What AI can and can't do in design
The honest map of the territory — drawn before you spend a rupee on a subscription — so you aim these tools where they pay and never where they bite.
A ₹4 crore mistake that started with a beautiful, confident, wrong picture.
Imagine presenting an AI render to a client who falls in love with it — a soaring double-height glass facade facing due west. They sign off. Only at working-drawings stage does someone run the numbers: that west glass turns the living room into an oven for six months a year, and the cantilever holding the upper floor needs a beam no one budgeted for. The picture was gorgeous. It was also structurally and environmentally illiterate — and the model said so with total confidence. The skill this lesson builds is reading that line, every time, before it costs you.
Two columns: the green list and the red list
Where AI genuinely earns its keep
Map these to the plausibility-machine idea from the last lesson and the pattern is obvious. AI is strong wherever 'plausible and fast' beats 'precise and slow':
Divergent ideation — twenty facade directions, ten palettes, a dozen plan parti diagrams in the time it took to make one. Visualisation — turning a grey SketchUp model or a napkin sketch into an atmospheric render to align a client. Mood and style — boards, references, 'show me Scandinavian-meets-Chettinad'. Drafting language — first drafts of briefs, specs, emails, meeting notes, scope sections. Summarising and searching — digesting a 90-page tender, pulling themes from client transcripts. Repetitive transformation — recolouring, restyling, removing clutter, upscaling.
What unites the green list: in every case a knowledgeable human reviews the output before it matters, and an error is cheap and visible.
Where it will fail you — quietly and confidently
The red list is everything where plausible and true diverge, and where being wrong is expensive or dangerous:
Code and compliance — setbacks, FAR, fire norms, NBC, local bye-laws. Models hallucinate regulations fluently. Exact dimensions and quantities — an AI floor plan's measurements and a buildable drawing's are not the same thing. Structure and physics — it will draw cantilevers, spans and junctions that cannot stand. Facts, products and prices — it invents plausible product names, IS codes, vendors and rates. Citations — it fabricates references that look real. Judgement and responsibility — what the client actually needs, what's appropriate, what you'll sign your name to.
The danger is not that AI fails here. It is that it fails confidently, in fluent prose and beautiful images, with no error message. There is no red underline under a hallucinated setback.
If being wrong here would embarrass you, cost money, or endanger someone — verify it outside the model. Always.
Treat AI as a brilliant, fast, tireless intern — never as the architect of record
The most useful single frame in this whole course: AI is the best intern you have ever had. Astonishingly fast, endlessly patient, well-read, great at first drafts — and absolutely not someone you would let seal a drawing, sign off compliance, or talk to a client unsupervised.
You would never hand an intern's first-day work straight to a client without reading it. You would never let them certify a structure. But you would absolutely use them to explore options, draft the boring paragraphs, and do in an hour what would have taken you a day. Apply exactly that supervision to AI. The human stays in the loop — reviewing, correcting, deciding, owning the result. The moment the human leaves the loop is the moment the ₹4-crore picture gets signed.
Your registration is the firewall. Use AI freely through concept, options and visualisation, then route everything that touches compliance, structure or dimensions through your normal professional process — calculations, checked drawings, code review. A practical rule: nothing an AI produced goes into a sanction set, a structural document or a contract without a human professional re-deriving it. The render persuades; the stamped drawing builds.
Your red list is shorter but real: AI invents furniture that isn't for sale, quotes fabric and stone that doesn't exist at that price, and ignores real clearances, ergonomics and services. Use it to sell a _direction_ — mood, colour, spatial feel — then build the actual scheme from a real, sourced, costed FF&E schedule. Never let a client believe a generated sofa is a product you can buy next week; set that expectation early and you'll never be caught out.
Without a senior to catch you, you _are_ the loop — so build the verification habit deliberately. Keep a personal 'red list' taped to your monitor: code, dimensions, structure, products, prices, citations. Every time AI gives you something on that list, the next action is to confirm it from a real source. Slower at first; career-saving by the second month. The students who get burned are the ones who mistook fluent for correct.
AI floor-plan generators (Maket, ARCHITEChTURES)
Great for options, not for sanction
Genuinely useful for rapid layout options and feasibility. But the dimensions, areas and code-compliance need a professional's check — most don't know NBC or your local bye-laws.
