Are we at the dawn of the AI-created city?

Just over a cen­tu­ry since The Man­i­festo of Futur­ist Archi­tec­ture declared the city must be rethought and rebuilt like an “immense and tumul­tuous ship­yard” – “every­where dynam­ic”, and the house like a “gigan­tic machine”, it may be that author Anto­nio Sant’Elia had things the wrong way around. 

Because although his machine-fetishis­ing sketch­es inspired our com­mon vision of a sci­ence-fic­tion future – as in Fritz Lang’s 1927 film Metrop­o­lis, with its tech­no­log­i­cal Tow­er of Babel an impos­ing cen­tre­piece – it might be the gigan­tic machines that are mak­ing our hous­es.

Archi­tec­ture and AI vision­ar­ies – form­ing espe­cial­ly around MIT in the 1950s, through to the lat­er work of MIT Media Lab co-founder Nicholas Negro­ponte – and design pio­neers have long thought about automat­ing the cre­ation of our envi­ron­ments. Now the tech­nol­o­gy is catch­ing up to their ideas, and a rad­i­cal shift into AI-assist­ed design is tak­ing hold, with impli­ca­tions that could rad­i­cal­ly trans­form the form, feel and func­tion of the places we inhab­it.

Com­plete­ly auto­mat­ed design is not quite there yet. This crop of gen­er­a­tive, AI-assist­ed tools is rather new. But there are signs that we could be on the cusp of a rev­o­lu­tion in how our build­ings, towns and cities are cre­at­ed. Will these begin to take on a homo­ge­neous shape, recog­nis­able as AI-planned spaces? And is this the begin­ning of the ‘copy-and-past­ed’ city – or do we already inhab­it those, with the iden­ti­cal new-build prop­er­ties that seem to crop up every­where? 

AI helping hand 

Advo­cates argue that AI-based city design could remove bur­den­some man­u­al labour, allow­ing archi­tects, design­ers and plan­ners to focus on cre­ativ­i­ty. On the oth­er hand, could AI accel­er­ate more of the same – a ruth­less­ly effi­cient approach to stuff­ing more peo­ple into build­ings and max­imis­ing rents. What­ev­er hap­pens, AI-assist­ed design appears set to rad­i­cal­ly change the future of archi­tec­ture.

Despite the long tail of think­ing around auto­mat­ed design, the draft­ing process was large­ly man­u­al until very recent­ly, even in soft­ware like the ubiq­ui­tous Auto­CAD or the build­ing infor­ma­tion mod­el­ling tools that added more con­text to designs and have become dom­i­nant. “There was always the dream of automat­ing design and urban plan­ning but, lit­tle by lit­tle, it hap­pened over the last decade or so,” com­ments Imdat As, archi­tect and co-edi­tor of The Rout­ledge Com­pan­ion to Arti­fi­cial Intel­li­gence in Archi­tec­ture.

The machine learn­ing rev­o­lu­tion has helped cre­ate the con­di­tions for ade­quate com­put­ing pow­er, with the imi­ta­tion-think­ing enabled by neur­al net­works final­ly mak­ing gen­er­a­tive design com­mer­cial­ly viable. Nudg­ing this AI-assist­ed world into real­i­ty are new tools backed by Sil­i­con Val­ley, such as Delve, owned by Google sub­sidiary Side­walk Labs, and Space­Mak­er, which was recent­ly acquired by com­put­er-aid­ed design giant Autodesk for $240m (£196m).

There was always the dream of automat­ing design and urban plan­ning

Unlike the painstak­ing­ly craft­ed line-by-line process­es nor­mal­ly asso­ci­at­ed with archi­tec­ture pro­pos­als, these tools allow the user to view and play with a huge range of vari­ables – pri­ori­tis­ing or adjust­ing nuances we may take for grant­ed, like noise lev­els, tem­per­a­ture or win­dow views – and then gen­er­at­ing design options. With the tra­di­tion­al approach, plan­ning teams are lim­it­ed by their time and their tools, so pro­pos­als rarely exceed a selec­tion of three to five designs. But by using AI-assist­ed tools, plan­ners can explore hun­dreds, if not thou­sands of options, with their sub­tle dif­fer­ences illus­trat­ed on a 3D map so var­i­ous stake­hold­ers can view progress or col­lab­o­rate as plans evolve. 

A huge shift in architectural planning

Space­Mak­er co-founder and CEO Håvard Hauke­land says that this is one of his platform’s key ben­e­fits: because urban plan­ning is very much about com­pet­ing inter­ests, projects can end up bogged down by meet­ings where peo­ple are spend­ing more time putting for­ward their cas­es, rather than explor­ing mul­ti­ple solu­tions. The lat­ter, he argues, is “bet­ter for the city, bet­ter for the peo­ple liv­ing in the apart­ments or using the office spaces, and it’s usu­al­ly bet­ter for the eco­nom­ics of the project and the devel­op­er”.

