They ask: Can you explain supervised learning?
You say: Supervised learning is training a model on data where somebody has already written down the right answer. Thousands of images, each one tagged cat or dog by a human. The model looks at the examples and learns the pattern that separates them. It is exactly what school is. A person who already knows the answer stands next to you while you learn it, and then tests you.
The tag is called a label, and the labelling is done by people, by hand, before any training starts.
The answer that ends the conversation early: you say "it is when you train with labelled data" and stop. That is the definition repeated back. Everyone in the pile says that. Say what a label is, who makes it, and what it costs, and you are the only one who has thought about where the data comes from.
They ask: What makes a training dataset good?
You say: Two things. Correct labels and enough variety. If I want a model that recognises cats, I cannot show it one cat a thousand times. I need different sizes, different colours, different lighting, different angles, different backgrounds. Otherwise it learns the background instead of the animal.
And every one of those labels was checked by a person. That is slow, it is expensive, and it is the part nobody talks about. When you hear that data is valuable, this is why. Not raw data. Verified data.
The answer that ends the conversation early: "more data is better." More of the same data is not better. It makes the model more confident about the narrow thing it already learned, which is worse than useless, because now it is wrong with conviction.
They ask: Why is machine learning changing knowledge work specifically?
You say: Because of what a trained model is. Think about a doctor reading X-rays. Roughly forty years to make that person, and what makes them good is that they have seen hundreds of thousands of images and built an instinct for the pattern.
A model trained on a million labelled scans has that pattern too. It did not need the forty years. And once it exists, copying it costs almost nothing, so it can go anywhere.
That is the shift. It is not that machines got smart. It is that pattern recognition built over a career stopped being scarce.
The answer that ends the conversation early: "AI is going to replace doctors." It is not, and saying it makes you sound like you read a headline. The instinct got cheap. Deciding what to do about what the scan shows did not.
They ask: What is unsupervised learning?
You say: It is what you do when you have no labels. You give the model the data with no answers attached and ask it to find structure. It groups things that resemble each other. Those groups are called clusters, and nobody tells the model what they mean.
Watch a four year old with a bin of toys. Tell them to tidy up, give them no categories, and you come back to blocks in one pile, stuffed animals in another, cars in a third. Nobody defined those groups. The child saw that things resembled each other.
That is useful when you do not know what you are looking for. Customer behaviour, fraud patterns, groups in data you have never examined.
The answer that ends the conversation early: treating it as a weaker version of supervised learning. It answers a different question. Supervised learning asks which of these known things is it. Unsupervised asks what groups are in here at all.