The creative line: Where AI should stop in architecture

Imagine a digital nation edging towards the one-billion mark, logging in each week for instant answers. That is the scale of ChatGPT now, with OpenAI reporting more than 900 million weekly active users worldwide, which is roughly twice the population of the European Union.

And it is not alone at that scale. Google says Gemini has passed 750 million monthly active users, and Claude is widely reported to be gaining users rapidly, with Claude Code and related developer capabilities shifting it from occasional chat to practical daily use.

The potential is obvious. A tool that can process enormous amounts of information in seconds could accelerate research, streamline development, and expand what we are able to learn and build.

Yet as more design studios experiment with Artificial Intelligence (AI) in architecture, a curious pattern is emerging. It is not the one we were promised.

Despite the hype, the outcomes have been surprisingly underwhelming.

Not because the tools lack power. Quite the opposite. AI is proving to be remarkably capable. The issue sits elsewhere. It is not what the technology can do, but how we are choosing to apply it.

Innovation only earns its place when it meets a genuine need more successfully than what already exists. Speed, on its own, is not a purpose. In architecture, the purpose is a better life. Comfort, dignity, belonging, safety, delight, and long-term value.

It’s also worth remembering that we are not built like machines. Humans arrive with five senses and a body so complex that science is still catching up with its full interconnections. No technology can replace those senses as they exist within us. AI may support the process, but it must never become the yardstick of quality.

Many architects now lean on AI in early-stage design, from massing studies to façade concepts. At first, it feels like a breakthrough. Quicker turnarounds, polished visuals, fewer hours spent on first iterations. It promises efficiency, and it delivers.

But over time, something troubling starts to surface. The work begins to look eerily similar across studios and project types. Refined, yes. Sophisticated, certainly. But also generic, as though drawn from the same limited pool of possibilities.

This explosion of instant, accessible AI-powered design solutions comes with other serious risks too. The same ease that makes these tools so appealing, with their immediate solutions, effortless responses and seamless assistance, could gradually weaken our capacity for independent thinking.

The problem isn't the technology. It's that we've confused efficiency with creativity, and in doing so, we're removing the very friction that makes design meaningful.

Artificial intelligence (AI) in architecture refers to the integration of machine learning, computer vision, and other data-driven tools into the design and construction process.

AI influences the architectural design process by rapidly producing multiple design options based on set parameters and identifying patterns across vast datasets.

While this accelerates certain stages of design development, it also raises critical questions about authorship, originality, professional accountability, and the risk of favouring efficiency over thoughtful, context-sensitive creativity.

Recent research from Apple delivered a sobering reality check to the AI industry. 

In a study titled "The Illusion of Thinking", Apple researchers tested leading “reasoning models” using controllable puzzle environments where complexity could be increased while the underlying logic stayed consistent. They reported that beyond certain thresholds of complexity, frontier models suffered what they described as a “complete accuracy collapse”.

Apple’s team also found limits in exact computation. In their setup, models often failed to follow explicit algorithms reliably and produced inconsistent reasoning traces across puzzles.

These conclusions have been debated. A published critique argues that parts of the apparent collapse may be shaped by evaluation design choices, including output length constraints and scoring issues, rather than serving as definitive proof that the models cannot reason.

Even so, newer research points to a related weakness from a different angle. In January 2026, researchers introduced DeepPlanning, a benchmark designed to test long-horizon agentic planning under verifiable constraints.

The point is simple: real-world tasks demand proactive information gathering, local reasoning, and global optimisation across many dependent steps. The researchers report that frontier models still struggle to satisfy these holistic constraints reliably. That matters for the built environment because development is a long-horizon problem too, with thousands of linked decisions where small errors compound over time.

For architecture, the implication remains serious. If these systems degrade under high compositional complexity, they are not safe to treat as autonomous decision-makers in contexts that demand nuanced, context-sensitive judgement. They may be powerful assistants, but they still require professional oversight and accountability.

Here’s where many architectural practices are getting lost: they’re confusing options with ideas.

AI tools may now give architects more options than ever before. Façades in seconds, layouts on command, hundreds of variations at the click of a button. But options aren’t the same as ideas.

Hear me out.

Good design isn’t about choosing from a menu. It’s about making judgements: what’s appropriate for this specific context, how people will actually move through the space, how the building will weather, adapt and settle into its community over time. AI may generate possibilities, but it doesn’t understand the lived conditions that turn possibility into a meaningful architectural concept.

And design isn’t just visual. It’s multisensory. We don’t experience buildings as renders. We experience them through glare and softness of light, the acoustics of a corridor at night, the feel of a handrail, the smell of damp air or fresh ventilation, and the way a plan supports daily rituals. Most AI output is still overwhelmingly image-led, which makes it dangerously easy to design for the screenshot rather than the lived moment.

That’s the risk. When architects start treating AI output as design insight rather than design input, they drift into curating content instead of creating meaning.

In a discipline where genuine breakthroughs come from wrestling with constraints and working through messy, uncertain processes, algorithmic shortcuts are seductive. But if speed becomes the goal, substance is what gets left behind.

