Generative Design & AI

Start with what AI actually does, because the mechanism explains both the power and the risk. A generative AI model is not given rules; it is trained on data - large collections of existing plans, maps, satellite imagery, street networks, renderings - and it learns the statistical patterns in that data: what tends to go with what, how streets usually meet blocks, what a plausible density gradient looks like. Having learned those patterns, it can generate new output by inference - producing fabric that is statistically similar to its training set without copying any single example, and without anyone having written a rule that says 'streets branch here'. Crucially, no explicit logic is authored, and often no one can say exactly *why* the model produced a particular result.

This is a genuinely powerful shift over rule-based procedural methods. Explicit rules can only capture what a human can articulate; a great deal of what makes urban fabric feel right - subtle relationships of grain, rhythm, connectivity and mixture - is tacit, hard to write down, and easy for an experienced eye to recognise but not to codify. A model that learns from thousands of real examples can absorb some of that tacit pattern and reproduce it, generating fabric that a rule-writer would struggle to specify. For rapidly producing precedent-informed options to explore, and for surfacing possibilities grounded in how cities have actually been built, AI is a real extension of the generative toolkit.

But the same mechanism plants the risks. Because the model learns *the past*, it inherits whatever the past contained - including its inequities and its blind spots. Because it works by statistical resemblance, it produces a confident *surface* without any *understanding* of what a city is for, whether its output is buildable, or whom it serves. And because no rules are written and no clear reasoning is available, its output is opaque - you cannot inspect the logic, only judge the result, which makes its errors far harder to catch than a rule-based generator's. So the very feature that makes AI powerful - learning tacit pattern from data rather than from rules - is exactly what makes it dangerous: it generates plausibility, learned from the past, with no understanding and no visible reasoning. Everything else in this lesson follows from that.

The single most important thing to understand about AI-generated urban fabric is that plausibility is not goodness, and the gap between them is where the danger lives. An AI model is, in effect, a machine for producing output that *looks* right - confident, polished, detailed, resembling real precedent, persuasive to a client or a public. That is what it was trained to do: generate statistically plausible fabric. But looking right and being right are different things, and the model has no grip on the difference. It can produce a masterplan that reads beautifully and is undeliverable, that renders convincingly and would displace the people who live there, that resembles a good district and is blind to this actual site, this slope, this community, this history.

Worse, generative models can confidently invent - the phenomenon often called hallucination in language models has a direct urban analogue: fabric, connections, or features that are plausible in appearance and simply fabricated, unsupported by the site or the brief, presented with exactly the same confidence as everything else. There is no flag on the output that says 'this part is invented' or 'this would not stand up'. The surface is uniformly convincing, which is precisely the trap: a confident wrong answer is far harder to catch than an obvious one. A crude generator that produces obvious nonsense is safe, because you discard it. A sophisticated AI that produces beautiful, expert-looking, subtly wrong fabric is dangerous, because you trust it.

This connects straight back to the previous lesson's warning about games and film, but sharpened. Procedural generation made a plausible surface; AI makes a *more* plausible surface, good enough to fool an expert, which raises the stakes rather than lowering them. The discipline is a kind of trained skepticism: treat every AI-generated urban proposal as a plausible hypothesis to be verified, never as a finding. Ask what it cannot know - the real site, the ownership, the community, the buildability, the justice of it - and check the output against all of it with human judgement and real analysis. The polish is not evidence; if anything, the more convincing the output, the more deliberately you must probe it, because the whole risk of AI in urbanism is that it is so very good at looking good.

The second severe risk is bias in the training data, and it is not a peripheral glitch but a structural consequence of how these models work. An AI model trained on the past learns the past - and the past of urban development is not a neutral record. It reflects whose city was built, mapped, recorded and valued, and whose was not. If a model is trained mostly on the formal, titled, well-documented city - the plans that got drawn, the neighbourhoods that got mapped, the developments that had power and money behind them - then it learns *that* as the template for a city, and it will tend to reproduce and normalise those patterns, including the inequities baked into them, treating the way things have been as the way things should be.

In the Indian context this is acute and concrete. A very large share of the city - the informal settlements, the unmapped organic quarters, the self-built fabric that houses hundreds of millions - is under-represented or simply missing from the data that models train on. A generative model cannot produce what it never saw: if the informal city is absent from its training, the model will generate the formal city as the default, and the informal fabric becomes invisible - not designed for, not seen, quietly erased by omission. The model does not decide to exclude it; it never learned it existed. That is bias by absence, and it is arguably more dangerous than obvious prejudice because it hides as neutrality.

The deeper problem is that AI launders this bias in the objectivity of the algorithm. A human planner who ignored the informal city could be challenged; 'the AI generated this' sounds neutral, technical, above politics - so the same erasure, produced by a model, is harder to contest and easier to wave through. What was a contestable political choice about whose city matters becomes an apparently objective output. This is the equity-and-power theme of the whole course, arriving through AI: what the model learned, from whose data, and whom it therefore serves, is never neutral, and the veneer of algorithmic objectivity makes the bias more dangerous, not less. The competent response is to interrogate the training data - whose city is in it, whose is missing - to refuse to treat AI output as objective, and to actively defend the informal and organic city the model cannot see, precisely because it cannot see it.

How, then, should an urbanist use AI in generative design without being captured by it? The honest position is neither refusal nor surrender, but disciplined, skeptical use. AI is genuinely valuable for exploration and ideation - generating a wide, precedent-rich field of possibilities fast, breaking fixation, and surfacing options grounded in how cities have actually been built. It can be a superb thinking partner for the early, divergent phase of design, where you *want* many provocations and expect to discard most. Treated as a generator of hypotheses to reason about, it extends what a studio can explore.

The discipline is everything that comes after generation. Verify, never trust: treat every AI output as a plausible claim to be checked against the real site, real buildability, real ownership and real community, with human judgement and genuine analysis - because the output's polish is not evidence and its confident inventions look exactly like its sound parts. Interrogate the data: ask whose city the model learned from and whose is missing, and assume the informal and organic city is under-represented, especially in India. Refuse the objectivity: never let 'the AI generated it' launder a contestable choice as a neutral technical output. And keep a human in the loop who is accountable - the responsibility for what gets proposed and built cannot be delegated to a model that has no stake, no understanding and no accountability.

The boundary is the same one the whole course insists on, made sharper by AI's persuasiveness. Generative AI helps you *explore and think*; it does not decide. The actual planning and land-use decisions, the statutory approvals, and the social, equity and political judgements about a city's future - including the decisive questions of whom the plan serves and who might be displaced - belong to the planning authority, the democratic and participatory process, the affected communities, and the governing planning law and development-control regulations, in India the master-plan and development-plan process, the applicable DCR and the National Building Code of India. AI raises the stakes because it makes wrong answers beautiful; the defence is to keep human judgement, real verification and democratic decision firmly in charge of a tool that is powerful precisely at producing what merely looks right.