What is AI, actually?

The short answer, and the word it turns on

Artificial intelligence is software that works out how to produce its output rather than following instructions a person wrote for it.

Everything else is detail. A payroll system applies rules somebody specified: if the employee is in this band, deduct this percentage. Every step exists because a human decided it should, and you can read that decision in the code. An AI system is handed examples instead of rules, derives a relationship from them, and then applies that relationship to inputs nobody anticipated. Nobody wrote the rule it uses. In many cases nobody can state it.

That is inference, and it is the hinge on which every serious definition of AI now turns, including the legal one.

The practical consequence of the right-hand column is the one people underestimate. When ordinary software is wrong there is a defect somewhere and it can be found. When a model is wrong, the error is frequently a faithful reflection of what it was shown, and the fix is a data problem, a retraining cycle and a fresh evaluation rather than a patch.

What "learns from data" actually means

Two phases, and conflating them causes most of the confusion in governance conversations.

Training is where the learning happens. The system is shown a large number of examples and adjusts internal numbers, called parameters or weights, until its outputs on those examples are close enough to what was wanted. This is expensive, is done once or occasionally, and produces an artefact: a file of numbers.

Inference is what happens afterwards, every time the system is used. The file of numbers is fixed. An input arrives, it is passed through, an output comes out. Nothing is learned. The system does not remember your question tomorrow unless something outside the model was built to store it.

This is why "the AI learned from our confidential document" is usually the wrong worry and occasionally exactly the right one. Pasting a document into a chat interface does not by itself change any weights. Whether that document is retained, logged, reviewed by humans or later used for training is a contractual question about the vendor, not a technical property of the model, and it is answered by reading the terms rather than by reasoning about how models work.

The types of AI, and which taxonomy is real

Search for the types of AI and you will meet two classifications, presented with equal confidence. They do not deserve equal confidence.

The four types, which is one person's essay

Reactive machines, limited memory, theory of mind, self-aware. This scheme is everywhere: vendor pages, training courses, certification syllabuses.

It comes from a single article by Arend Hintze, then an assistant professor at Michigan State University, published in The Conversation on 14 November 2016. It is a thoughtful piece of writing. It is not a standard, it was not produced by a standards body, and it does not reflect a research consensus.

More to the point, Hintze wrote plainly that the last two categories do not exist. Theory of mind, systems that form representations of what other minds believe and want, is described as research yet to be achieved. Self-aware AI is presented as a goal that would first require understanding consciousness.

A taxonomy in which half the categories are empty is a description of an imagined future, not a way of sorting the things in front of you. It is fine as a thought experiment about where the field might go. It is close to useless for deciding what to do about the model your marketing team started using last month, and being asked to memorise it as though it were the periodic table is how a lot of AI training wastes people's time.

Narrow, general, super: better, and still slippery

The second classification sorts by breadth. Narrow AI does one kind of task. General AI would match human capability across the range of cognitive tasks. Superintelligence would exceed it.

Narrow is real and is all anybody has deployed. The other two are projections. The difficulty is that "general" has never had an agreed meaning, which makes claims about proximity to it unfalsifiable.

That is not a complaint from outside the field. In November 2023 Meredith Ringel Morris and colleagues at Google DeepMind, including the company's co-founder Shane Legg, published a paper precisely because the term was being used incompatibly. They catalogued nine existing definitions, from the Turing Test to "economically valuable work" to the coffee test, and proposed a scale measuring performance against generality instead.

Read that across rather than down, because the shape of it is the most useful thing on this page.

Narrow systems already reach the top row. AlphaFold predicts protein structure better than any person. Stockfish plays chess beyond any human. In their own lanes these are not approaching human capability, they passed it and kept going.

General systems sit on the bottom rung. The models that dominate public conversation, the ones people mean when they say AI, are placed at Level 1 of 5 by researchers at the laboratory building them, and every level above is marked not yet achieved for general capability.

Both facts are true simultaneously, and holding them together is most of what it takes to reason about this technology sensibly. Systems that are superhuman within a boundary, and unreliable the moment they are asked to leave it, are not well served by a debate conducted as though there were a single dial marked "how intelligent".

AI, machine learning, deep learning, generative AI

These are nested, not alternatives, and the nesting is worth getting right because contracts and policies routinely use the terms as though they were interchangeable.

Artificial intelligence is the outer category: systems that infer rather than follow instructions. It includes older symbolic approaches such as expert systems, which encoded human knowledge as logic and dominated the field into the 1990s.

Machine learning is the subset that derives behaviour from data. This is the dominant method now, so much so that most people use the two words interchangeably, but they are not the same claim.

Deep learning is the subset of machine learning using neural networks with many layers. It is what made the last fifteen years look different from the previous forty.

Generative AI is the subset of deep learning that produces new content, text, images, audio or code, rather than a classification or a score. Large language models are the best known instance.

A governance consequence: a policy written about "generative AI" does not cover the CV screening model in your recruitment system, which is machine learning but not generative. That model is far more likely to attract legal attention, because what regulators care about is the decision being made about a person, not the novelty of the output.

