What is artificial intelligence?

Artificial intelligence is the field that builds systems able to carry out tasks that, done by a person, we would call intelligent: recognizing, deciding, planning, translating, generating. Since 2012, most of what works comes from learning: instead of somebody writing a rule for each case, the system is shown examples and extracts a pattern from them.

The part most introductions leave out is that this definition was never settled. It rests on what counts as intelligent, and that shifts as machines take over more of it.

What underpins nearly everything today is machine learning, whose dominant form is deep learning, which stacks layers of neural networks. Each of those terms is narrower than the one before it, and swapping them as though they were synonyms is the most common confusion in the subject.

The four layers, and where the drawing misleads

The nested-circles diagram is the most reproduced picture in the subject, and it gets the containment right: all generative AI is deep learning, all deep learning is machine learning, all machine learning is AI. What it hides are the three exceptions that surface the moment anyone works with this.

There is AI that learns nothing. A chess engine running tree search, a constraint solver building an on-call rota, a rule system approving a credit application: all three sit in the AI literature from before machine learning was practical, and none of them has weights to adjust.

There is machine learning that is not deep, and it still wins plenty of real cases. Logistic regression, decision trees and gradient boosting are machine learning, train in minutes, explain themselves reasonably well and remain competitive on tabular data, which is the shape most corporate data comes in.

And there is generation that is not deep learning. Count-based language models were generating text decades before deep networks — badly, but generating. In the other direction, most deep learning in production generates nothing at all: it classifies, detects and ranks.

The drawing is a teaching tool, not a taxonomy. Use it to place the terms and drop it when the question becomes which architecture to build.

The types of artificial intelligence

The lists of “three types” or “four types of AI” in circulation come from promotional material rather than research. The reactive, limited-memory, theory-of-mind and self-aware ladder is popular and has no technical use: nobody designs a system by picking which rung it lands on.

Two distinctions do get used. The first is by technique: symbolic AI, where behaviour comes from rules and structures people wrote, against learning-based AI, where it comes from data. The second is by scope: task-specific systems, which is everything running in production, against artificial general intelligence, which is a hypothesis with no agreed operational definition.

That lack of agreement is not an oversight. It is the subject of the work on definitions that comes next.

A definition that never settled

In 1950, Alan Turing opened a paper by proposing to consider the question “can machines think?” and, in the next sentence, threw it out. Defining “machine” and “think” from common usage would, he wrote, mean answering the question with a Gallup poll, which is absurd. In its place he proposed a substitute question, framed as a game he called the imitation game¹.

That manoeuvre is the origin of a habit the field keeps to this day: when the definition will not close, swap the definition for a test. The term “artificial intelligence” appeared a few years later, in the proposal for the Dartmouth summer research project, held in 1956 and treated ever since as the field’s founding marker².

Seventy years on the argument is still open, and there is serious work on it. A 2019 paper systematically analysed the problem of defining artificial intelligence and started from a point that usually goes unnoticed: the definition you adopt influences the path the research takes, so it is not a preamble, it is a technical choice. The author set out four criteria for a good working definition — being close to the term’s common usage, drawing a sharp boundary, leading to fruitful research, and being as simple as possible — assessed the field’s representative definitions against them, and then offered his own: intelligence as adaptation with insufficient knowledge and resources³.

Hold on to the shape of the problem. There is no canonical definition of AI that everyone accepts; there are definitions serving different purposes. When somebody says “that isn’t real AI”, they are almost always using a different definition from yours, and the argument goes nowhere until one of you says which.

What changed: from written rule to learned pattern

For decades, building an AI system meant extracting rules from an expert and writing them down as code. To recognize a cat in a photograph, somebody had to describe what makes a cat look like a cat in terms of edges, textures and proportions. That worked in closed domains with few variables, and broke the moment the world produced a case nobody had foreseen.

The shift was not about cleverness, it was about authorship. A 2013 review of representation learning put the point precisely: the success of machine learning algorithms depends on how the data is represented, because different representations entangle or expose the underlying factors of variation; domain knowledge can help design that representation, but learning with generic priors works too⁴. In practice: instead of a human deciding which features matter, training finds out.

The moment this stopped being a bet has a date and a number attached. In the 2012 ImageNet competition, a deep convolutional network with 60 million parameters and 650,000 neurons won with a top-5 test error rate of 15.3%, against 26.2% for the second-best entry⁵. Eleven percentage points, in a competition where progress had been arriving in fractions of a point, reorganized the field within months.

Five years later a second shift did the same for text. A 2017 paper proposed an architecture that dispensed with recurrence and convolution entirely and used only attention mechanisms, reaching 28.4 BLEU on WMT 2014 English-to-German, over 2 points above the previous best including ensembles, and training at a fraction of the cost⁶. That architecture, the Transformer, is the base of the language models and the generative AI that followed.

