Artificial intelligence is the most used and least understood term of recent years. It is applied to everything: the algorithm that decides which video to show you next, the model that writes an essay, the system that detects tumours in X-rays, and the robot that assembles cars in a factory. The fact that the same label is used for such different things is not an accident — it is the consequence of a term that was born as an aspiration before becoming a technical description.
To use it well, you first need to understand what it is. And for that, two myths that dominate the conversation need to be dismantled.
The problem with the definition
The first myth is that AI is intelligent in the human sense of the word. It is not. It does not think, reason or understand. What it does is process very large quantities of information and find statistical patterns within it that allow it to make predictions. That the result resembles intelligent reasoning is a consequence of the volume of data and the sophistication of the method — not because there is anything resembling consciousness or understanding.
The second myth is the opposite: that AI is “just statistics” and therefore has nothing special about it. This is also wrong. The scale of what current models can do — generate coherent text, translate languages, synthesise medical research, write code — has no historical precedent. The fact that the mechanism is statistical does not make it trivial.
The reality lies between the two myths.
What AI actually does
A useful, if imperfect, definition: AI is a set of computational techniques that allow systems to improve their performance on a task through experience, without being explicitly programmed for each case.
The important part is the contrast with classical programming. In traditional programming, a human writes explicit rules: if X occurs, do Y. The system follows those rules. If reality falls outside the anticipated rules, the system fails.
AI works the other way: the system receives many examples of inputs and desired outputs, and learns for itself what transformation relates one to the other. The rules are not written — they emerge from training.
This has an enormous consequence: AI systems can work in domains where writing explicit rules is impossible. Nobody could write rules to recognise a face in all its possible variations of light, angle and expression. But a model trained on millions of labelled images can do so with high accuracy.
Pattern, prediction, probability
All modern AI — from recommendation models to large language models — operates on the same fundamental logic: find patterns in data and use those patterns to make predictions.
An image recognition model learns: “these pixels together, in these arrangements, tend to correspond to a cat.” A language model learns: “after these words, in this context, this other word is likely to come next.”
The prediction is never deterministic: it is probabilistic. The model does not say “the next word is X.” It says “the probability that the next word is X is 47%, that it is Y is 23%…” What appears as certainty in the output is a choice among probabilities.
This has direct practical implications. When AI produces convincing text about something incorrect, it is not consciously lying: it is generating what its parameters indicate is statistically probable given that context. Factual correctness is not part of the mechanism.
Why now
AI is not new. The first machine learning models are more than 70 years old. What has changed is the combination of three factors that until recently were not all available together:
Data. The massive digitisation of human activity — text, images, video, transactions — has produced quantities of training data without historical precedent. Modern models learn from practically everything humans have written or photographed.
Computation. GPUs, originally designed for video games, turned out to be perfect for the massive linear algebra required by neural network training. Their cost has dropped dramatically while their power has increased.
Architecture. In 2017, a team at Google published “Attention Is All You Need,” the paper that introduced the transformer architecture. That architecture, which allows models to process relationships between elements distant in a sequence, turned out to be the missing piece for scaling language learning to previously impossible dimensions.
The confluence of these three factors — not any one alone — is what makes 2017–2024 qualitatively different from everything before.
What changes when you understand it well
Understanding what AI is — and what it is not — is not an academic exercise. It has immediate practical consequences.
If you know that models predict based on statistical patterns, you understand why they are brilliant in domains well represented in their training data and fail in new or infrequent situations. You know when to trust and when to verify.
If you know there is no real understanding behind the text a model generates, you understand why it is not enough that it sounds convincing: you need to evaluate the content, not just the form.
If you know that AI emerges from training with human data, you understand why it reflects the biases in that data and why neutrality is not automatic.
The starting point of this course is not technical: it is conceptual. You do not need to know mathematics to use AI with judgement. You do need to understand what it does and what it does not do. Everything else comes after.
