What you'll learn
- What the term artificial intelligence actually describes, and why it is harder to define than it looks
- The difference between narrow AI (everything that exists today) and general AI (which does not)
- How to tell an AI system apart from an ordinary program, a set of rules, or simple automation
- How AI, machine learning and deep learning fit inside one another
- The everyday systems that already use AI, many of which are never advertised as AI
- What AI does not have — understanding, intention, feelings or experience — and why the words we use about it mislead
Key terms and definitions
| Term | Meaning |
|---|---|
| Artificial intelligence (AI) | A computer system that performs tasks normally thought to need human intelligence, such as recognising speech, translating languages or making predictions |
| Narrow AI | A system that performs one kind of task, however well. Every AI system in use today is narrow |
| General AI | A hypothetical system that could handle any intellectual task a person can. It does not exist |
| Machine learning (ML) | An approach to AI in which a system learns patterns from data instead of being given explicit rules |
| Deep learning | Machine learning using neural networks with many layers; the approach behind most recent advances |
| Algorithm | Any set of step-by-step instructions for solving a problem. Most algorithms are not AI |
| Automation | Getting a machine to do a task without a person doing it each time. Automation need not involve AI at all |
| Generative AI | AI that produces new content — text, images, audio, code — rather than only classifying or predicting |
| Anthropomorphism | Describing a non-human thing as though it had human thoughts and feelings |
Core concepts
Defining artificial intelligence
A workable definition is this: artificial intelligence is a computer system that performs a task we would normally expect to require human intelligence. Recognising a face, understanding a spoken sentence, translating Spanish into English, suggesting what you might want to watch next — these were all once things only people could do.
That definition has an awkward consequence, and it is worth facing early: it depends on what we currently find impressive. Optical character recognition — reading printed text from a photograph — was a serious AI research problem for decades. Today it is a feature of your phone's camera and nobody calls it AI. This pattern is common enough to have a name, the AI effect: once a problem is solved reliably, people stop regarding it as intelligence and start regarding it as ordinary software.
So "is this AI?" is often less useful than two sharper questions:
- What task is it doing?
- Does it learn the task from data, or was it given the rules?
The second question is the one that separates modern AI from everything that came before.
Narrow AI and general AI
Narrow AI does one kind of thing. A system that plays chess at world-champion level cannot order your lunch. A system that transcribes speech cannot diagnose a rash. A large language model can write about an extraordinary range of subjects, but it is still doing one task — producing likely text — and it cannot drive a car or feel the weather.
General AI would be a system that could turn its ability to any intellectual problem, learn new domains on its own, and transfer understanding from one to another the way a person does. No such system exists. This matters practically, not just philosophically: when a product claims to have "general intelligence", that claim is marketing rather than a description, and the right response is scepticism.
The confusion is understandable. A chatbot that answers questions on history, chemistry and cooking feels general. But breadth of subject matter is not generality of ability. The task has not changed; only the range of content has.
AI, machine learning and deep learning
These three terms are often used as though they were interchangeable. They are nested:
- Artificial intelligence is the widest circle — any system doing a task associated with intelligence, including older systems built entirely from hand-written rules
- Machine learning sits inside it — systems that derive their own rules from examples
- Deep learning sits inside machine learning — ML using many-layered neural networks
An expert system from the 1980s that diagnosed illnesses using hundreds of rules written by doctors is AI, but it is not machine learning: a human wrote every rule. A spam filter that was shown a million emails and worked out for itself which features predict spam is machine learning. A model that generates text is deep learning.
Keeping the nesting straight prevents a common error in written answers: "AI and machine learning" implies they are two separate things, when one contains the other.
Telling AI apart from rules and automation
This is the distinction most often got wrong, so it is worth a clear test. Ask: where did the rules come from?
| System | Rules written by | AI? |
|---|---|---|
| A traditional thermostat switching on below 18 °C | A person | No — one fixed rule |
| A washing machine running a timed cycle | A person | No — automation, not intelligence |
| A spell-checker comparing words against a dictionary list | A person | No — look-up against rules |
| Predictive text suggesting your next word from what you usually type | Learned from data | Yes |
| A thermostat that learns your household's routine and pre-heats accordingly | Learned from data | Yes |
| A spam filter trained on labelled examples | Learned from data | Yes |
Notice that the thermostat appears twice. The physical job is almost the same; what differs is whether a person specified the behaviour in advance or the system inferred it from observation.
Two slogans to retire:
- "AI is just a lot of if-statements." A rule-based program with thousands of conditions is still following rules a person wrote. A machine learning model has no such list — it has adjustable numbers fitted to data, and nobody wrote the behaviour it ends up with.
- "If it's automated, it's AI." A supermarket self-checkout automates a job without doing anything intelligent. Automation describes who does the work; AI describes what kind of work it is.
