What Machine Learning Is (and Isn't)
P16.ml-literacy.01 · Audience: guest, it-ml, language-pro
You do not need the maths to lead a machine-learning project — but you do need an accurate picture of what a model actually hands back, and where it quietly fails. This module opens the ML Literacy track with four ideas that decide whether a leader is easy or hard to mislead: what a model produces, the two learning styles, why "it worked on our data" is a trap, and why a highly accurate model can still be worthless.
One mental model. A machine-learning model is a pattern-matcher that learned from past data to guess about new data. It produces predictions, not understanding — so it can be fluent, useful, and confidently wrong all at once.
Machine learning finds patterns in past data and uses them to make predictions on new data it has not seen. What it produces is a prediction or a score — not understanding, not reasons, not guarantees.
Worked example — read a churn model's actual output. A churn model hands back one row per customer. Here is what it does and does not say:
| The model outputs | What that means | What it does NOT give you |
|---|---|---|
| Customer 4471 → churn risk 0.82 | "82 of 100 similar past customers left" | why this customer might leave, or any guarantee |
The 0.82 is pattern-matching at scale, not a reason. And because the model only knows the world it was trained on, it will be confidently wrong when the world changes — a new competitor, a new pricing rule — while sounding just as sure as before.
🗣️ From a linguist's perspective: A model as a translator who has never left the country
⚡ Interview Ref — the quick-scan Reference face
In plain language: what a model really produces, the two main learning styles with business examples, why "it worked on our data" is a trap, and why high accuracy can still be worthless. No maths required.
1 — What ML is, and what a model actually produces (b16)
Machine learning finds patterns in past data and uses them to make predictions on new data it hasn't seen. That's the whole trick: learn from examples, then guess about the future. What a model actually produces is a prediction or a score — "this customer is 82% likely to churn", "this invoice looks like fraud", "this house should sell for €340k". It does not produce understanding of your business, reasons or a real explanation of why, or guarantees that it's right. It's pattern-matching at scale, not human-style reasoning — and because it only knows the world it was trained on, it will be confidently wrong when the world changes while still sounding just as sure as before.
2 — Supervised vs unsupervised, with business examples (b17)
The difference is simply whether you already have the answers to learn from. Supervised learning learns from labelled examples to predict a known target — past customers tagged churned / stayed, so it predicts churn yes/no for current customers (predicting a price works the same way). Unsupervised learning finds structure without labels — it sifts your customer base into natural segments, or spots anomalies, without ever being told what a segment or a fraud is. Rule of thumb: if you can point to a column of right answers, it's supervised; if you're exploring to discover groups or oddities, it's unsupervised.
3 — Train/test split; why "it worked on our data" fails (b18)
A model is only trustworthy if it does well on data it didn't learn from, so hold back a test set the model never sees during training and judge it only on that. Scoring a model on its training data flatters it: it can simply memorise those exact cases and look near-perfect, the way a student who saw the exam answers scores 100% without understanding anything. So "it worked on our data" often just means it overfit. One more trap to name in the room: data leakage — when a clue about the answer sneaks into the training data (a field only filled in after a customer has already churned). It makes the model look brilliant in testing and then collapse in production.
4 — Why a highly accurate model can be useless (b19)
A single big accuracy number is one of the most misleading things in this whole field — a model can be 99% accurate and completely useless. The classic case is imbalanced classes: if only 1% of transactions are fraud, a model that never catches any fraud is 99% accurate and worth nothing. What actually matters is the cost of each kind of error — missing real fraud usually hurts far more than a false alarm — and whether the model helps the actual decision. Ask about precision and recall and, above all, the business value.
Interview one-liners
- A model produces predictions, not understanding, reasons, or guarantees — pattern-matching at scale that's confidently wrong when the world changes.
- Supervised learns from labelled answers (churn, price); unsupervised finds structure with no labels (segments, anomalies).
- Judge on a held-out test set, never the training data — "it worked on our data" usually means it overfit or leaked.
- High accuracy can be useless on imbalanced data; ask about precision, recall, error cost, and business value.
📚 Go Further
Plain-language explainers on what machine learning can and can't do for a business.
| Type | Resource |
|---|---|
| Book | The Hundred-Page Machine Learning Book — Andriy Burkov (skim the intro chapters) |
| Read | Google's Rules of Machine Learning — practical, non-mathematical guidance |
| In-app | Working with ML: Trust, Questions & Project Value — turning this literacy into decisions |
Ask the mentor about this module
Ask a question about this content. The mentor explains and grounds its answer in what you are studying; asking is recorded as a learning signal, not a grade.
Where next?
Later in ML Literacy for Leaders