Code-checking AI (UpCodes Copilot, VitruAI)
Assistive, not authoritative
Speeds up cross-referencing, mostly US/intl codes. Treat its answers as a first pass to verify against the actual gazetted regulation, never as the ruling itself.
LLMs for 'facts' (ChatGPT, Claude, Gemini)
Confident fabricators
Will invent IS codes, product specs, vendor names and citations that look completely real. Anything load-bearing — literally or legally — gets verified from the primary source.
“Newer, more powerful models will soon fix the accuracy problem — hallucination is just a temporary bug.”
Models do get better, and error rates fall — but generating the plausible is what these systems fundamentally do, not a flaw to be patched out. Even a near-perfect model gives you no guarantee, and you cannot tell a 99%-right answer from a 100%-right one by looking. For anything where being wrong is costly, verification outside the model stays mandatory no matter how good the model gets.
Workshop — build your own green/red audit on a real task
Don't take the two lists on trust — generate them from your own work. You'll run one real task through an AI and colour-code every line of the output, turning the green-list/red-list idea into a habit you can apply to anything.
Free: any chat AI. Bring one live task from a current project.
Pick ONE real task and prompt the AI with it, e.g.: "Write the scope-of-work section for a 3BHK apartment interior fit-out in [CITY], including materials, finishes and an indicative budget per item." Then, line by line, tag the output: [G] green = mood / draft / structure I can trust [R] red = a dimension, code, product, price, or fact I must verify
- 1Run your task and paste the full output into a notes doc.
- 2Go line by line and tag each one [G] or [R]. Be strict: every number, product name, price and code is [R] until proven otherwise.
- 3Pick the two most load-bearing [R] items — the ones a client would act on — and verify them against a real source (a vendor site, the actual code, a real price list). Were they right, close, or invented?
- 4Count your tags. The ratio of G to R on this kind of task tells you how much supervision it needs every time.
- 5Write your keep-line: one sentence naming what you'll let AI own for this task going forward, and what you'll always verify yourself. Tape it where you work.
You’ll walk away with
A colour-coded audit of a real deliverable plus a written 'keep-line' for that task type — a reusable rule for exactly how far to trust AI on the work you actually do.
A quick stress-test, if you have five minutes.
- 01Ask an AI to 'list the relevant IS codes and fire-safety norms' for your project type, then check how many of the codes it cites are real and current.
- 02Ask it for the price of a specific named product (a tile, a faucet). Then find the real price. Note the gap — and that it never flagged any uncertainty.
AI is a brilliant intern: trust it to diverge, visualise, draft and summarise, where a human reviews before it matters and errors are cheap. Distrust it on code, dimensions, structure, facts and judgement, where it fails confidently and errors are expensive. Keep a human in the loop, and the authorship — and the liability — stays yours.
Green list: ideation, visualisation, mood, drafting, summarising, repetitive edits. Red list: code, dimensions, structure, facts, products, prices, citations, judgement. The rule: AI is the intern, never the architect of record. Never let the human leave the loop where being wrong is costly.
Can AI design an entire building or interior on its own?
It can generate something that looks like a complete design — plans, renders, even a spec — but not a buildable, compliant, responsible one without a professional. The output skips structure, code, climate, real products, cost and the client's actual needs. Used as a starting point under expert direction it's powerful; used as a finished answer it's a liability.
Why does AI 'hallucinate' facts and codes?
Because it generates the most plausible-sounding text, not retrieved facts. A made-up setback of '3 metres' sounds exactly as confident as a correct one, because the model is matching patterns of how such sentences usually read — not consulting your city's bye-laws. It has no internal sense of true versus plausible, so it can't warn you when it's wrong.
So is AI even worth it for serious practice?
Yes — enormously, if you aim it correctly. Indian firms report meaningful time and cost savings using AI on the green list (ideation, visualisation, drafting). The value is real; the discipline is knowing the red list cold so a confident, wrong output never reaches a drawing, a client or a site unchecked.
You know what the tools are and where they're trustworthy. The last piece of groundwork is the map — where, across a real project from first sketch to site, each kind of AI actually belongs.