A startup needs a spam filter, and you are the training data. Label 10 emails — spam or inbox. Your model learns only from your labels: no second chances, no corrections. Then it goes live on real email... and the real world has surprises for it. How accurate will YOUR model be?
This is exactly how supervised learning works: humans label examples, the model learns the mapping, and every labeling mistake becomes a model mistake. Module 02 explains the theory — here you'll feel it.
No feedback yet — the model trusts you completely. Every label you give becomes its truth.
Your model never saw the emails — it only saw your labels. Every mistake you made during training became a rule it faithfully learned. In real ML, labeling quality is destiny.
The model learned "free = spam" as a pattern, so HR's free-lunch email got flagged. It matched the feature, not the meaning. That's a false positive — and the seed of overfitting.
The polite phishing email sailed through because nothing like it existed in your 10 examples. Models fail quietly on inputs unlike anything they trained on — which is why coverage matters more than volume.
Supervised learning is the most commercially important form of machine learning, and it is remarkably simple to describe: humans label examples, and a model learns the mapping from input to label. Spam filters, fraud detection, medical triage, and content moderation all work this way. This game removes the maths and leaves the mechanism, so you can feel why real machine learning projects succeed or fail.
During the training phase you label ten emails as spam or inbox, and you get no feedback whatsoever. That silence is deliberate and accurate: a model has no independent access to the truth. It cannot check your work. Whatever you assert becomes its ground truth.
The consequence shows up immediately in the live phase. Label the “FREE iPhone” scam as legitimate, and the model waves through the near-identical “FREE cruise” scam a minute later — and the game traces that failure straight back to your label. This is the practical meaning of the phrase garbage in, garbage out. In industry, labelling quality is usually the single largest determinant of model quality, which is why serious teams spend far more on annotation guidelines and reviewer agreement than on model architecture.
Every spam email in the training set contains the word FREE, so the model learns a rule: FREE means spam. Then HR announces free lunch in the break room on Friday, and the filter quarantines it.
The model is not being stupid. It is doing exactly what it was built to do: match statistical patterns in the features it was given. It has no concept of a scam, a colleague, or an office. It matched a surface signal that happened to correlate with the label in training and does not generalise outside it. This is a false positive, and it is the visible face of overfitting — learning the incidental characteristics of your training set rather than the underlying concept.
Real systems fail this way constantly, and the failures are rarely as legible as a word. A model that learns to detect pneumonia partly from which hospital’s scanner produced the image has learned something real about its training data and nothing useful about pneumonia.
The second edge case is a polite, well-written phishing email: someone claiming to be from IT, asking you to share your login so they can migrate your mailbox tonight. No capital letters, no money, no suspicious links, no urgency. Nothing in your ten training examples looked remotely like it, so the model lets it through without hesitation.
This is a false negative, and it is the more dangerous failure mode, because nothing signals it. A false positive is annoying and visible: someone complains that HR’s email went to spam. A false negative is silent — the attack simply arrives. Models do not degrade gracefully on inputs unlike anything they trained on; they produce a confident answer drawn from irrelevant patterns.
This is why coverage matters more than volume. Ten thousand examples of the same three kinds of spam teach a model less than two hundred examples spanning twenty kinds. Real spam filters are beaten precisely at these coverage gaps, which is also why attackers keep inventing new formats.
If you evaluated this model on the same ten emails it learned from, it would score 100% every time and tell you nothing. The model has already seen the answers. Genuine evaluation requires data the model has never encountered, which is why practitioners hold out a test set and never train on it. A model that scores brilliantly on training data and poorly on test data is overfitting; the gap between the two numbers is one of the most informative diagnostics in machine learning.
The concepts here — labels, generalisation, overfitting, false positives and negatives, coverage — are the foundation of everything in applied AI. Module 02: How Machines Learn covers them with no code required. The Data Mindset goes further into datasets, distributions, missing values, and feature engineering, and Model Evaluation covers precision, recall, F1, and how to measure whether a model is actually any good.