What artificial intelligence is
An explanation without jargon: what a model does, why it can do it, and where it differs from ordinary software.
What this is about
A calculator calculates because somebody wrote down how to calculate. A spell checker corrects because somebody supplied rules and a dictionary. Both are ordinary software: people write rules, the computer executes them.
Artificial intelligence works differently. You show a system a great many examples and let it work out for itself what patterns are in them. What sits in the system afterwards is not rules you could read but billions of numbers.
An image that holds
Picture somebody who never learned a language but has read every text ever written in it. That person does not know what a noun is. They do know very precisely which word follows which, and can therefore write fluently and mostly sensibly.
A language model is exactly that. At every step it predicts which piece of text fits best next, and repeats that until the answer is finished.
What it is not
- No consciousness and no intent. A model wants nothing.
- Not a reference work. It has no register of what it knows.
- Not a search engine. Without a connection to sources it answers from whatever stuck during training.
- Not a calculator. Numbers are treated as text, which is why arithmetic is not a strength.
The three ingredients
- 01
A great many examples
Text, images or measurements. The volume is why it only started working in recent years: the internet supplied it at that scale for the first time.
- 02
A blueprint with many dials
A neural network with billions of parameters. The blueprint fixes how the arithmetic runs, not what comes out.
- 03
Compute to set the dials
Millions of times: predict, measure the error, move every dial a tiny step. That is training, see gradient descent.
What changes in practice
| Ordinary software | Model | |
|---|---|---|
| Rules | written down, readable | derived, not readable |
| Behaviour on the unknown | error message | plausible guess |
| Result for the same input | identical | usually, not guaranteed |
| Changing it | change the code | retrain or change the prompt |
| Checkability | line by line | only by sampling |
The last row has the greatest consequences. A model cannot be checked line by line. It is measured on examples, which means: with an error rate you know rather than a guarantee you hold.
What "learning" means formally
Learning means searching for a function that is rarely wrong on unseen data. Since the underlying distribution is unknown, the error on the sample is minimised instead. The gap between the two is the actual subject of the field, see Overfitting, bias and variance.
What follows for governance
- No specification in the classical sense. There is no document against which behaviour could be fully verified. Measurement against a fixed evaluation set takes the place of verification.
- No determinism without extra effort. Floating-point arithmetic on GPUs is not bit-reproducible. What gets logged is the result plus version, not reproducibility.
- No complete explanation. Attribution methods show which inputs contributed, not why. For a legal justification that is not enough.
- Changes are system changes. A new model, a different system prompt or a different number format alter behaviour and re-trigger the duties.
The boundary that matters under the AI Act
The AI Act defines an AI system by its capacity to infer outputs from inputs that can influence its environment, with some degree of autonomy. What decides is therefore not the technology but the use. The same library can be unproblematic in one case and high risk in another. See Risk classes.
The exercise that shows more than any explanation
FREE ACCOUNT
Exercise: find the error
A twenty-minute exercise that shows you, in your own field, how a model gets things wrong.
Worksheet2 items
Related courses and sources
Elements of AI
A free introductory course from the University of Helsinki, no prior knowledge required. Around 30 hours, with a certificate.
If you want somewhere to go after our beginner path, this is the obvious next step. It overlaps our foundations in places but goes considerably deeper into how things work.
Teachable Machine
Train an image classifier in the browser, without code, in ten minutes. Demonstrates overfitting faster than any explanation.
For a first look without code; it demonstrates overfitting in ten minutes.