Learning models¶

Supervised learning¶

Supervised learning models or algorithms is like teaching a child with flashcards—except the child is an algorithm, and the flashcards are thousands of labelled data points. You show it pictures of cats and dogs, tell it which is which, and hope it doesn’t start calling all four-legged creatures “probably a moggy.” The algorithm’s job is to find patterns (or invent them) that map inputs (like images) to outputs (labels). It’s brilliant at memorising but can throw a wobbly when it sees something new, like a Sphynx cat (“Is this a dog? A rat? A tiny alien?”).

Real-life¶

Your email spam filter is a classic case. It’s trained on millions of emails marked as “spam” or “not spam,” learning that words like “free,” “urgent,” and “Nigerian prince” are red flags. But sometimes it gets overzealous—like when it sends your boss’s important email to the junk folder because it contained the phrase “meet me tomorrow” (clearly a phishing attempt). Meanwhile, actual spam about “enlarging your [REDACTED]” slips through because the algorithm decided it was “probably a medical newsletter.”

Security & privacy risks (moderate)¶

If the training data contains sensitive info (e.g., medical records), a malicious actor could reverse-engineer the model to extract personal details. For example, a health insurance predictor might inadvertently leak that “people aged 30–40 with a history of buying gluten-free pasta are high-risk.” Models can also be poisoned—imagine a spam filter fed fake “not spam” labels for scam emails, turning it into a spam butler.

Unsupervised learning¶

Unsupervised learning models or algorithms is the AI equivalent of chucking a bunch of puzzle pieces into a room and telling the algorithm, “Sort that out.” No labels, no guidance—just raw data and existential dread. The algorithm might group similar things together (clustering), shrink data into simpler forms (dimensionality reduction), or flag odd outliers (anomaly detection). It’s like a detective solving a crime without knowing what a crime is—just “Hmm, this bloodstain is statistically interesting.”

Real-life¶

Netflix’s recommendation system uses unsupervised learning to group users into “taste clusters.” You might be in the “watches true crime while eating cereal at 3 AM” cohort, which is why Netflix keeps suggesting Making a Murderer and ads for melatonin. Retailers also use it for fraud detection—like when your credit card gets blocked for buying a single banana in Spain (clearly, only a criminal would do that).

Security & privacy risks (moderate)¶

Clustering can accidentally reveal sensitive groupings—like an algorithm noticing that “people who buy pregnancy tests also buy unscented lotion” (hello, Target’s infamous pregnancy prediction mess). Anomaly detection can also backfire; imagine a system flagging donations to LGBTQ+ support groups as “suspicious activity” because they’re statistically rare.