Autonomous Algorithmic Architects: Wicked Problems of Machine Learning in Architecture
Image credit: Archinect / Midjourney
How should the architectural profession consider and respond to futures made possible by advances in artificial intelligence? Can we propose a willful, designed route that acknowledges the inevitability of intelligent machines intersecting with every dimension of design and construction, while maintaining a proper role for human architects? Architect and Yale School of Architecture Associate Dean Phil Bernstein reflects on the opportunities, threats, and resulting strategies for the confluence of human and machine intelligence in the profession in this article adapted from his book Machine Learning: Architecture in the Age of Artificial Intelligencepublished by RIBA in 2022.
Could a computer design an entire building well? Computer scientist Mark Greaves describes the advances in computational creativity in natural language generation with tools like GPT-3 (and now GPT-4) as having ‘fluency and expressivity’:
“Using modern machine learning techniques, machines are starting to successfully perform creative, original tasks in domains like language that were once uniquely the realm of humans. There have certainly been limited achievements based on more traditional AI, which have been called ‘creative’, such as the famous ‘hand of God’ move played by Deep Blue in its chess match against Garry Kasparov. But these are quite rare… These systems seem to exhibit a level of creativity and expressiveness and linguistic artistry that machines hadn’t reached in the past. And, in the realm of game playing, ML-based AlphaGo has shown real creativity as well.”1
How should the profession of architecture consider and respond to futures made possible by advances in artificial intelligence?
Of course, recent advances with tools like Midjourney or DALL-E take machine “originality” closer to architecture. But creativity and coherence do not equate with competence, however, and therein lies at least part of the answer to the question that we must ask ourselves as architects facing the oncoming wave of AI: how should the profession of architecture consider and respond to futures made possible by advances in artificial intelligence? Having moved reluctantly through the eras of both CAD and BIM, can we propose a willful, designed route – a professional strategy – that acknowledges the inevitability of a preponderance of intelligent machines in every dimension of design, construction, and built asset operation while maintaining a proper role for human architects?
New Tools and New Anxieties
If designers solve, as described by Peter Rowe (quoting Horst Rittel) ‘wicked problems’,2 with open-ended beginnings and no fixed conclusion, competent practice requires heuristics across a broad spectrum of technical and aesthetic issues. While these are human strategies, computers are increasingly able, empowered by machine learning, to learn these techniques, and when they do so, professional strategies and methods – and the value of designers themselves – will be inalterably transfigured. But there are challenges.
As Stanford computer scientist, Roy Amara, is purported to have said, ‘We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run’
Architects are often anxious and ambivalent at best about new technologies: the conviction that our work as architects is a uniquely valuable contribution, paired with the paranoia that capable machines will mercilessly replace us – the source of our profession’s angst about machine intelligence and its putative disastrous effect on the design process. However, as Stanford computer scientist, Roy Amara, is purported to have said, ‘We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run,’3 which is certainly the case with architecture’s current concerns about artificial intelligence. With the recent explosion of technologies like language transformers such as ChatGPT and image generators such as Midjourney, it feels like things are coming faster than usual, and these systems operate with increasing autonomy from their human creators.
Opportunities That Precede the Threats
We can speculate on the likely set of opportunities that architects will see in the next decade as such systems become more capable and available. These categories, which we will call ‘autonomous outputs’, form a speculative framework from which we can begin to build strategies for the implications for the profession, and might include the following:
Design task automation: Procedures and protocols that require the direct intervention of the designer as likely to be autonomous in the future. For example, while most code checking is a manual process today, that procedure can be supported by submitting a digital model to a code-checking tool that uses AI to evaluate code compliance, combining a more traditional ‘architectural’ model with a technoscientific counterpart. Tomorrow, that AI-based code checker could be lurking in the background of a BIM process, anticipating code-related problems as the designer creates her scheme. Similar autonomous protocols might reach across the entire delivery and operation life of a building, including the following:
Alternative generation, exploration: Design alternatives were once created entirely by direct manipulation of design information, like the cardboard models shown below. Today this process is assisted by scripting, a form of ‘automating’ the control of certain digitally controlled parameters of a model to vary its characteristics and thereby create a variety of solutions. Those scripts are sometimes combined with analytical software – such as energy analysis – to evaluate and optimize.