Hauke­land adds that this approach could rep­re­sent a huge shift in archi­tec­ture and plan­ning – and one that can vir­tu­al­ly elim­i­nate what those in the indus­try colour­ful­ly term ‘Oh Shit Moments’: when a design has already been fixed, but the team had for­got­ten to car­ry out essen­tial tasks like noise analy­sis, there­by poten­tial­ly set­ting project dead­lines back – some­times quite lit­er­al­ly to the draw­ing board.

That was a fate avoid­ed by the grow­ing Kivistö dis­trict in Van­taa, Fin­land, where a new rail­way line will see its pop­u­la­tion swell to 45,000 in the com­ing decades (a rel­a­tive­ly large pop­u­la­tion in Finnish terms). 

For plan­ners, it was cru­cial to bal­ance new, dense urban neigh­bour­hoods with­in it with the district’s prox­im­i­ty to nature and its sil­ver birch-lined streets, while also remain­ing on course for a 2050 car­bon neu­tral tar­get. Even at a late stage, design­ers were able to use Space­Mak­er to refine plans for inte­ri­or court­yards, reduc­ing wind effects and plac­ing a sun­ny ter­race for future res­i­dents. “The soft­ware almost down­right per­suades one to try dif­fer­ent options,” com­ment­ed town plan­ner Ville Lep­pä­nen.

Mean­while, in Sofia, Bul­gar­ia, city plan­ners applied Delve to a city unit to map out future devel­op­ment strate­gies for the area. They told RIBAJ that the test project pro­vid­ed “valu­able and impor­tant insight into the many pos­si­bil­i­ties that para­met­ric plan­ning offers”, with one of the key ben­e­fits being able to “rapid­ly change input data and gen­er­ate out­put results that may not have been even con­sid­ered before”.

This kind of exper­i­men­ta­tion was just not pos­si­ble at this pace or scale pre­vi­ous­ly. Michele Peli­no is prin­ci­pal ana­lyst, edge com­put­ing and the inter­net of things, at For­rester Research. She notes that although future-look­ing cities like Sin­ga­pore had exper­i­ment­ed with dig­i­tal twin tech­nol­o­gy – where vir­tu­al copies are made of exist­ing places and are then sub­ject to com­put­er sim­u­la­tions – applied gen­er­a­tive design is new ter­rain and how it will shape our build­ings and cities is yet to be deter­mined. 

Model renaissance

With the right prompts and some patience, algo­rithms appear capa­ble of help­ing to cre­ate stun­ning por­traits and fan­tas­ti­cal worlds, as evi­denced by OpenAI’s Dalle2 sys­tem: an eas­i­ly nav­i­ga­ble foun­tain of artistry that any­body can use. This AI-gen­er­at­ed art is a win­dow into what will become pos­si­ble on a larg­er scale with our build­ings, espe­cial­ly when com­bined with 3D print­ing, com­ing togeth­er to encom­pass one auto­mat­ed process. So says Eleanor Wat­son, IEEE AI Ethics engi­neer and AI Fac­ul­ty at Sin­gu­lar­i­ty Uni­ver­si­ty. 

This could, she says, build works of incred­i­ble com­plex­i­ty but at no extra mar­gin­al cost – and would be an oppor­tu­ni­ty to push back against the “stark util­i­tar­i­an mass-pro­duced sim­plic­i­ty, inof­fen­sive and time­less yet dull and soul­less”. “We might soon see a renais­sance, where­by plain­ness becomes passé, in a world where beau­ty has become next to free,” she says.

First, though, there are many more com­plex­i­ties to cre­at­ing a build­ing (and even more on the scale of towns and cities) than gen­er­at­ing a pleas­ing por­trait image. With all their vari­ables, loca­tion-spe­cif­ic con­sid­er­a­tions, the con­text-depen­dent nature of floor­plans, and the algo­rith­mi­cal­ly impos­si­ble-to-pin-down over­all feel­ing of a place, it may be some time yet before machines are help­ing to bring about that renais­sance.

For now, it is the nuts-and-bolts stuff where the lat­est AI-gen­er­a­tive tools excel; the design is not quite end-to-end – mean­ing, push­ing a but­ton won’t instant­ly gen­er­ate you a build­ing or dis­trict to your lik­ing. 

You want to encour­age pos­i­tive muta­tions and that’s what AI and machine learn­ing make pos­si­ble

But even these opti­mi­sa­tions have the poten­tial to change the look and feel of spaces, adds Peli­no. 

Being able to analyse, cal­cu­late and map pre­dict­ed tem­per­a­tures, for instance, could help devel­op­ers avoid cre­at­ing urban heat islands, and cre­ate cool­er con­di­tions for res­i­dents as build­ings and cities evolve. And as sus­tain­abil­i­ty becomes a more press­ing con­cern, it may be proven that our approach­es to build­ings and cities are woe­ful­ly inad­e­quate – and that AI-imag­ined geo­met­ric mod­els sur­prise design­ers with the opti­mised shapes they take. 

