Definition
Most software behaves the way it does because a person wrote the rule down: if the balance is below zero, decline the payment. Artificial intelligence is the part of computing where the specific answer was never written by anyone — the programmer supplies a method for producing answers, a search through possibilities or a function fitted to examples, and the answers fall out of that method. That line is real and checkable, unlike the textbook version ("machines doing things that would require human intelligence"), which defines the word using the word.
Both sides are called AI. In the older, symbolic tradition a person writes the rules and the machine explores their consequences at inhuman speed: IBM's Deep Blue, which beat world chess champion Garry Kasparov in 1997, searched around 200 million positions per second (IBM) using an evaluation function tuned by grandmasters. Nobody wrote down which move to play in a given position, but humans did write down what "good" meant.
On the other side — what nearly everyone means by AI today — nothing is written down at all. No engineer wrote a rule describing what a cat looks like or when a sentence is sarcastic. A model was shown a very large number of examples and the rule was inferred, ending up as billions of numbers no human chose or can read. That is why machine learning is almost a synonym for AI: since about 2012 it is how nearly all of it is built.
One more thing is worth knowing: the AI effect. As soon as a capability becomes reliable, people stop calling it AI. Optical character recognition, spam filtering, route planning and speech transcription were each frontier research problems and are now plumbing nobody labels. Larry Tesler summarised it as "AI is whatever hasn't been done yet" — the term names the current frontier, not a fixed body of techniques.
How It Works
A modern AI system is a loop: a model holding a large number of adjustable numbers, a measure of how wrong it is, and a procedure for making it less wrong. Those numbers are parameters, or weights; before training they are random and the output is noise. Training shows the model an example, scores the output with a loss function, and shifts every parameter in the direction that reduces the error — gradient descent. Repeat across billions of examples and they settle into a configuration that behaves usefully. No person decided any parameter's value, which is the whole difference from ordinary programming.
The scale gap between the eras makes the point concrete. MYCIN, an early-1970s Stanford system for diagnosing blood infections, held roughly 600 rules, each elicited from a physician and typed in by hand. GPT-3, released in 2020, had 175 billion parameters — about 290 million times as many adjustable pieces, and enough that writing them out at one per second, without sleeping, would take some 5,500 years. Modern AI is not the same idea with more rules; its knowledge was never in a form a person could have written.
What the loop optimises for is generalization: performing on data it has never seen. This is also why AI fails differently from normal software: a code bug fails identically every time and traces back to a line; a fitted model degrades quietly on inputs slightly unlike its training data, with no line to look at.
Why the approach changed
The field was named 70 years ago, at a two-month workshop at Dartmouth College in the summer of 1956 organised by John McCarthy, Marvin Minsky and Claude Shannon. For three decades the dominant bet was symbolic AI: encode human knowledge explicitly as rules. Commercialised in the 1980s as expert systems, it worked in narrow domains and broke everywhere else: rules had to be extracted from experts one at a time and did not survive messy real-world input. Funding collapsed twice — the "AI winters" after the UK's 1973 Lighthill report and, later, the death of the Lisp machine market around 1990.
The statistical turn began quietly in speech recognition and translation, where probabilities estimated from data began beating hand-written grammars. The decisive public moment was 2012: AlexNet, a convolutional neural network trained on 1.2 million labelled photographs across 1,000 categories, won the ImageNet competition with a 15.3% top-5 error rate against 26.2% for the runner-up (Krizhevsky et al., 2012). An 11-point jump in a contest that had been moving a point or two a year settled the argument. By 2015 the same methods reached 3.57% error, below the roughly 5% measured for a trained human annotator (Russakovsky et al., 2015).
Two results from 2017 set up everything since. The transformer architecture (Vaswani et al., "Attention Is All You Need") made training parallel enough to absorb internet-scale text, leading straight to the large language model era and the scaling laws that predict performance from compute and data. And DeepMind's AlphaZero, given only the rules of chess and no human games, reached superhuman play through self-play within 24 hours — deleting the last hand-written component from the domain of symbolic AI's most famous victory.
