A working definition
Artificial intelligence is the field of building machines that perform tasks we normally associate with human intelligence: perceiving images, understanding and producing language, reasoning toward a goal, and making decisions under uncertainty. That definition is deliberately broad because “intelligence” is a moving target — tasks once considered hallmarks of AI, like beating a grandmaster at chess, become ordinary engineering once solved. A more practical modern definition is narrower: AI is software that learns patterns from data and uses them to make predictions or generate content, rather than following only rules a programmer wrote by hand.
A short history
The field was named at a 1956 workshop, but the conceptual roots go back to Alan Turing, whose 1950 “imitation game” (the Turing test) asked whether a machine could converse indistinguishably from a human. Early decades were dominated by symbolic AI — encoding human knowledge as explicit logical rules. This produced useful expert systems but proved brittle: the real world has too many exceptions to write down. Progress stalled in periods called “AI winters.” The current era took off in the 2010s when deep learning — large neural networks trained on huge datasets with powerful GPUs — began beating older methods at vision and language. The 2017 transformer architecture and the large language models built on it, from GPT to Claude to Gemini, are the latest chapter.
The main branches
It helps to see AI as a few overlapping approaches. Symbolic (rule-based) AI represents knowledge as explicit symbols and logic — transparent and precise, but hard to scale to messy real-world data. Connectionist AI, better known as neural networks and machine learning, learns statistical patterns from examples instead of rules; this is what powers nearly all of today’s headline systems. Hybrid (neuro-symbolic) AI tries to combine the two — the pattern-matching power of neural networks with the reliability and explainability of logic — and is an active research direction for tasks that need both flexibility and correctness.
What AI can and cannot do
Modern AI is superb at narrow, well-defined tasks where lots of data exists: transcribing speech, translating languages, flagging fraud, recommending content, and drafting text. It is poor at tasks requiring genuine common sense, reliable factual accuracy, causal reasoning, or transfer to situations unlike its training data. Crucially, every system in use today is narrow AI — skilled at one domain and unable to generalise. The broader goal of artificial general intelligence, a system as flexible as a human across any task, does not yet exist and remains contested as to when, or whether, it will.
Why the definition keeps shifting
Because “intelligence” is defined relative to humans, AI suffers from a constant goalpost shift: once a capability works reliably, we stop calling it AI and call it software. Optical character recognition and route-finding were once frontier AI; now they are unremarkable. Keeping this in mind makes you a sharper reader of AI news — the right questions are always what specific task does this system do, how reliably, and on what data was it trained? rather than the vaguer is it intelligent?