AI you already use without noticing
Most AI is invisible, because the systems that shout about it are a small minority of the ones you touch:
- Recommendation — what a streaming service puts on your home screen, what a shop suggests alongside your basket
- Filtering and ranking — spam detection, which search results come first, what appears in a feed
- Recognition — unlocking a phone with your face, tagging people in photos, reading a number plate
- Speech and language — voice assistants, automatic subtitles, translation
- Navigation — predicting traffic and choosing a route
- Generation — writing, summarising, drawing, producing code
Realising how much of this is routine is part of the point of the subject. AI is not only the chatbot you deliberately open; it is already shaping what you are shown, and that is a reason to understand it rather than to be alarmed by it.
What AI does not have
A language model will write "I think the main cause was industrialisation". It is extremely easy to read that as thought. It is not.
A model of this kind holds no beliefs, wants nothing, understands nothing in the way you understand the sentence you are reading, and has no experience of the world it writes about. It has learned which words tend to follow which, across an enormous quantity of text, well enough to produce writing that is often accurate and sometimes wrong in ways it cannot detect.
We talk about these systems in human terms because the shorthand is convenient — the model knows, it decided, it is trying to. The shorthand is harmless in conversation and dangerous in reasoning, because it quietly implies two things that are false: that the system has checked what it says against reality, and that it would know if it were wrong.
Describing systems accurately is a skill, and it is examinable:
- not "the AI understood my question" but "the system produced a relevant response"
- not "it knows the capital of Peru" but "that fact appeared consistently in its training data"
- not "it lied to me" but "it produced a confident statement that was false"
The third is the most important, because lying requires an intention to deceive, and there is nothing there to intend.
Worked examples
Example 1: Classifying a system (3 marks)
A school fits a lighting system that switches lights off when a room has been empty for ten minutes. The manufacturer advertises it as "AI-powered". Is this description justified? Explain your answer.
- The system follows a single fixed rule — no movement for ten minutes, switch off — written by its designers (1 mark)
- It does not learn from data and its behaviour does not change with experience, so it is automation rather than AI (1 mark)
- The description is not justified; "AI-powered" is being used as marketing language (1 mark)
Example 2: Comparing two systems (4 marks)
Compare a spell-checker that flags words missing from a dictionary with predictive text that suggests your next word. Explain which is AI.
- The spell-checker compares each word against a fixed list, which is a look-up against rules written by people (1 mark)
- Predictive text has learned patterns from large quantities of text and from what this user tends to write (1 mark)
- Predictive text is therefore AI and the spell-checker is not (1 mark)
- The deciding difference is where the rules came from, not how useful or complicated the system is (1 mark)
Example 3: Judging a claim (3 marks)
A company advertises an app as having "general artificial intelligence". Explain why this claim should be treated sceptically.
- General AI would handle any intellectual task and transfer learning between domains (1 mark)
- No system of that kind currently exists; all working AI is narrow (1 mark)
- A system that answers questions on many subjects is still performing one task, so breadth of content is being mistaken for generality of ability (1 mark)
Common mistakes and how to avoid them
- Using "AI" and "machine learning" as separate things. Machine learning is one way of building AI. Say "machine learning, a type of AI".
- Calling any automation AI. Ask where the rules came from. If a person wrote them, it is not AI however convenient it is.
- Assuming AI means a chatbot. Generative AI is recent and visible; classification, ranking and prediction have been in everyday use far longer.
- Treating fluent writing as evidence of understanding. Fluency shows the text is likely, not that it is true or understood.
- Saying a system "lied" or "chose to". These imply intention. Write "produced a false statement" and "output was".
- Believing general AI exists because a system seems versatile. Range of subject matter is not range of ability.
Using this in practice
When you meet a new system — in the news, in an advert, or in an app you have just installed — three questions will usually place it correctly:
- What task is it performing? Name the task plainly: ranking, recognising, predicting, generating.
- Where did its behaviour come from? Written rules, or patterns learned from data?
- How would I know if it were wrong? This is the question people skip, and it matters most for systems that generate rather than classify.
Answer those three and you can describe almost any system accurately, without either dismissing it or overestimating it — which is the habit the rest of this course builds on.
Quick revision summary
- AI is a system doing a task normally thought to need human intelligence; the label shifts as technology becomes ordinary
- All AI today is narrow; general AI does not exist, and claims that it does are marketing
- AI ⊃ machine learning ⊃ deep learning — each is contained in the one before
- The test that separates AI from automation is where the rules came from: written by people, or learned from data
- Most AI you use is invisible — recommendation, filtering, recognition, navigation — not just chatbots
- These systems have no beliefs, intentions or understanding; describe what they produce, not what they "think"