TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Ethics & Provenance · AGE 17—20 · FACTUAL · EVOLVING
The tilted mirror
The default house is placeless, and shipping it builds the tilt into a real building.
The idea
An image model's training data exerts a strong pull on what it can picture, and the pull is tilted: the open web photographs some places far more than others. The model is not a census of the internet, though; filtering, captions and post-training shape what comes out. In the Lab's own informal tests, a bare prompt for "a house" drifted toward North America and northern Europe. Rarely photographed places, many Indian regions among them, are rarely imagined. And "Indian" is not one look: a Malnad house, a Kutch bhunga and a Chettinad mansion share no visual essence. Fix it regionally.
Why it matters
The default is not neutral. It is a placeless default, and if you ship it, you build the tilt into a real building. Replacing it with a costume of jalis and arches is a second tilt, not a cure.
See it in the studio
Brief a guesthouse in Mysuru: warm, contextually grounded, bamboo and lime-wash, soft afternoon light. The output is lovely: pale wood, white walls, low ceilings. You could find it in any hotel chain and nowhere in Karnataka. The deep eaves, the jali, the high ceiling, the cross-ventilation — the whole reason the type exists — are gone.
Watch for this
Calling a placeless result "warm and modern." That phrase is how the tilt gets approved. And the opposite: calling a result "Indian" because it has a jali. Ask instead what the image has left out, and whether what it added belongs to this region.
Try it
Run a bare prompt for "a house" in English. Run it again in Kannada script — ಒಂದು ಮನೆ — and again in Romanised Kannada, ondu mane. Note where each result seems to be from, and whether the language of the prompt moved it at all. Then brief precisely for one named regional type — a Malnad house with deep eaves and a Mangalore-tile roof, or a Deccan wada round a courtyard — with named materials, the climatic moves, a reference architect. List what the precise brief had to restore.
Prove it
Show the default drift with a bare prompt and fix it with a precise regional one. Then explain why your own well-documented context is training data the world is missing — and why "Indian-looking" is not the same as right for the place.
How it works
Nobody designed the skew. Nobody decided that vernacular Karnataka should be rare in the data; the platforms that fed the scrape simply circulate some images a hundred times more than others, and every step after the scrape — filtering, captioning, post-training, the product's own guidance — can sharpen or soften the lean. The Lab's tests were informal: a handful of models, English prompts, no published protocol. A proper study — models, versions, dates, prompt language, seeds, sample size, a coding scheme — is planned for after this edition; until it is published, the sentence above is an observation, not a measurement. Published measurements exist for demographic bias in generated portraits. The specifically architectural version — vernacular dwellings from the Global South under-represented relative to Western ones — has not yet been measured in a peer-reviewed study that the Lab knows of, which is why this card says in our tests and is marked EVOLVING. Adding more data does not fix it: a bigger scrape of the same web carries the same lean at larger scale.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.