Artificial intelligence is the field of building computer systems that do things which, when people do them, seem to require intelligence: recognising a face, translating a sentence, planning a route, writing a paragraph. Most modern AI does this by learning patterns from data rather than following rules a programmer wrote down.
That's the short answer, and it's honest as far as it goes. The longer answer is more interesting, because it explains something the short one hides: why nobody can agree on the definition, why people argue about whether today's systems are "really" intelligent, and why those arguments never resolve.
The definition has never held still
Here is the pattern the field doesn't advertise. A problem is declared to require intelligence. Researchers work on it for decades. Eventually a machine does it. And then, with remarkable speed, everyone agrees the problem never required intelligence in the first place.
Chess was the canonical example of machine intelligence for forty years — until Deep Blue beat Kasparov in 1997, at which point chess became "just search." Reading printed text was an AI problem; now OCR is a checkbox in a scanning app. Filtering spam, recognising speech, translating languages — each one was the frontier, each one fell, and each one stopped being called AI roughly the moment it started working reliably. Researchers call this the AI effect, and it has a corollary: AI is, in practice, defined as whatever machines can't do yet.
The retreat isn't entirely unreasonable. When chess fell, we learned something real: that chess can be won by brute calculation plus clever pruning, without anything resembling understanding. Each time the boundary moves, it's partly because the machine revealed the task was shallower than it looked. Edsger Dijkstra made the sharpest version of this point back in 1984: asking whether a machine can think is about as interesting as asking whether a submarine can swim. Submarines move through water superbly. Whether that counts as swimming is a question about the word, not the machine.
Keep that in mind for everything that follows. Most arguments about what AI "really is" are arguments about vocabulary wearing a lab coat.
What counts as AI right now
The term itself was coined in the 1955 proposal for the Dartmouth workshop, where John McCarthy and colleagues conjectured that every feature of intelligence could "in principle be so precisely described that a machine can be made to simulate it." Seventy years on, the field that proposal launched covers three broad approaches — and understanding the difference between them explains most of what you read about AI.
| Approach | How it works | Where it wins | Where it fails |
|---|---|---|---|
| Symbolic AI | Humans write down knowledge as explicit rules and logic; the machine reasons over them | Problems with clean, complete rules: theorem proving, scheduling, route-finding | Anything requiring knowledge experts can't articulate — which turned out to be almost everything |
| Machine learning | The machine finds patterns in example data instead of being given rules | Prediction from structured data: fraud, pricing, recommendations | Needs data that resembles the future; inherits every bias in its examples |
| Deep learning | Machine learning with many-layered neural networks that learn their own features | Messy perceptual data: images, speech, and — via transformers — language | Data-hungry, expensive, poorly understood, and confidently wrong in ways that are hard to predict |
The first approach dominated for thirty years and lost. Not because the idea was silly — expert systems built this way genuinely worked, and one famously outperformed Stanford medical faculty at diagnosing infections — but because of a wall the field named the knowledge acquisition bottleneck. Getting rules out of a human expert takes months, and the deeper problem is that experts cannot say most of what they know. You recognise your friend's face effortlessly and can't write down how. That gap has a name, Polanyi's paradox — we know more than we can tell — and it's the single best explanation for why learning from data replaced writing down rules.
Nearly everything called AI today is the second and third row. When a news story says a company "uses AI," it almost always means a model trained on data. When it says "generative AI," it means deep learning models that produce text, images or audio — large language models like the one behind ChatGPT, or diffusion models behind image generators.
How it got here, in five acts
The compressed version — the full interactive timeline runs from 1943 to now.
1950: the question gets a test. Alan Turing's famous paper opens by proposing to ask whether machines can think, then immediately abandons the question as meaningless and substitutes a game: can a machine's conversation pass for a human's? The Turing test is remembered as a definition of intelligence. It was actually a way of dodging the definition — Turing's point was that if you can't distinguish the behaviour, the metaphysics stops mattering. The field has been living inside that dodge ever since.