Construction automation: As robotics accelerate with AI-assisted systems for control, enhanced by computer vision and bolstered by the advent of industrialized construction processes borrowed from manufacturing, construction tasks once performed exclusively by human workers will be augmented, and in some cases, replaced, by autonomous devices on the job site.
Autonomous building function: If the architect’s and engineer’s initial BIM is a first functional descriptor of a building that might include the performance objectives of its systems the operating air, water, waste and signal systems of that building generate another digital collection of data that, in concert with real-time analytics, can be used to calibrate and optimize those systems. Companies that provide such building infrastructure, such as Johnson Controls, build not just, for example, an air distribution system but also digital controls for the system that communicates with AI-based monitors that memorialize and optimize the system output and use of energy (thereby reducing carbon). Of course, the resulting analytical data sets, interpreted by AI, can also provide insight into the design of subsequent buildings and their component equipment.
The Augmented Architect
Let us assume that an architect in the AI future has a complement of ML-enabled tools at their disposal, along with significant advancements in the resolution, precision, and flexibility of modeling platforms that one hopes would be the logical successors of today’s BIM. Such systems would be tightly tied to design modeling/representational platforms and their data and receive training from other information sources like engineering systems, real-world data collection about context from LIDAR or GIS, construction management sources that describe process and results from contractors and building operations data from existing projects controlled by sensor-driven building management and control systems. These systems are likely to be semi-autonomous, cloud-based agents that operate in the background of the architect’s process, appearing when the architect demands some piece of insight or analysis.
An immediate result of such a change in design would be the importance of evidence in supporting design decisions. While the credibility of design decisions in the pre-AI era (ours today) stems primarily from the (presumably) sound judgment and intuition of an experienced architect4, in a world of ubiquitous machine intelligence those same judgments will need to be substantiated, at least in part, by evidence and analysis to back them up. The built environment has traditionally disgorged a collection of ambiguous, heterogeneous data sets, but the ability of AI systems to divine and understand patterns within it gives architects the opportunity to generate and leverage just such evidence. And since many of today’s clients rely on AI data systems to run their enterprises, architects will be expected to do the same to substantiate the decisions that form the design.
Opacity will make it impossible, in my view, for architects or others to rely on these systems without some sort of third-party validation of their results.
When structural engineers began to rely on software for routine calculations, the credibility of those results relied not so much on the regulation of or promises by the technology vendors but rather on that the engineer herself was responsible for the output of those systems and any errors that might occur because of their use. Just as BIM has now become a tool that, under the duty of care, an architect may be expected to use on a project, AI-produced results will become part and parcel of the architect’s professional judgment, despite the opacity of the underlying algorithms that produce results.
Opacity will make it impossible, in my view, for architects or others to rely on these systems without some sort of third-party validation of their results. Should the building industry, with architects as important contributors, decide to build a global data trust to drive AI, a component of that trust would include entities who would extensively test and certify the results of these systems before releasing them into the wild. The future leaders of BuildingSMART, for example, have a much bigger enterprise on their hands.
AI Monkey Business
To make all this work, today’s “BIM monkeys” will give way to “AI Monkey Trainers.” Those systems will require specialized understanding of inputs, outputs, data demands, and relationships of the AI system to the broader infrastructure of design information. These are skills that architects trained as generalists are unlikely to understand, nor, frankly, have much engagement with. The outputs of such systems will be of great interest; the process by which they are generated, not so much. While it would be nice to simply ask the architectural version of Alexa, ‘How much carbon is embodied in my project?’ the route to that answer is likely significantly more complex and will require specialists to enable it.
New Value Propositions and Business Models
It is hard to imagine another modern enterprise, even one so reluctant to really modernize like architecture, whose business models are essentially unchanged from their 18th-century precedents. Yet architecture, like much of the construction industry, remains tied to a fundamental value strategy of lowest first cost, where services are bid and purchased in a way not dissimilar to steel, sheetrock, or carpeting: maximum pressure on competitive price, with far less attention paid to the value delivered, particularly over the cycle of a project’s lifespan.
Despite these improvements, the centuries-old methods for computing architectural computation remain largely intact, suggesting that these improved services have not translated into business terms, nor profit.