Types
The standard way to sort AI is by breadth of capability, and only the first category exists.
Artificial narrow intelligence (ANI) covers every system ever deployed: competent in one domain, helpless outside it. Breadth of topic is not breadth of capability — a language model discusses everything while doing exactly one thing, predicting text.
Artificial general intelligence (AGI) would match human ability across essentially all intellectual tasks, including learning a genuinely new skill without being retrained. It does not exist, there is no agreed test for it, and serious forecasts differ by decades.
Artificial superintelligence (ASI) would exceed human ability across the board, and is a subject of philosophical and safety argument rather than engineering.
Real-World Applications
The most persuasive AI in your life is the AI you never think about. Your inbox is readable because a classifier sorts spam and phishing from real mail — Google reports that Gmail keeps more than 99.9% of it out. Your phone camera merges several exposures and decides what the picture should look like using learned models, not optics. Photo apps group faces and let you search for "beach" using computer vision; the keyboard predicting your next word and the captions on a video are trained models too.
Google Maps predicts arrival times using graph neural networks built with DeepMind, which the two reported improved ETA accuracy by up to 50% in cities including Berlin, Jakarta, São Paulo and Tokyo. A card declined at a checkout was scored by a fraud model in milliseconds. Netflix's own engineers have written that around 80% of watched hours begin from a recommendation, not a search.
Then there is the visible work: generative AI that drafts, translates and writes code; voice recognition in assistants and call centres; radiology triage that flags findings before a human reads the scan; protein folding prediction that turned years of lab work into minutes of compute; warehouse robotics; and AI agents chaining tool calls to finish multi-step tasks.
Key Concepts
- Machine learning — the fitting itself: supervised learning from labelled examples, unsupervised learning from structure alone, reinforcement learning from consequences rather than answers.
- Neural networks and deep learning — layered arithmetic that builds its own features from raw input; "deep" only means many layers.
- Foundation models — one large pre-trained model adapted to many tasks, replacing one model per problem.
- Training versus inference — training fits the parameters once at great expense; inference runs the finished model and is what you pay for on every query.
- Multimodal AI — text, images, audio and video in one shared representation, which is why you can paste a screenshot into a chatbot.
- Natural language processing — grammars to statistics to language models: the clearest illustration of the field's whole arc.
Challenges
A fitted model gives no guarantee about any individual answer. Language models produce hallucinations — fluent, confident and wrong — because they were trained to produce plausible continuations, and plausibility only correlates with truth. There is no assertion to check and no stack trace to read, which is why serious deployments wrap the model in retrieval or human review.
Because nobody wrote the rule, nobody can read it. Explainable AI reconstructs after the fact why a model decided what it did, and remains a research area rather than a solved capability — a direct problem anywhere a decision must be justified, in credit, hiring or medicine. Models also inherit their data: bias in AI is usually not a bug someone introduced but a faithful reproduction of a pattern in the corpus — harder to fix because the system is working as designed.
The costs are physical and concentrated: frontier training runs draw enough electricity that AI energy consumption is a grid-planning question, and only a handful of organisations can fund them. Evaluation is straining too — benchmarks leak into training data, so a rising score can mean contamination rather than capability, while AI safety work on alignment chases failure modes discovered after release.
Future Trends
The most consequential recent shift is where the compute goes: rather than only enlarging pretraining, systems now spend effort at answer time — test-time compute, where a model reasons through a problem in steps before replying. Capability costs money per query, not only per training run.
The unit of deployment is moving from a single response to an agent that plans, calls tools, reads results and iterates, which raises the cost of a wrong answer because the output now takes actions instead of appearing on a screen. Capability is also spreading downward: small models distilled from large ones run on phones as edge AI, and open-weight releases keep a usable free tier close behind the frontier.
Finally, the rules are arriving. The EU AI Act entered into force in August 2024, with the bulk of its high-risk requirements applying from 2 August 2026, so AI governance is now an engineering constraint rather than a policy conversation. Expect this page's definitional argument to repeat, too: whatever counts as frontier AI in 2026 will, if it works, be called ordinary software by 2031.