img(height="1" width="1" style="display:none" src="https://www.facebook.com/tr?id=2939831959404383&ev=PageView&noscript=1")

Six things architects should know about artificial intelligence

Header Image

Words:
Phil Bernstein

If you’re worried that AI might put you out of a job, Phil Bernstein has some reassurance and realism to put your mind at ease

If news feeds are to be believed, the time that our robot overlords will take over the world is nigh, or at least right over the horizon. ‘The game is over!’ tweeted Google DeepMind’s Nandi de Freitas, extolling the virtues of the firm’s latest project, GATO, which attempts to achieve artificial general intelligence (AGI) – a human-level ability to deal with the general complexities of the world. Unlike its predecessors which slavishly focus on a single goal, GATO has a vague ‘understanding’ of 604 different tasks and so could, theoretically, scale to manage the world without a lot of training in specific contexts.

Perhaps the predictive capabilities of your word-processor may one day find themselves in your BIM tool, and eventually grow to replace you and your livelihood. But right now, here are six things that architects should know about artificial intelligence, before you decide to close up shop and open a fish and chips stand.

1 AI systems are powerful, but only in certain specific contexts, and architecture isn’t one of them. The current technology of deep learning neural networks, based on complex (and opaque) correlation models, is excellent at processing and generating language, and 2D images, and even a combination of the two. If you need a fruit-based piece of furniture for your next project you’re in fine shape, but managing something as complex as a building—particularly represented by something more robust than an image—is some distance in the future. This is because these provocative images (or sentences) are generated by statistical correlations rather than any innate understanding of their content. DALL-E understand neither avocados nor chairs in any sense more than the math that connects the words with the pictures of both. Since making buildings is significantly more complex enterprise, we’re probably safe for the moment.

OpenAI’s DALL-E system generates provocative images from text inputs; this one was “an armchair in the shape of an avocado.” See https://openai.com/blog/dall-e/ “DALL·E: Creating Images from Text”
OpenAI’s DALL-E system generates provocative images from text inputs; this one was “an armchair in the shape of an avocado.” See https://openai.com/blog/dall-e/ “DALL·E: Creating Images from Text”

2 Your clients, especially the large corporate and institutional organizations, are likely to be using AI systems and may soon ask you to do so too. In much more limited contexts, such as evaluating financial decisions, generating targeted marketing plans, scanning chest X-rays, or setting interest rates for car loans, AI has become quite useful in data-based performance and decision making. This suggests that those same clients might soon wonder why a particular aspect of your design—its energy performance, dimensions to accommodate occupancy, embodied carbon or  even supply chain ethics—is based on AI analysis. ‘Is this building beautiful or appropriate for context’ is unlikely, however, to be on that list.

HYPAR’s AI-enabled generative design platform (image courtesy of HYPAR).
HYPAR’s AI-enabled generative design platform (image courtesy of HYPAR).

3 There’s a lot of promising, but as yet unfulfilled, experimentation in AI in the broad building industry, but nothing that yet rises to the level of a great opportunity or a huge threat. A range of tech players that sell to architects are working on AI-based tools, everyone from Autodesk (SpaceMaker and Construction IQ) to start-up Hypar (AI-assisted generative design tools) has an oar in the AI water. And nobody believes they can build any modern tech without an AI strategy, at minimum. Expect some of the products and companies to reach useful maturity—and many to fall by the wayside—in the next three years.

4 Machine learning/deep learning systems need tons of data to train, and the stuff we create in the building industry is neither large enough nor sufficiently curated to be consumed usefully by today’s systems. Across the various disciplines that collaborate to create a building, there’s a lot of data generated in various incompatible formats that doesn’t create a coherent whole, and is specific to the demands of individual creators. The GATO system needed a ‘mere’ 1.2 billion training parameters, groomed carefully as inputs. It will be some time before the necessary data standards, risk and privacy controls, and basic administrative infrastructure is in place to make such data available for training the algorithms.

 

A possible structure of an Industry Data Trust that could collect information across sources and projects.
A possible structure of an Industry Data Trust that could collect information across sources and projects.

5 AI systems don’t really ‘know’ anything about their output in the sense that while the images of the avocado chairs are compelling, the system couldn’t answer a single question about chairs, sitting, avocados, or even fruit in general, nor make technical or value judgements about any of those questions (‘is that chair comfortable, or even nice?’). At the scale of an entire building, this is even more challenging, and therefore implausible for an AI to replace even the basic judgements needed to create an architectural design. Beyond reliable decision-making, we are also not ready to rely exclusively on algorithms to make important decisions, nor is there any plausible way today to hold them accountable for such decisions. This is all to say that the essential value proposition of architects—making synthetic decisions in a complex environment of trade-offs—isn’t in danger of being eliminated by robots in the near future. But…

6 …what AI systems are good at is evaluating proscribed problems with defined data sets, and systematically generating alternatives. This is a useful tool for architects who almost always rely on a combination of analysis, judgement, and intuition to make design decisions. To the extent that such decisions can be informed by algorithms that probe a wider array of sub-options, and inform the architect about their efficacy, they augment a human designer’s power to understand problems and solve them.

All of this is to say that the day that your client will access her artificial architect via a web browser and a pair of VR glasses is far in the future. A little closer to reality is a world where AI plays an increasing role in our daily business, with specific opportunities for architects to leverage data and analysis toward more robust and predictable results. Rather than a threat, maybe this is an opportunity to offer more valuable services to those clients and keep the robots at bay.


Phil Bernstein is an associate dean and professor adjunct at the Yale School of Architecture. His book Machine Learning: Architecture in the Age of Artificial Intelligence was published by RIBA earlier this year

 

Latest

Tuesday 20th September2022, 9 am – 11:30 am

PiP Design for Sustainability Webinar 2022

Could alien landings help us understand the climate emergency better, asks Eleanor Young

Eleanor Young gets sustainability inspiration from Bristol's Martian house

A new collection offers six palettes in eight colours and designs to give architects and specifiers a host of options for creating bespoke floors for every sector

Arturo Colour Collection offers architects six palettes to choose from

Right-wing sympathisers are infiltrating architectural criticism to peddle reactionary views online, says Will Wiles

Dangerously right-wing polemics have infiltrated some seemingly harmless criticisms

Reorganising the ground-floor layout made the terraced house on Copeland Road better suited to modern family life, says practice director Tom Gresford

Reorganising the layout made the terraced house better suited to modern family life, says Tom Gresford