BIM has allowed all members of the delivery team to generate, organize, integrate, and exchange design information at much higher levels of resolution and transparency, at least in theory. It also, in some minimal way, begin to bridge the information gulf between design and construction; builders who saw the value of 3D data began to request it to assist their work. Other digital technologies have improved information exchange and client-facing images of projects (think renderings or even virtual reality models). Yet despite these improvements, the centuries-old methods for computing architectural computation remain largely intact, suggesting that these improved services have not translated into business terms, nor profit. The MacLeamy Curve suggests that the real value of design work lies early in the delivery process, despite the relatively small degree of effort entailed there compared to production and delivery stages. Perhaps AI will begin the value shift.
A willingness to examine innovative business strategies for new services, organizational strategies, and even new products can translate the threat of AI into an opportunity to improve both our performance as professionals and our business results if we apply the same sort of creative thinking often reserved for the design studio to this problem. But how?
Strategies for AI in Architecture
Any strategy for guiding the development and use of AI systems in architecture should serve two goals: improve the quality of the built environment, and enhance the relevance of the human architects who are best suited to make those improvements through design.
Given that the development of increasingly capable modes of automation is inevitable, I propose that the profession embrace five strategies to guide its future:
Explicitly guide the definition and creation of technologies that will frame future practice. Given that the next generation of technology may well define the future of architectural practice, the profession must establish means to declare its needs and direction in a way that does not defer to the business whims of software providers, whose motivations will ultimately prioritize profits and shareholders. Architects have spoken with individual voices as customers rather than in a united fashion as a collective of important users. We should organize, collaborating with regulators, clients, designers, and builders, to declare an industry technology strategy that prioritizes the most important data and AI/ML capabilities and then demand the industry provide them.
Expand the remit of design to include explicit performance. Starting with life safety at the beginning of the 20th century, and energy performance at the beginning of the 21st century, the range of performance parameters that architects must address will continue to expand. AI systems, driven by data, can empower architects to integrate a broader set of these parameters into their design processes, connecting the generation of solutions to performance models of, for example, occupancy, economics, epidemiological implications, embedded carbon, and even embodied labor.5 These considerations do not supplant the importance of design in its traditional sense, but rather expand it, while simultaneously enlarging the effectiveness of building and the credibility of architects.
Data Trust. Image via Phil Bernstein
Create the data infrastructure that can serve as a platform for design. The AEC process today is awash in digital data—models, analyses, project documentation, lidar scans, cost data, and the like. The potential of these resources is wholly unrealized without a strategy to organize and access them, particularly in the industry’s contentious and risk-averse delivery models. The collaborative organizations described above could guide software strategy and create policies and platforms for the collection, organization, access, and use of this data, ostensibly through a global building data trust managed by a third-party fiduciary and accessible to all.
Change the relationship between design, construction, and asset operation. Digital data created by the various players who design, build, and operate assets are often incompatible and rarely shared.6 AI platforms, which could develop, manage and integrate the data relationships between these various representations and processes, can be a catalyst for allowing architects to cross the traditional boundaries that separate project definition, design, construction, and asset operation and making knowledge reciprocally available across the design intent – construction execution divide. The tools may make the opportunity, but practitioners must want to embrace it.
Shift the value propositions of design. Artificial intelligence tools strategically deployed in the service of performance-enhanced design solutions could be the catalyst for changing the fundamental business propositions of practice, converting the value of the architect’s services from deliverables and fixed fees to outcome-based delivery models and related services. AI/ML systems could radically accelerate the capability of today’s algorithmically driven software tools to predict the future state of project performance, generating the best value by virtue of simulation. As soon as reliable AI-based tools –resulting from the implementation of the previous four strategies – become widely available, architects could embrace the largest challenges of architecture and society, and finally escape the tyranny of commodified fees, limited resources, and public skepticism about the value of the buildings they design.
Best to decide now where the best opportunities lie—and declare the same to the vendors who will build the tools—before being washed ashore by successive waves of more capable, autonomous algorithms.
These new technologies are moving quickly into the mainstream, and it’s likely that architecture’s typically plodding embrace of digital innovation will be pushed hard from behind by the demands of clients, pull from contractors, and the temptations of easy productivity gains. Best to decide now where the best opportunities lie—and declare the same to the vendors who will build the tools—before being washed ashore by successive waves of more capable, autonomous algorithms.