What AI & ML actually are — for a designer
Strip away the hype and the jargon, and there's one simple idea underneath. Once you see it, every tool in this course makes sense.
It does not know what a wall is. And it just drew you forty of them.
That paradox sits at the heart of every AI tool you are about to use. The model that produces a breathtaking courtyard render has never stood in a courtyard, never felt the heat off a west wall, never read the National Building Code. It is, underneath the magic, a very large pattern-matching machine that has seen millions of images and learned what usually comes next. Understand that one sentence and you will use these tools with confidence instead of superstition — knowing exactly when to trust them and when to take the pencil back.
AI, ML, LLMs, diffusion — five words, one nested idea
They are not rivals. They sit inside each other.
The vocabulary sounds like competing products. It is actually a set of Russian dolls.
Artificial intelligence (AI) is the big outer doll — any machine doing something we would call 'intelligent'. Machine learning (ML) sits inside it: instead of a programmer writing rules, the machine learns patterns from examples. Inside ML sits deep learning — ML using many-layered 'neural networks', loosely inspired by the brain. And inside that sits generative AI: models that don't just classify or predict a number, but generate new things — text, images, plans.
The two generative families you will meet daily: large language models (LLMs) like ChatGPT and Claude, which generate text; and diffusion models like Midjourney, Stable Diffusion and FLUX, which generate images. Same core idea, different medium.
When someone says 'the AI', they almost always mean one specific model doing one specific thing. Ask which.
A prediction machine, trained on examples, that produces the most plausible next thing
Here is the model that unlocks everything. These systems were trained by being shown enormous amounts of data — text scraped from the web, or hundreds of millions of captioned images — and tuned until they got very, very good at one task: predicting what plausibly comes next.
An LLM, given 'the load-bearing wall should be at least', predicts the next word, then the next, producing fluent text. A diffusion model, given the words 'sunlit Kerala courtyard house, laterite walls', starts from visual noise and repeatedly nudges it toward an image that matches what such captions usually looked like in its training.
Neither is looking anything up. Neither has a fact-checker inside. They produce the most plausible output — which is usually useful, occasionally wrong, and never the same as true. That gap between plausible and true is the single most important thing a designer must hold in mind.
The mental model tells you, in advance, where the tool will shine and where it will lie
Because these are plausibility machines trained on the past, you can predict their behaviour:
They are brilliant at the visual and the typical — mood, atmosphere, style, a first draft of a layout, a fluent paragraph — because that is exactly what 'most plausible' produces. They are unreliable at the precise and the local — your city's setback rule, an exact dimension, a real product code, a structural calculation — because the plausible-sounding answer and the correct answer are two different things, and the model cannot tell them apart.
So a designer's instinct becomes: use AI to diverge, to visualise, to draft; use your training to converge, to verify, to decide. That division of labour is the whole course in one line.
Think of it as the difference between a sketch and a survey. AI gives you sketch-speed exploration across more options than you could ever draw by hand — but it carries no warranty on FAR, structure or compliance. You remain the registered professional who signs the drawing. The model is a tool in your hand, exactly like a pencil or Revit; the authorship, and the liability, stay yours.
For interiors this is freeing: most of your early work _is_ mood, style, palette and 'what if we tried…' — precisely where a plausibility machine excels. You can show a client six directions for a living room before lunch. Just remember the model invents products, finishes and prices that look real but aren't. Treat its images as concept art to align taste, then specify from real, available, costed products.
If you are a student or a one-person studio, this is the great leveller — capabilities that used to need a viz team or a big office are now a subscription. The catch is that the tool amplifies judgement; it doesn't supply it. A weak designer with AI makes plausible-looking mistakes faster. Spend your learning on the fundamentals this platform teaches — light, proportion, materials, code — and let AI accelerate a foundation you actually have.
ChatGPT, Claude, Gemini
Large language models (text)
Generate and reshape text: briefs, specs, emails, summaries. Claude holds very long documents; ChatGPT is strong at tables. All can sound confident and be wrong.