They're essentially outsourcing the very thing that makes architecture more than just arranged matter: the thoughtful response to human need and cultural context.

Research using AI image generators like Midjourney has confirmed what many architects intuitively sense: initial prompts consistently produce homogenised designs that reflect the dominant architectural styles present in the AI’s training data.

The result is what experts now call "AI Slop", a proliferation of generic imagery that adopts a neutral, globalised aesthetic, free from contextual friction.

This isn't just an aesthetic problem; it's a cultural one. Architecture that could come from anywhere effectively comes from nowhere. It carries no memory of place, no response to climate, no acknowledgement of local building traditions or social patterns. The very qualities that make architecture culturally significant are precisely what get smoothed away by AI's pattern-matching approach.

In plain terms, AI doesn’t invent new design languages or revolutionary ideas. It repackages what already exists. It analyses patterns, recombines familiar elements, and generates polished variations.

While the fact that AI lacks true creativity is a concern, it’s not the only one. The deeper risk is that architects, by increasingly leaning on these tools, may erode their own capacity for the kind of original, challenging thinking that leads to truly transformative design — the kind that breaks patterns rather than simply rearranging them.

But don’t just take my word for it. A study by ArchDaily and Ulises Design Studio, which explored how AI envisions contemporary homes across 15 countries, initially produced remarkably similar results regardless of location. Only when researchers deliberately refined their prompts to include specific regional elements did the outputs begin to show cultural diversity. 

The lesson is clear: AI's default mode is homogenisation.

We talk extensively about what AI means for design efficiency and creativity, but we don’t talk nearly enough about the burden of responsibility.

When a building fails, whether functionally, socially, or environmentally, who is accountable?

This isn’t an abstract question. Architecture is inherently public. It shapes how people live, connect, and belong to their communities. Buildings outlast their creators and affect generations of users who had no voice in their conception.

That’s precisely why planning matters. It is society’s way of testing private design decisions against public values: neighbourliness, heritage, access, safety, and the wider shape of the city. We may automate parts of compliance, but legitimacy still depends on human judgement that can be explained, challenged, and trusted.

In the UK, this accountability is codified through the Construction (Design and Management) Regulations 2015, which establish clear legal duties for what are known as "duty holders" throughout a project's lifecycle. 

The Principal Designer, for instance, must "plan, manage, monitor and coordinate health and safety in the pre-construction phase," whilst ensuring that designs "eliminate, reduce or control foreseeable risks that may arise during construction and the maintenance and use of a building once it is built." Yes, these are professional obligations but they're also legal responsibilities enforceable by criminal law.

The CDM framework recognises that design decisions have profound consequences for human welfare, establishing that designers must consider not just immediate construction safety but long-term building performance and maintenance. 

When AI systems generate design solutions, who carries this legal responsibility? The algorithm cannot be prosecuted for failures; the human professionals who implement its outputs must be held accountable.

UNESCO's AI ethics guidelines explicitly state that AI systems must not displace ultimate human responsibility and accountability. Yet the architecture profession risks doing exactly that when it delegates design intent to algorithms without maintaining meaningful human oversight.

Phillip Bernstein, deputy dean at the Yale School of Architecture, echoes this concern. “Architects have several responsibilities beyond simply designing buildings,” he explains. “We are responsible for public health and safety, which is why we are licensed to practise… Can we really delegate these professional responsibilities to an algorithm? Would society want us to do that?”

His point is clear: Professional accountability cannot be passed to a machine, and doing so risks not only legal uncertainty, but a fundamental erosion of architectural responsibility.

This reinforces the urgent need for architects to retain agency and judgement in the design process, particularly when AI tools are introduced into workflows that carry legal, social, and ethical implications.

The European Union's AI Act emphasises the critical importance of human oversight in high-risk systems, particularly those affecting fundamental human rights. Buildings - especially housing, healthcare facilities, and educational institutions - surely qualify as systems with profound impacts on human welfare.

This doesn't mean rejecting AI wholesale. Instead, it means being far more deliberate about where we deploy it. 

In my opinion, the most promising applications aren't in the creative realm but in the administrative and repetitive tasks that currently consume enormous amounts of architectural time and energy.

Building regulations compliance checking, for instance, represents a promising area where AI can assist in initial reviews, flagging potential issues and streamlining preliminary assessments.

Whilst AI cannot replace the nuanced understanding required for complex compliance decisions, it can accelerate the identification of clear-cut violations, allowing architects to focus their expertise on ambiguous cases and design refinements rather than routine checking.

Similarly, AI shows promise in supporting structural analysis, energy performance evaluation, and material quantity calculations.

These computational tasks benefit from AI's pattern-recognition capabilities for initial assessments, though architects must still apply professional judgement to validate results and make final determinations. In other words, AI can handle the heavy lifting of data processing, freeing architects to concentrate on design refinement and creative problem-solving.

The BIM process offers particularly valuable opportunities for AI assistance. AI-powered tools can support clash detection between building systems, assist in generating preliminary cost estimates from model data, and help streamline the preparation of construction documentation. However, these tools work best as intelligent assistants rather than autonomous decision-makers.