In time, as tech­nol­o­gy march­es for­ward, new, sur­pris­ing, aes­thet­ic forms may crys­tallise. 

Stephen Bar­rett, part­ner at archi­tec­tur­al design­ers RSHP, believes that AI will be able to take the more mech­a­nis­tic day-to-day activ­i­ties of plan­ners and design­ers, and “auto­com­plete” some of the labo­ri­ous process­es. There are “great advan­tages” to this, he says: it frees up time and space to work on the inter­est­ing stuff, to inno­vate and cre­ate. 

At present, for instance, an AI-gen­er­at­ed Picas­so scene will cre­ate a rough car­i­ca­ture of the artist’s style based on the inputs fed to it. But how­ev­er impres­sive it might be, it is no more than an approx­i­ma­tion – a kind of evolved copy – of exist­ing images and aes­thet­ics rather than some­thing alto­geth­er new. 

So, deter­min­ing the future with algo­rithms needs to be con­sid­ered “very care­ful­ly”, he says: “It’s a lit­tle bit like Mod­ernist archi­tec­ture in the mid-to-late 20th cen­tu­ry. It was meant to be a form of archi­tec­ture free of bag­gage and val­ues but, in the end, you could argue it was fair­ly post-colo­nial dom­i­nant – and you have the same build­ings in São Paulo as you have in Riyadh as you have in Sin­ga­pore or Moscow or wher­ev­er.

“You want to encour­age pos­i­tive muta­tions and that’s what the rapid pro­cess­ing and mul­ti­ple iter­a­tions of AI and machine learn­ing make pos­si­ble,” he adds. “But also, to ensure that the out­put is intel­li­gent, and not sim­ply a reflec­tion of the lim­i­ta­tions of the inputs.”

So design­ers will have to tread care­ful­ly and remain con­scious of algo­rith­mic bias, where soft­ware repro­duces errors due to the prej­u­dices of the soft­ware design­ers. 

Flair over form for the foreseeable

Hauke­land argues that more design options and there­fore more vari­ety can hard­ly ever be a bad thing. But if you look at his­to­ry, he adds, rev­o­lu­tions in archi­tec­ture have occurred across thou­sands of years and in the end, there’s always some­thing new, bet­ter, or smarter that builds on the past. “It’s easy to stand here at the begin­ning of a new era and say this will change every­thing,” he says. “But in 10, 20 or 30 years’ time, we might say that gen­er­a­tive design and AI was super inter­est­ing but it had its weak­ness­es. So I don’t think humans have come to the end.”

Imdat As won­ders what design­ers would work on if AI were to pro­duce 90% of the build­ings. 

“The top 10 design­ers – the Zaha Hadids, and so on – will always be there, with new ideas, new aes­thet­ics. Those will be the design­ers who come up with a new design idea,” he says. “And what if they trained in-house AI sys­tems? Instead of, say, 10 build­ings a year, they might build a mil­lion, all over the world. The pow­er of AI for an archi­tec­tur­al com­pa­ny could be amaz­ing. The busi­ness struc­ture could be: if you use my AI sys­tem, you pay roy­al­ties. It could change archi­tec­tur­al prac­tice mod­els. I think there will be those types of changes.”

While machines are learn­ing to nav­i­gate envi­ron­ments, they can nev­er know the expe­ri­ence of doing so

Present­ly, it’s unlike­ly that res­i­dents would notice at all if com­put­ers helped to build their hous­ing, angling a com­plex five or six degrees to max­imise live­abil­i­ty. And it may be too ear­ly to tell if algo­rith­mi­cal­ly designed build­ings and cities will leave tell­tale AI marks in their floor­plans or on their façades, but As is hope­ful that “its own sort of urban mor­phol­o­gy” will emerge.

Bar­rett won­ders if the “step-change rev­o­lu­tions you get with acci­den­tal, errat­ic, unique or eccen­tric human inputs” could ever occur with arti­fi­cial intel­li­gence thrown into the mix. “If you took a city like Paris and ran that as an exist­ing data set, you’d have more Paris,” he com­ments, “which is not a bad thing. But would you ever have had a Pom­pi­dou Cen­tre?”

One thing does, though, seem cer­tain. Giv­en the effi­cien­cy gains, AI-assist­ed design will play an increas­ing­ly impor­tant role in that future plan­ning, devel­op­ing, build­ing. But just as with soft­ware and the com­plex data sets that inform or medi­ate our lives, keep­ing a human in the loop is like­ly to be a fix­ture in the fore­see­able, “mar­shalling, judg­ing, and con­tin­u­ing to select”, says Bar­rett. 

“After all, while machines are learn­ing to nav­i­gate envi­ron­ments, they can nev­er know the expe­ri­ence of doing so,” adds Wat­son. “We must nev­er sac­ri­fice the feel of an urban land­scape at the altar of effi­cien­cy, nor cause a mal­func­tion in any person’s enjoy­ment of a resource.” 

---