Midjourney, Stable Diffusion, FLUX, Firefly
Diffusion models (images)
Generate images from text or other images. Midjourney for aesthetics, FLUX for realism + editing, Firefly for commercially-safe training data, Stable Diffusion for open/local control.
Veras, Maket, InteriorAI, Forma
Applied tools built on the above
Most 'design AI' products are a friendly wrapper around these engines, tuned for a task — rendering, floor plans, staging, site analysis. Knowing the engine underneath demystifies the product.
“AI understands architecture — it just made a perfect render, so it clearly knows what it's doing.”
A convincing image is evidence of pattern-matching, not understanding. The model learned what 'a beautiful modern villa' tends to look like; it has no concept of gravity, climate, cost or code. It can render a cantilever that could never stand and a kitchen no one could cook in, with equal confidence. Beauty in the output is not correctness in the building.
Workshop — make the 'plausible, not true' gap visible
The fastest way to internalise this whole course is to watch a model be confidently wrong, once, with your own eyes. Fifteen minutes, free tools, no setup. Do it before you read another lesson.
Free: any chat AI (ChatGPT / Claude / Gemini) + any image AI.
Prompt 1 (to a chat AI): "What is the exact front setback, in metres, required for a 30x40 ft residential plot in [YOUR CITY], and cite the specific bye-law clause." Prompt 2 (to an image AI): "modern 3-storey Indian villa, structurally accurate, flat roof, large west-facing glass facade, photorealistic"
- 1Run Prompt 1. Note three things: the number it gives, how confident it sounds, and the clause it cites.
- 2Open your city's actual building bye-laws (or your municipal portal) and check that number and clause. Mark it: right, close, or invented.
- 3Run Prompt 2. Zoom in and hunt for what couldn't be built: floating beams, columns that miss the floor below, a cantilever with nothing holding it, doors to nowhere. Count them.
- 4On that west-facing glass: ask yourself what it would do to the room's temperature in May. The model never considered it.
- 5Write two sentences in your notes — one on what each tool was genuinely good at, one on exactly where plausible parted company from true. That instinct is what you're building for the next 39 lessons.
You’ll walk away with
A concrete, personal example of an AI being fluent, confident and wrong — the mental anchor for every 'verify this' decision you'll make for the rest of the course.
Two more quick experiments, if you have five minutes.
- 01Ask the same chat AI for a real IS code number for, say, 'concrete mix design' — then check whether the code it names actually exists and covers that.
- 02Ask an image AI for the same villa twice with the identical prompt. Notice it gives two different buildings — there is no single 'right answer' inside it, only plausible ones.
AI tools are pattern machines that produce the most plausible next thing from what they were trained on. That makes them superb partners for diverging, visualising and drafting — and unreliable witnesses for anything precise, local or true. Every later lesson is just this principle, applied to a specific task.
AI contains ML contains deep learning contains generative AI. LLMs make text, diffusion models make images, and most design products wrap these engines. They predict the plausible, not the true — so diverge with AI, converge with your own judgement.
Do I need to know maths or coding to use AI as a designer?
No. To use the tools in this course you need design judgement and clear language, not calculus or programming. Understanding the concepts — what a model is, why it hallucinates — makes you far better at directing the tools, but you will never write a line of code to render a courtyard or draft a spec.
What's the difference between ChatGPT and Midjourney?
They are different families. ChatGPT is a large language model — it generates and manipulates text (briefs, emails, specifications). Midjourney is a diffusion model — it generates images from descriptions. You'll use language models for words and diffusion models for pictures, often together in the same project.
Is 'AI' going to replace architects and interior designers?
It is replacing specific tasks — first-pass visualisation, drafting boilerplate text, generating layout options — not the role. Design is judgement, responsibility, client relationship and accountability under your professional registration, none of which a plausibility machine can hold. The designers at risk are the ones who do only the automatable tasks; the ones who direct the tools become far more productive.
Now that you know what these tools are, the obvious next question is the practical one: where exactly is the line between what they can and can't be trusted to do?