AI and the built environment

It's fair to say that most of us did not do terribly well at predictions for 2022. Few saw the Ukraine war coming and while I did think that China would relax their zero-covid policy, I thought it would be under far more controlled circumstances than has been the case. Closer to home it was a messy year for planning and the built environment disciplines. While we started and ended the year with the same minister in charge at DLUHC, the interregnum and lengthy hiatus for any kind of policy-making during the summer and autumn helped no-one. 

I have little optimism that 2023 will bring clarity rather than chaos, but looking more broadly at the policy and political field, I wonder if this may be the year that AI and related technologies start to really impact the space. And whether this will be a net positive or negative for trying to create really good places that work for communities, are climate-adapted as well as emissions-neutral, and have the longevity and adaptability - measured in centuries not decades - that is increasingly the aspect of design that interests me the most.

The chatGPT model is broadly speaking about generating text, and code (which is broadly text-like). The AI 'art' programmes generate 2D visuals which may depict 3D scenes but do not reflect any underlying 3D model. Both these AI models use a machine learning process that is trained using huge quantities of pre-existing information - books, journal articles, transcripts for chatGPT, pictures and photographs for the AI art generators. Training an AI model on 3D information is harder, but it is already being done. Self-driving cars, for example, are trained on huge amounts of real-world data (camera footage, LIDAR, etc) but also in synthetic simulation environments that are quicker and cheaper to generate. Many of these environments themselves are being generated through AI using 2D-3D automation.

Design by algorithm is already here. There are hundreds of platforms offering to generate infinite numbers of floor plans for your house remodel. At the multi-residential scale, Architechtures is already on the market and Finch is gearing up with new funding announced last year. These are currently aimed at multistory residential development which is ideal for AI due to the complex and rigid rules that must be met (room sizes, furniture arrangements, fire escape distances) and the complexity of developing (and proving) compliant floor plans manually. But they also generate building forms and can easily be used to test multiple different design options with, for example, planners - without any of the time-consuming drafting that most projects still rely on. At the city scale, there's Cityplanner which, it has to be admitted, looks pretty basic currently, and for roads there's Site3D, and no doubt there are a plethora of other tools out there in development.

Machine learning is fundamentally different from these rule-based platforms as it generates new things based on what it has deduced from old things that it has been trained on. As most of the old stuff that already exists isn't compliant with current requirements, it's going to be far harder to effectively use a machine learning approach for building design. But no doubt, the rules-based algorithmically driven tools are only going to get more common.

All this leaves me feeling distinctly out of date as I contemplate going back to an office powered by 2D and 3D CAD along with a lot of hand sketches and indeed physical models made of cardboard. It also feels a million miles away from the reality of construction as we currently experience it on the scale of projects we are involved with (up to £6-7m in construction value). This is still dirty, inefficient, powered by people, with a site cabin holding racks of printed drawings covered with notes in biro. (We are, however increasingly seeing subcontractor drawing submissions that are being generated by some kind of fairly bad algorithm, because they are ridden with errors and are hideously difficult to read due to their terrible graphic design and layout on the page. The art of creating a clear and well-ordered fabrication drawing, showing exactly what needs to be understood, no more and no less, is dying fast. I found a beautiful example of this craft in our Xmas office tidy-up. It dates from 1967.)

Are small practices like ours doomed to obsolescence as we are undercut by automatic plan generators? My prediction is that, if you want certainty, expertise in the heritage sector is going to be a good bet. But I also saw two amazing works of analog-digital fusion over the holiday period that makes me think there may be a more complex future, albeit one that only some will be willing to pay for. Guillermo del Toro's Pinocchio, and the RSC/Improbable's Totoro, swept us - and so many others - up on such waves of joy, in their very different ways. Pinocchio could have been a claymation, fully digitally produced film - it wasn't, but I suspect if a viewer didn't know this, they might not have been able to tell. Totoro, of course, being in the theatre, was undubitably physical and tangible. Totoro's round body shook and wobbled gloriously in all three dimensions, to the tip of every piece of stringy fur that made up his mop-like coat.