Generative vs Parametric

Start with the word. A parameter is an adjustable input - a number you can turn up or down - and parametric urbanism is design through a model whose behaviour is governed by such inputs. The crucial move happens before you touch any knob: you *build the model*. You encode the relationships that make an urban fabric hang together - blocks are this size, streets this wide, buildings set back this far and rise to this height, density rises toward the transit stop, plots subdivide by this ratio. You are, in effect, writing down your design intent as a system of rules and dependencies. Only then do you tune it.

And tuning is where parametric earns its keep. Change the block-size parameter and every block re-forms; raise the density number and the massing thickens consistently across the whole plan; rotate the grid angle and the streets, plots and buildings all follow, coherently, in an instant. Nothing is placed by hand twice. What would have been days of redrawing becomes a slider you drag while the consequences ripple through the model in front of you. This is the genuine gift of parametric urbanism: it lets you *explore a family of designs you have specified*, testing "what if the blocks were bigger?" or "what if height stepped down toward the river?" and seeing the full, coordinated result rather than one sketch.

But be precise about the limit, because it is the whole point of the distinction. Parametric urbanism does not invent form. It cannot surprise you with a configuration you did not build the capacity for. The family of outcomes is exactly the family your model defines; the knobs only move you within it. If your model has no parameter for a diagonal desire-line across the grid, no amount of tuning will ever produce one. This is not a weakness - it is what makes parametric models *legible and controllable*, which for a public, contested artefact like a city is often exactly what you want. You can point to every rule and say why it is there. The danger is subtler: because you defined the family, it feels neutral and complete, and it is easy to forget that the model's silences - everything you did not give a parameter - are design decisions too, made by omission. A parametric model is a mirror of your assumptions, turned into a machine.

Generative urbanism inverts the relationship. Instead of you defining the form and tuning it, *algorithms propose the form itself*. You still set the terms - rules, constraints, goals, a starting seed - but you do not draw the result. The computer does, and it does so many times over, generating candidate after candidate and searching among them. Where parametric hands you a knob, generative hands you a population of proposals, most of which you never conceived.

The mechanisms vary, and the course meets them properly later. Procedural rules grow form the way an organism grows - a street network branches and subdivides by local rules, the way an L-system or a shape grammar unfolds a pattern step by step. Optimization searches for arrangements that best satisfy goals you set, breeding and mutating candidates the way a genetic algorithm does, keeping what scores well and discarding the rest. And increasingly AI proposes plausible urban fabric learned from vast quantities of existing city. What unites them is the essential act: the algorithm *produces and explores* many forms - hundreds, thousands - often surfacing configurations no human would have drawn, which a designer then evaluates and selects among.

The reason this matters is scale. A city has far too many interacting parts, and far too many possible arrangements, for a person to survey by hand. Generative methods let you *search a space too large to draw* - to sample widely across the possible and find candidates a hand would never reach. That is a genuinely new power, and it is why the field is exciting. But notice how the trap changes shape here. In parametric urbanism the risk was the model's silences. In generative urbanism the risk is the objective: the algorithm searches toward whatever goal you encoded, relentlessly, and it will happily find a form that scores brilliantly on your metric while being hostile to everything you forgot to measure. The more powerful the search, the more ruthlessly it exploits exactly the gap between what you optimized and what actually makes a city live. Generative methods do not decide what matters; they only find what best satisfies the mattering you specified - which is why the goals you set, and everything you left out of them, quietly become the city.

In real practice the two are rarely used alone; they interlock, and seeing how is what turns the distinction from a definition into a working tool. The clean way to hold it: a parametric model defines the moves, generative search explores them, and evaluation picks among the results. Each does the part it is suited to, and a human sits over all of it deciding what any of it is for.

Here is the loop in slow motion. First, you build a *parametric* model of the fabric - the adjustable system of blocks, streets, densities and heights that expresses your design intent and its dependencies. This defines the space of possible moves: the dimensions along which the design can vary. Second, you let a *generative* search run over those parameters - an optimization or exploration algorithm that turns the knobs for you, not one combination at a time but across thousands of combinations, sampling the space the parametric model opened. Third, you *evaluate*: analysis measures how each candidate performs (daylight, walkable reach, density, cost), optimization ranks the trade-offs, and - this is the part no algorithm does - a human judges which results are actually any good, which metrics were even the right ones, and whether the whole exercise is answering a real question or a convenient one.

This is why the parametric-generative pairing is so productive: the parametric model makes the design space *precise and controllable*, and the generative search makes it *traversable at scale*. You get both legibility and reach. But the combination also compounds the danger, and honesty demands naming it. A tight parametric model plus a powerful generative search plus a clean-looking optimization produces output with tremendous authority - charts, scores, a ranked frontier of options - and that authority is exactly what makes it easy to forget that every part of the machine inherited the assumptions and omissions you fed it. The search did not question the goal; the model did not question its own silences; the metric did not ask who it served. The machinery amplifies whatever you put in, including the mistakes. So the more sophisticated the pipeline, the more the human judgement at the end - and the democratic judgement beyond it - has to be defended, not delegated.

It would be tempting to treat the parametric-generative distinction as tidy vocabulary and move on. Do not. Blurring the two words does real damage - to how the field talks, how it is sold, and how the public is asked to trust it - and keeping them distinct is a piece of professional honesty, not pedantry.

Consider what the confusion hides. When someone presents a "computed" masterplan, the distinction tells you *who decided the form*. If the process was parametric, a human placed every move and the computer only kept the plan consistent as numbers changed - the authorship is human, the machine is a coordinator. If the process was generative, an algorithm proposed forms and a metric selected among them - the authorship is shared with a search process steered by whatever goal was encoded, and the crucial question becomes what that goal was and who chose it. Collapse the two and you cannot ask that question, which is precisely why marketing collapses them: "generative" sounds like objective discovery, so calling a merely parametric tool "generative" borrows a glamour of machine intelligence, while calling a goal-driven optimization "just parametric" hides the value-laden search inside it. Either slippage launders a claim.

There is a deeper reason too. The two carry the optimization trap differently, so a critique aimed at the wrong one misses. Parametric urbanism's failure mode is the *unquestioned model* - its family of outcomes silently excludes everything you gave no parameter, and it can feel complete when it is merely bounded. Generative urbanism's failure mode is the *unquestioned objective* - a powerful search will exploit any gap between your metric and the city's real life, the more ruthlessly the better it works. If you cannot name which method is in front of you, you cannot aim the right question at it: "what did your model leave out?" for the parametric, "what did you optimize for, and for whom?" for the generative. And in the Indian context, where a naive generative masterplan can literally not see the informal, organic city it was never given the data or categories to represent, getting this distinction right is not vocabulary - it is the difference between a tool that serves a just city and one that automates its erasure. Keep them distinct, and hold both to account. The binding choices - about the goals, the omissions and the plan itself - remain human, democratic and open to challenge, never settled by which word the tool wore.