Recent advances in AI-enhanced BIM include virtual assistants that can interpret natural language queries to extract information from complex building models, allowing architects to ask questions like "What is the structural load capacity of this wall?" and receive rapid initial responses that still require professional verification.

A strong design still needs a pathway to consent. In the UK, that pathway is planning permission, with policy, consultation, officer judgement, committee decisions, conditions, and ongoing compliance shaping the outcome.

By 2030, the planning “interface” is likely to feel less like a pile of PDFs and more like structured information moving through connected systems. Think constraint screening, validation, evidence logs, condition tracking, and faster technical checks. This is where AI genuinely helps, not by replacing planning judgement, but by stripping out avoidable friction.

But planning is not a rules engine. Much of it sits in the balancing exercise: harm versus benefit, design quality, impact on neighbours, heritage significance, local character, and the wider public interest. These are not questions you can simply compute. They are questions of legitimacy.

That’s why the future interface will be two-layered. One layer is machine-facing, dealing with what is measurable, repeatable, and auditable. The other is human-facing, where decisions are explained in plain English, open to challenge, and grounded in trust.

And that’s exactly why the creative core must remain protected. If we let machines steer early intent, we won’t just lose originality. We will erode the human purpose behind the planning system in the first place.

The key is protecting what matters most: The early creative stages that give a project its soul. 

This is where architects develop design intent, respond to site conditions, and translate client needs into spatial concepts. It's messy, iterative, and fundamentally human work that benefits from friction, not efficiency.

When we use AI to bypass this essential struggle, we're saving time, sure, but we're also eliminating the very process through which architectural meaning emerges. The constraints, the false starts, the moments of uncertainty - these aren't inefficiencies to be optimised away. They're integral to how architects develop contextually appropriate responses to design challenges.

VR in architecture offers a compelling example of technology enhancing rather than replacing human creativity. Virtual reality tools allow architects to experience their designs at full scale during development, providing invaluable spatial understanding that informs design decisions. 

Unlike AI-generated concepts, VR serves as a sophisticated evaluation tool, helping architects test their ideas without compromising the creative process that generates those ideas. Recent research shows that VR creates "a collaborative environment that minimises misunderstandings and costly post-construction modifications" whilst maintaining human control over design intent.

Research consistently shows that human oversight remains indispensable for nuanced decision-making. Humans possess emotional intelligence, cultural knowledge, and ethical reasoning that AI simply cannot replicate. These qualities are particularly crucial in architecture, where every decision affects how people inhabit and experience space.

So what does all this mean?

In short, the architecture profession stands at a crossroads. We can either rush headlong into AI adoption, treating every new tool as an opportunity for efficiency gains, or we can be more strategic about where artificial intelligence adds genuine value.

The most successful residential architects will likely be those that draw clear boundaries: AI for administration, humans for creation. AI for analysis, humans for synthesis. AI for options, humans for ideas. 

This method is already showing promise in practices that integrate AI-enhanced BIM processes with traditional design methods, using technology to eliminate tedious documentation tasks whilst preserving creative control.

This approach requires discipline. It means saying no to the allure of AI-generated concepts, even when they look impressive. It means maintaining the slow, thoughtful work of early design development, even when algorithms promise faster results. For architects, this might mean using AI to optimise thermal performance and structural sizing whilst insisting on human-led spatial planning and aesthetic decision-making.

But it also offers something valuable: the possibility of using AI to eliminate the tedious work that prevents architects from spending more time on the creative and social dimensions of their practice. 

Instead of replacing architectural thinking, AI can create space for more of it.

All in all, Apple's research reminds us that current AI systems, for all their sophistication, are fundamentally limited in their reasoning capabilities. They excel at pattern matching but struggle with genuine innovation. They can optimise known solutions but can't imagine entirely new approaches to persistent problems.

For architecture, this suggests a clear division of labour. 

Let AI handle the administrative burden, the compliance checking, the documentation coordination, the performance optimisation. But preserve the creative core for architectural services and professional judgment: the spatial imagination, the cultural sensitivity, the ethical reasoning that transforms building programmes into meaningful places.

So the aim is not to resist technology. It is to put it in its proper place.

Use AI to interrogate decisions, not to replace them. To stress-test ideas, not to supply intent. To support practice, not to set the standard. Because once the benchmark becomes what a model can produce quickly, we start optimising for the wrong outcome.

Buildings are not consumed as images. They are lived through the body, across five senses, over years. If we forget that, we will trade meaning for momentum.

And meaning is the job. Architecture is not merely the delivery of efficient structures. It is the responsibility to shape environments where human life can flourish. That duty does not disappear because a tool got clever. If anything, it becomes sharper. It’s a duty we embrace in every project we design, a dedication to creativity, empathy, and human-centred thinking that’s evident throughout our portfolio.

AI will keep transforming the profession. It already is. The only real question is whether we stay in control of what matters most: creative intelligence grounded in human need, expressed through spatial imagination. That line is worth defending.