1956–1974: symbols and optimism. Dartmouth names the field. Early programs prove theorems and solve puzzles, and the pioneers predict human-level machines within a generation. The predictions fail, funding collapses, and the first AI winter arrives — a pattern of boom, overpromise and bust that has now happened twice and that everyone in the field quietly watches for.
1980s: expert systems, briefly. Rule-based systems become AI's first real industry, then collapse against the knowledge bottleneck. Second winter.
2012: the deep learning moment. A neural network called AlexNet demolishes the ImageNet image-classification competition, running on gaming GPUs. The ideas were decades old — the perceptron dates to 1958 — but three things finally aligned: enough data, enough compute, and the training tricks to push gradients through deep networks. Within five years, deep learning had absorbed computer vision and speech recognition almost entirely.
2017 onward: transformers eat the field. A new architecture built on attention turns out to scale in a way nothing before it did. Feed it more data and compute and it gets predictably better — a finding formalised as scaling laws — and at sufficient scale it produces the fluent, general-purpose systems that made AI a household topic in 2022. Whether those gains keep coming is the trillion-dollar open question, and "scaling laws" was always a generous name for an empirical trend.
The eleven fields, and where things stand
AI isn't one thing; it's a family of research areas at very different levels of maturity. This site organises 202 concepts into eleven fields — you can see how every one of them connects on the map — and the honest one-line status of each looks like this:
| Field | What it covers | Honest status |
|---|---|---|
| Foundations | The core ideas: learning, search, benchmarks, history | Settled vocabulary, unsettled questions |
| Machine Learning | Learning from data, and measuring whether it worked | Mature; the measurement half is routinely done badly |
| Deep Learning | Neural networks and how they train | Works spectacularly; why it works is genuinely unexplained |
| Language & LLMs | Models that read and write text | The current frontier, moving monthly |
| Generative AI | Producing images, video and audio | Capable, with unresolved arguments about training data |
| Computer Vision | Making sense of images and video | Deployed everywhere; benchmark scores oversell it |
| Speech & Audio | Hearing, transcribing and producing sound | Solved in quiet rooms, hard everywhere real |
| AI Agents | Models that act — call tools, take steps, pursue goals | The gap between demo and product is the whole story |
| Safety & Ethics | Making systems do what we intend, fairly | Far behind capability, and not for lack of trying |
| Tools & Ecosystem | The hardware and software everything runs on | Where the actual moats and costs live |
| Applied AI | The unglamorous AI already running the world | Recommenders and forecasts move more money than chatbots |
One of those rows deserves emphasis because it cuts against the news cycle: the most economically significant AI on earth is probably not a chatbot. It's the recommender systems deciding what a few billion people watch, buy and read next — deployed for two decades, discussed almost never.
What it can't do
Any honest answer to "what is AI" has to include what current AI isn't, because the failures are as characteristic as the successes.
It's confidently wrong. A language model is trained to produce plausible text, and truth is only correlated with plausibility. The result is hallucination: fluent, specific, wrong. This isn't a bug being patched out; it falls out of how the systems are built, which is why it persists across every model generation.
It doesn't reliably plan. Agent demos imply models can break a goal into steps and execute them. The published evidence says they mostly can't plan in any robust sense — performance collapses when problems are rephrased to avoid memorised patterns. What works in production is narrow, heavily scaffolded, and checked by humans or hard guardrails.
Its report card is inflated. Models are compared using benchmarks, and benchmarks have two chronic problems: they get into the training data, which turns a reasoning test into a memory test, and they're routinely mistaken for the ability they approximate. A model that scores like a lawyer on a bar exam does not perform like a lawyer in a courtroom. The gap between benchmark and deployment is where most AI disappointment lives.
Nobody fully knows why it works. This one surprises people. Classical learning theory says models with billions of parameters should memorise their training data and fail on anything new. They don't — deep networks generalise — and the theory that predicts they shouldn't has not been repaired. The field's most powerful tool is, at a foundational level, an open research problem. That's also the fairest way to calibrate both the hype and the dismissals: we are all reasoning about a system whose success is unexplained.
So is it actually intelligent?
You now have everything needed to see why this question never resolves.
