2. What is AI?

Most of what is sold as AI is a function that maps inputs to outputs, fitted on examples. That is less magical and more useful than the marketing.

Learning objectives

Use “AI”, “machine learning”, and “deep learning” correctly.
Distinguish supervised, unsupervised, and generative tasks.
Explain why a model that fits the training data can still be worthless.
State what a model cannot tell you, no matter how accurate it is.

The nesting dolls

These three words are not synonyms.

┌─ Artificial Intelligence ──────────────────────┐
│  any system that performs tasks we'd call      │
│  "intelligent" — including hand-written rules  │
│                                                │
│  ┌─ Machine Learning ───────────────────────┐  │
│  │  systems that learn patterns from data   │  │
│  │  rather than being told the rules        │  │
│  │                                          │  │
│  │  ┌─ Deep Learning ───────────────────┐   │  │
│  │  │  ML using many-layered neural     │   │  │
│  │  │  networks. Needs a lot of data.   │   │  │
│  │  └───────────────────────────────────┘   │  │
│  └──────────────────────────────────────────┘  │
└────────────────────────────────────────────────┘

BLAST is arguably AI — it performs an intelligent task using hand-written rules. It is definitely not machine learning: nobody trained it on examples.

Say what you mean

If you fitted a logistic regression, say “logistic regression”, not “we applied AI”. Reviewers notice, and precision costs you nothing.

The three kinds of task

Supervised — you have labelled examples and want to predict the label for new cases. This sequence is genotype A; that one is genotype D; what is this new one? Requires labels, and your model inherits every bias in them.

Unsupervised — you have no labels and want structure. Do these 338 samples fall into groups? Nothing guarantees the groups mean anything biological.

Generative — you want the model to produce new content: a protein structure, a plausible sequence, a paragraph of text. Large language models live here.

	Needs labels?	Typical question	Failure mode
Supervised	Yes	“Which class?”	Learns the label’s bias
Unsupervised	No	“Is there structure?”	Finds structure in noise
Generative	Sometimes	“Make me one”	Fluent and wrong

The one idea that matters most

A model that performs well on the data it was trained on has demonstrated nothing. It may have memorised.

Overfitting

A model with enough parameters can fit any dataset perfectly, including the noise. It will then fail on new data. The only way to know whether you have learned a pattern or memorised a table is to test on data the model has never seen.

The remedy is discipline, not cleverness: split your data before you start, hold out the test set, and do not look at it until the end. If you tune your model against the test set, it is no longer a test set.

This is not a technicality. It is the single most common reason published models fail to replicate.

What a model cannot tell you

An accurate model tells you two variables carry shared information. It does not tell you:

Which causes which. Correlation is symmetric; causation is not.
Whether a third thing causes both. This is confounding, and it is the subject of the hands-on case study — where a perfectly reasonable analysis produces a result that cannot be interpreted, because of how participants were recruited.
Whether the question was answerable. A model always returns a number. The number is not a judgement about your study design.

Accuracy is a trap on imbalanced data

If 95% of your samples are HIV-negative, a model that predicts “negative” for everyone is 95% accurate and completely useless. Always ask what the base rate is before being impressed by an accuracy figure.

Check yourself

You cluster samples and find three groups. What have you learned?
Your classifier is 99% accurate on training data and 60% on held-out data. What happened?
A tool reports 95% accuracy predicting a condition affecting 5% of people. Is it good?

We work through these in the exercise.