The definition that actually binds you

Every classification above is descriptive. One definition is operative: if a system meets it, a body of law attaches to it. Article 3(1) of the EU AI Act reads as follows, and the Commission's guidance of 6 February 2025 sets out seven constituent elements within it.

a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments

Two things in that annotation are commonly got wrong, and both are worth checking against your own assumptions.

Adaptiveness is not required. A great deal of writing treats continuous learning as the test, on the reasoning that a system which never changes is just software. The Commission guidance closes this off directly: because the provision says may exhibit adaptiveness, a system that never updates after release is still an AI system. Almost every commercially deployed large language model has frozen weights, and all of them are in scope.

Inference is the test, not the technology. The definition names no architecture. It does not require a neural network or mention machine learning. A 1980s expert system can meet it. Some very modern software does not. The Commission guidance states that systems based on rules defined solely by natural persons to automatically execute operations fall outside, which is what keeps spreadsheets, query engines and threshold alerts out of scope however elaborate they become.

Meeting the definition brings a system into scope. It does not tell you where in scope. That is decided separately, largely by Article 6 and by what the system is used to decide, and the same model can be high-risk in one deployment and barely regulated in another. A language model summarising internal meeting notes and the same model screening job applicants are the same technology in entirely different regulatory positions.

Which obligations land on you then depends on what you are in relation to the system, and most organisations misjudge this in the same direction.

Applying the test to three real systems

The definition is only useful if it survives contact with things you actually have. Here are three systems most organisations run, each taken through the seven elements in order. The answers are less obvious than they look, and two of the three land somewhere people do not expect.

A spam filter using a trained classifier. Machine-based, yes. Autonomy, yes: it files messages without asking. Adaptiveness is present in many products and irrelevant either way. It has an implicit objective, keep unwanted mail out of the inbox, which nobody had to write down for it to count. It infers, because nobody wrote the rule that this particular message is spam; the model derived it from labelled examples. It produces a decision. That decision moves mail, which influences a virtual environment. All seven elements are met, so a spam filter is an AI system under the Act. People find this surprising and then object that it is obviously harmless. Both things are true. Being in scope is not the same as being high-risk, and conflating the two is the most common error in early AI Act conversations.

A CV screening tool that ranks applicants. The same seven answers, and it reaches them faster because the inference is the entire product. The difference is what happens next: this system is used to make a decision about a person's access to employment, which is the sort of use Annex III concerns itself with. Identical answers on the definition, an entirely different regulatory position afterwards. This is the pair to keep in mind whenever somebody argues that the definition is too broad. The definition is a door, not a verdict.

A rules-based fraud alert that fires above a threshold. Machine-based, yes. Some autonomy, yes, it fires without a human pressing anything. But element five fails. A person decided the threshold and a person wrote the conditions, so the system executes logic specified by natural persons rather than inferring how to produce its output. The Commission guidance puts this class of system outside the definition explicitly. Add a model that learns which transactions are anomalous and element five is satisfied, and the same product moves into scope on the strength of that one change.

The pattern across all three: elements one, two, four, six and seven are satisfied by almost anything that runs and does something. Element three is optional by its own wording. Element five decides the case, and it decides it on a question you can usually answer in a sentence. Did a person write the rule this system applies, or did the system derive it?

If you cannot answer that about a system you own, that is worth knowing on its own, and it is normally a sign the vendor has not been asked.

What AI is not good at

A guide that only describes capability is marketing. The limits below are structural rather than temporary, in the sense that they follow from how the systems work rather than from this year's models being immature.

Knowing whether it is right. A model produces the output its training makes probable. Confidence in the phrasing is a property of the style, not evidence about the content. This is why a fluent, well-formatted, entirely incorrect answer is the characteristic failure rather than an unusual one.

Anything outside the distribution it was trained on. Performance degrades on inputs unlike the training data, and it degrades without warning or any change in tone. The superhuman narrow systems in the table above are superhuman inside a boundary and useless a step outside it.

Explaining its reasoning. A model asked why it produced an answer generates a plausible account of why such an answer might be produced. That account is another output, not an inspection of the process, and it can be confidently wrong about the very computation that generated it.

Stable behaviour over time. Vendors update hosted models. A prompt validated in March can behave differently in September with no change on your side, which is why anything consequential needs its own evaluation set rather than a one-off check at procurement.

Handling untrusted input safely. A system that reads a web page, an email or a document is reading instructions from whoever wrote it. Prompt injection is unsolved, not merely unpatched, and it is the reason giving a model both untrusted input and real permissions is the single riskiest pattern in current deployments.

Where to go next

If the question is now when the Act's obligations reach you, the deadlines moved in July 2026 and most published timelines have not caught up. The tool below filters the Act to the obligations that are actually yours.

If you want to watch inference happen rather than read about it, the explainer runs a model step by step and shows the intermediate state at each stage.

If the pressing question is what to do first, the AI governance roadmap starts where it should start, with an inventory rather than a policy.