How a system like that learns

Without any mathematics, the mechanism fits in a paragraph. The model is a function with a great many adjustable numbers, the weights. You show it an input, it produces an output, and a measure of error compares that output with the right answer. Training nudges the weights in the direction that reduces the error, and repeats that millions of times over different examples. In the end the weights encode regularities in the data that nobody wrote down.

Two practical consequences fall out, and both explain a great deal further on.

The first is that the data is the specification. The system’s behaviour is determined by the examples it saw, not by a requirements document. If the examples carry a bias, the system carries the same bias, and it will not surface in code review because it is not in the code.

The second is that training and use are separate phases. Training is expensive and happens once; running the trained model costs little per call and happens constantly. A model in production does not learn from what you type into it unless somebody built an explicit pipeline for that.

The map of the subject

Learning covers how a system extracts patterns: what machine learning is, what deep learning added, and the ways of learning — with labels, without labels, or from reward.

Architecture covers what these systems are made of: the neural network as the basic part, what people mean by a model, and what changes when the output is generated rather than classified.

Limits is where the subject becomes useful. Overfitting, bias and variance, hallucination and the argument about general intelligence are the four places where expectations usually come apart from what the system delivers.

Ethics and regulation deal with what gets done about all of it. Explainability and responsible AI are the technical and organizational answer; regulation is the legal one, and it varies by jurisdiction.

Where AI breaks

It does not know what it does not know. A model produces the most likely answer given the pattern it learned, and nothing in that mechanism separates “likely” from “true”. When information is missing, the output does not come back empty: it comes back plausible. That is the root of hallucination, and it is not a bug more data fixes.

It inherits bias from the data. A screening system trained on past decisions reproduces past decisions, bad ones included. The problem is not that the algorithm has an opinion; it is that it has none, and therefore copies whichever one was in the examples.

It breaks outside its training distribution. Excellent test-set performance guarantees nothing about inputs that do not resemble what was seen. That is why systems that work in a demo fail in production: the demo uses data close to the training set and production does not.

It is opaque by construction. Billions of weights tuned by optimization do not constitute an explanation. Interpretability techniques exist and they give hints, not justifications. In decisions affecting people, that is a legal problem before it is a technical one.

A benchmark score is not a capability. A high score measures performance on that set of tasks, in that question format. Generalizing to your case is a hypothesis you have to test, not a conclusion you get for free.

The cost is real and recurring. Training a large model consumes energy and hardware at industrial scale, and running it consumes per call. Plenty of projects that die do not die on quality; they die because the per-request bill does not work at the volume the product needs.

Where to start

1. Pick the problem by its shape, not by the technology. If you can describe the task as “given this, produce that” and gather a few thousand examples, it is a learning problem. If you cannot, more AI will not help.
2. Check whether a simple rule solves it. A lot of what gets presented as AI is an if statement with marketing attached, and an if is cheaper to maintain, easier to audit and never hallucinates.
3. Look at the data before the model. The quality and bias of your examples decide the outcome more than the choice of architecture does.
4. Start with an off-the-shelf model. In almost any text, image or audio case, calling an existing model ships faster than training one, and prompt engineering is how you steer that model without touching its weights.
5. Define how you will measure before you build. Without a set of cases with expected answers, every improvement is an impression.
6. Design the failure mode as deliberately as the happy path. What the system does when it does not know is a product decision, and it needs to be written down somewhere.

Step 5 is what separates an AI project from an AI demo, and it is the one almost every team discovers late.

Footnotes

1. Turing (1950) proposed replacing the question “can machines think?” with a behavioural test, arguing that defining the terms from common usage would reduce the answer to a matter of opinion polling. ↩

2. Kline (2010) documents, from unpublished archives, the origins of the 1956 Dartmouth conference, treated as AI’s official birthplace, and the field’s relationship with cybernetics. ↩

3. Wang (2019) establishes four criteria for a good working definition of artificial intelligence, assesses the field’s representative definitions against them, and proposes intelligence as adaptation with insufficient knowledge and resources. ↩

4. Bengio, Courville and Vincent (2013) review representation learning and argue that performance depends on how data is represented, because different representations entangle or expose the explanatory factors of variation. ↩

5. Krizhevsky, Sutskever and Hinton trained a convolutional network with 60 million parameters and 650,000 neurons that won ILSVRC-2012 with a top-5 error rate of 15.3% against 26.2% for the runner-up. ↩

6. Vaswani et al. (2017) proposed the Transformer, based solely on attention mechanisms, reaching 28.4 BLEU on WMT 2014 English-to-German, more than 2 points above the previous best, at a fraction of the training cost. ↩