In this article＋
01 / ARTIFICIAL INTELLIGENCE
Remove the word “intelligence” for a moment
A camera judging burnt buns from past examples is machine learning. A system inventing a breakfast from the ingredients left in the fridge is generative AI.
One executes instructions, one learns a boundary, and one produces a new arrangement from a learned distribution.
The cha chaan teng example works because we see it every day: some actions follow a card, some rely on experience, and some need a fresh sentence on the spot.
'Iced lemon tea, less sweet' hard-coded is a rule. Predicting that a customer probably wants less sweet after seeing thousands of orders is machine learning. Reading 'not too sweet today' and drafting a reply is generative AI.
02 / ARTIFICIAL INTELLIGENCE
AI is a goal, not one component
Artificial intelligence names the broad ambition of making machines perform work associated with perception, language, reasoning or decisions. Machine learning is one family of methods used to pursue it.
Generative AI is a further family focused on producing new content. Saying a product uses AI is like saying a restaurant uses electricity: the fridge, lights and oven qualify, while their abilities are not interchangeable.
Calling all three 'AI' is like giving the cashier, waiter and cook one job title; when something breaks, nobody knows who owns it.
Businesses often skip the uncomfortable question: what does a mistake cost? A cash register should not guess generatively; a reviewed marketing draft can. Match the tool to the consequence.
03 / ARTIFICIAL INTELLIGENCE
The hidden layer is who defines success
The neglected question is who defined success before training. If a burnt-bun detector is labelled using customer returns, it may learn customer tolerance rather than colour.
People shape correct through labels, sampling, objectives and evaluation. Data is not reality itself.
It is a partial record of reality, created by a process with omissions and incentives.
Ask what a wrong step would hurt, then decide how much intelligence you actually need.
A generative model can produce human-looking text without having a human purpose, common sense or accountability. It searches for a plausible next step in pattern space; it does not privately fact-check the world.
04 / ARTIFICIAL INTELLIGENCE
Generating is not the same as understanding
A generative system can make a coherent menu without knowing that the oven is broken, ingredients are expired or a guest has an allergy. Fluency describes the surface of an output; it does not guarantee an updated model of the physical world.
Underneath the conversation is a powerful mechanism for extending patterns, not an observer who automatically checks every condition.
05 / ARTIFICIAL INTELLIGENCE
One product can sit in all three layers
One phone feature can belong to all three layers. Face grouping uses machine learning; a newly written memory caption uses generation; the complete organising experience may be marketed as AI.
These labels are nested, not competing boxes. Ask whether behaviour comes from rules written directly, a boundary learned from examples, or a new output sampled from a learned distribution.
06 / ARTIFICIAL INTELLIGENCE
The real boundary is how it fails
The most useful boundary is the failure mode. Traditional software often fails reproducibly.
Machine learning tends to fail around thin or shifted areas of its training distribution. Generative AI can produce a different plausible-looking error each time.
The advanced choice is not automatically best. Choose the failure pattern the surrounding process can detect, contain and afford.
The useful question after the label
When a product claims AI, ask what output is predicted, what examples shaped it, who defined success and how a mistake appears. These questions reveal more than a model name.
They also prevent a common category error: expecting generative flexibility from a fixed rule system, or deterministic guarantees from a generator. The hierarchy matters because each layer needs a different test.
Traditional software is checked against specifications, learned systems against representative data, and generators against evidence, variation and downstream harm.
07 / ARTIFICIAL INTELLIGENCE
The cha chaan teng example hides a useful trap: the order looks the same, but the decision mechanism is not.
'Iced lemon tea, less sweet' hard-coded is a rule. Predicting that a customer probably wants less sweet after seeing thousands of orders is machine learning. Reading 'not too sweet today' and drafting a reply is generative AI.
Businesses often skip the uncomfortable question: what does a mistake cost? A cash register should not guess generatively; a reviewed marketing draft can. Match the tool to the consequence.
08 / ARTIFICIAL INTELLIGENCE
Generating is not the same as understanding
A generative model can produce human-looking text without having a human purpose, common sense or accountability. It searches for a plausible next step in pattern space; it does not privately fact-check the world.
So 'sounds fluent' and 'knows what it means' must stay separate. Slightly disappointing, perhaps, but useful.
09 / ARTIFICIAL INTELLIGENCE
Questions people actually ask
Does machine learning have to generate content?
No. Classification, prediction and ranking are machine learning without generating an essay.
Must AI look human?
No. Useful AI often works quietly in detection, matching or forecasting.
What should a business do first?
Name a measurable, reversible problem, then decide whether rules, prediction or generation fits it.
Sources
Sources
Sources support the mechanisms and limitations discussed here. Models, products and prices change; check the official page and date when a detail matters.