"Intelligence" itself is used constantly and defined by nobody — psychology has spent a century failing to agree on a definition for humans, and the AI debate inherits that failure wholesale. Turing saw this in 1950 and swapped the question for a behavioural test. John Searle's Chinese Room argument pushed back in 1980: a system could pass every behavioural test by manipulating symbols it doesn't understand, so behaviour can't settle the matter. Forty-five years later, that argument has neither been refuted nor accepted — the two camps talk past each other because they mean different things by understand.
Which is Dijkstra's submarine again. Today's models translate, summarise, write working code and hold coherent conversations. Those are facts about behaviour. Whether the behaviour constitutes intelligence is a fact about how we've decided to use a word — and the word has been redefined after every previous success. The prediction this history supports: if current systems' limitations become well understood, the boundary will move again, and what they do will retroactively become "just" statistics at scale. The AGI debate is largely this same argument, relocated to a term that's even less well defined.
What to watch
Three things, chosen because each is checkable rather than vibes.
Whether reasoning holds up. The newest models "think" before answering, and the gains on hard problems are real. The open question is how much is genuine reasoning versus sophisticated retrieval — and the visible thinking a model shows you is a plausible story, not a reliable account of what happened inside.
Whether agents cross the reliability gap. The industry's current bet is models that act rather than answer. Watch the agent evaluation numbers, not the demos: a system that succeeds 90% of the time per step fails most ten-step tasks.
Whether the measurement crisis gets fixed. The field's most cited claims — including the famous "emergent abilities" that appear suddenly at scale — have a way of dissolving under scrutiny; the emergence result was substantially a plotting artefact, a finding that won a best-paper award and changed far fewer headlines than the original claim. Until evaluation improves, treat every capability announcement, positive or negative, as provisional. The annual Stanford AI Index is a reasonable place to watch the aggregate picture.
If you want the same material at whatever depth suits you, every concept linked above is explained at five levels — start with Artificial Intelligence itself, browse a field, look anything up in the glossary, or open the map and wander.
Frequently asked questions
What is artificial intelligence in simple terms?
AI is software that does things which normally need human intelligence — recognising images, understanding speech, writing text, making predictions. Modern AI learns how to do these things from examples rather than being programmed step by step, the way you'd learn to recognise a dog from seeing dogs, not from reading a definition of one.
Is AI the same as machine learning?
No — machine learning is one approach to building AI, and today the dominant one. AI is the goal (machines doing intelligent-seeming things); machine learning is a method (learning patterns from data instead of following hand-written rules). Deep learning is in turn one family of machine learning, built on layered neural networks. Almost everything currently in the news is that innermost family.
Is ChatGPT artificial intelligence?
Yes, by any working definition of the term. ChatGPT is built on a large language model — a deep learning system trained on enormous amounts of text to predict what comes next, then tuned on human feedback to behave like an assistant. Whether that constitutes "real" intelligence is the vocabulary debate this article describes, and it has no settled answer.
Who invented artificial intelligence?
No single person. Alan Turing laid the conceptual groundwork in his 1950 paper on machine intelligence. The term itself was coined by John McCarthy in the 1955 proposal for the 1956 Dartmouth workshop, which is treated as the field's founding event. The neural-network line of work traces to McCulloch and Pitts in 1943 and Rosenblatt's perceptron in 1958.
What are the main types of AI?
The practical split is by approach: symbolic AI (hand-written rules and logic), machine learning (patterns learned from data), and deep learning (many-layered neural networks). You'll also see a split by ambition — "narrow AI" for systems good at one task, which is every deployed system today, versus AGI for hypothetical broadly capable systems. The narrow/general framing is common in media and rare among practitioners, mostly because AGI is too loosely defined to build toward.
What can AI not do reliably?
Current systems can't reliably tell you when they're wrong — they produce confident falsehoods. They can't robustly plan multi-step tasks without heavy scaffolding. They struggle with anything genuinely unlike their training data. And their reported abilities are measured on benchmarks that systematically overstate real-world performance. None of these is a minor engineering gap; each traces to how the systems fundamentally work.