persistent orientalism and reduced fidelity for non-English prompts. The implication is the same regardless of which tool is in front of the architect: the model is not neutral. Without upstream constraints, it slides toward whatever cultural shorthand its training data privileges.
Put together, the picture is unambiguous. Adoption is climbing; reliability is not. The failures cluster predictably across seven categories — structural implausibility, material-context mismatch, stylistic incoherence, cultural bias, program logic absence, scale drift, and long range spatial coherence loss — and almost all of them are seeded upstream in how the project is briefed and prompted, even though they only become visible downstream in the generated image. The diagram above maps the anatomy of that gap.
Inside the Framework
The Pre-Generative Reasoning Framework is built to close that gap by changing the order of operations.
Before any prompt is written, the framework requires the architect to declare a polarity — the specific relationship between the building and its environment. There are four: architecture as a subordinate insertion within nature, nature internalized as a spatial agent within architecture, architecture behaving as an ecological system, or nature itself functioning as architecture. The polarity decides hierarchy and agency. Everything downstream answers to it.
From the polarity, the architect introduces an anchor element — a single conceptual nucleus around which the project organizes itself — and constrains it through six classical design principles drawn from architectural pedagogy: balance, emphasis, movement, proportion, rhythm, and unity. From there comes the socio-cultural intent, the typological commitment, the spatial sequence, the material logic answerable to climate and terrain, the daylight treated as a narrative variable rather than a stylistic filter, and a cinematic perceptual plan that treats the camera as bodily navigation through space.
Only then is the image generated. And once it is, the generation itself is layered, never solved in a single prompt. Architecture is established first. Then textures. Then lighting. Then atmosphere. Then human entourage. Each layer is constrained by the layer above it. The model is not asked to invent the building; it is asked to render decisions that have already been made.
What Sets It Apart
This is the core inversion that sets the framework apart from how AI is most commonly used in architecture today. Most generic AI workflows let the prompt do the architectural thinking and treat the image as the answer. Plug-in libraries, model-specific LoRAs, and brand-name prompt formulas all work at the same downstream layer — they teach the architect how to talk to a particular model, not how to think before they call one. The Pre-Generative Reasoning Framework moves the discipline upstream of any model.
What makes it different in practice is what does not change when the tool changes. The polarity is the same whether the image runs on Midjourney, Stable Diffusion, or whatever comes next. The six design principles do not depend on prompt language; they come from the same pedagogy that has trained architects for generations. The framework’s specificity lives in architectural reasoning, not in any vendor’s release notes — which is why it has survived the model upgrades that made every prompt-engineering trick brittle, and why it ports across firms and project types without losing fidelity. The hallucinations do not vanish because the model has gotten smarter. They vanish because there is nothing left in the prompt for the model to hallucinate around.
The Outcomes That Followed
The same discipline, applied outward, produced its own kind of evidence. Work I shared publicly — visualizations built within the framework rather than around AI’s defaults — drew an audience past one million followers, attention from global galleries, and coverage in architecture publications including Illustrarch and ArchFeed. The kind of reach that, in this field, tends to follow work people find substantively new rather than work they find loud.
Around the same period, a project I developed within the methodology received an AI design recognition from Arch Hive back in 2023, and I later returned to that program as a juror — closing a small loop that taught me the framework worked at both ends of the table. It produced work strong enough to be recognized; it also produced a vocabulary clear enough to recognize good work in others.
An Invitation, Not a Threat
There is a version of this conversation that frames AI as a threat to architectural integrity. I think it is the opposite.
AI is the most exciting thing to happen to architectural representation in a generation. It compresses what used to take weeks into hours. It makes ideas visible at the speed of thought. It removes the friction between imagination and image — and for architects who have spent their whole training learning to think before they draw, that is not a threat. It is an invitation.
The tools we waited a generation for have arrived. The only question that matters now is whether we will meet them with the discipline they deserve. Architecture has always demanded intent. AI just makes the absence of intent louder — and for those of us who love this work, that is the most exciting part of all.
SOURCES & REFERENCES
[1] American Institute of Architects, Firm Survey Report 2024 — AI adoption rates across more than 1,200 U.S. firms. https://www.aia.org/resource-center/aia-firm-survey-report-2024
[2] McKinsey & Company — AI in Architecture, Engineering, and Construction (2024); The Economic Potential of Generative AI: The Next Productivity Frontier.
[3] Stanford HAI — Artificial Intelligence Index Report 2024, Technical Performance chapter (hallucination rate ranges). https://hai.stanford.edu/ai-index/2024-ai-index-report
[4] Autodesk — State of Design and Make Report 2024 (n=5,399 industry leaders, 18 countries). https://www.autodesk.com/design-make/research/state-of-design-and-make
[5] Diffusion-model failure-mode and locality-bias literature — Pix2Pix and HouseDiffusion architectural layout analyses (PLOS One, ScienceDirect, arXiv, 2024-2025).
https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0331240
[6] Cultural-bias / Western-default analyses in generative AI — Nature, Humanities and Social Sciences Communications (2024); MDPI Buildings (Islamic heritage study, 2024).
https://www.nature.com/articles/s41599-024-03968-5
[7] Arch Hive AI Architecture Competition 2023 — Artificial Nature winners announcement. https://amazingarchitecture.com/news/arch-hive-ai-architecture-competition-2023-artificial-nature winners-announced
Samuvel Benhursha is an architectural designer working at the intersection of design practice and AI methodology. sbxlabs.work · @sambenhur_architecture
