Social housing has a reputation problem that it mostly earned. The grand mid-century projects — the estates and the blocks — are shorthand today for how good intentions become bad places. And among the causes people cite, one keeps recurring across countries and decades: uniformity. Thousands of identical units, designed for a statistical household that no actual household matches, repeated until the repetition itself became the identity of the place — and of the people assigned to live there.
I think uniformity was never an ideology. It was a cost structure. Design effort was expensive and manual, so it was spent once and amortised over ten thousand copies. One plan, repeated, was the only way the arithmetic of mass housing worked. Every critique of the sameness of social housing is, underneath, a critique of what design used to cost.
That cost just moved. Which means a question most people consider closed might deserve reopening.
The variable that changed
I’ve written before about what generative design does to the cost stack of affordable housing: a house’s cost is a stack of numbers, each improvable by a few percent, and the improvements compound once design iteration gets cheap. The argument here is that same argument pointed at a different layer: what it costs for two units to be different.
Under manual design, variation is nearly linear in effort: every distinct unit is another drawing set, another structural check, another approval package, another chance for a coordination error on site. Variation is therefore the first thing value-engineering deletes. Under generative design, the marginal cost of variation collapses. A parametric housing model produces a space of buildings. Household of two or seven, ground-floor accessibility, a work-from-home room, a corner unit catching different light, deeper balconies on the noisy side: these become inputs, and each unit becomes a distinct, checked solution to its actual occupants’ constraints, generated from the same verified system.
The design cost is paid once, on the generator. The units inherit it. Personalization stops being a luxury good — that’s the “on paper” claim, and the reason the paper matters is that uniformity was the load-bearing objection to the whole enterprise.
What personalization would actually mean here
“Personalized social housing” can sound like marketing. The version generative tools make plausible is the difference between a unit designed for the median household and a unit fitted to this household — family size, mobility, work patterns, orientation, sun and noise — with the fitting done by a system rather than by a scarce architect’s scarce hours.
Mass production has always had this blind spot; I’ve made the parallel argument that the real use of new fabrication tech is serving the variance the median abandons. Housing is the extreme case, because the people social housing serves are precisely the people the median fits worst — larger families, older bodies, non-standard work, constrained mobility. The population with the most variance got the product with the least. That was a budget line, and the budget line just changed.
And there’s a second-order effect that might matter more than fit. Identical units are anonymous, and anonymity was part of how these places failed — nothing about your home was yours, and everything about the block announced that someone had processed you. Variation, even modest variation, is legibility: my window, my plan, my corner. Whether that measurably changes how people treat a place and each other is an empirical question — but the hypothesis that sameness contributed to failure is one most post-mortems of the era already accept.
The honest counterargument
“On paper” is doing real work in this argument, so let me put weight on the other side.
Social housing didn’t fail only because of design. It failed for reasons no generator touches: maintenance budgets that vanished after ribbon-cutting, concentration of poverty by allocation policy, stigma, isolation from transit and work, and management that treated residents as cases rather than tenants. Hand the mid-century planners perfect personalization technology and most of those failures happen anyway. Anyone selling AI as the fix for social housing is selling — and the pattern where a genuinely changed variable gets inflated into a universal solvent is exactly the judgment-eroding move worth resisting in every AI conversation.
Construction economics also push back. Personalization upstream in design is nearly free; personalization downstream — in structure and in services — still costs, because repetition is what makes sites fast and trades cheap. The honest near-term version is bounded variation: a standardized structural system and services core, with the generator exploring the space inside those bounds — layouts, orientations, unit mixes, envelopes. Bounded variation is exactly how form optimisation under fixed programs already works, and the bounds are where buildability lives. I’d defend the same structure-versus-flexibility split in my own work — it’s the logic behind designing the permanent layer conservatively and keeping every other layer changeable, applied at the scale of a housing program.
Verification is the actual bottleneck. Ten thousand distinct units are ten thousand chances for one to be wrong. The generator only earns trust if every variant it emits is automatically checked — code compliance, structure, egress, light, cost — so that variation never means risk. Building that checking system is harder than building the generator, and the leverage is downstream of it.
So the claim I’d defend is the narrow one: the strongest design objection to social housing rested on an economic constraint that no longer holds. The strong claim — AI solves social housing — is false and worth saying is false. What I’d want next is a pilot: one program, bounded variation on a standardized system, units generated against real household data, every variant machine-verified, and honest measurement of what the variation cost and what it changed for the people living there. Small and reversible, the way you’d trial any consequential change to a system you don’t fully understand.
The last era of social housing scaled a single answer; the tools now exist to scale the question. Which constraint in your own work is still priced as if design cost what it used to?