What is artificial intelligence?

Defining artificial intelligence

Artificial intelligence is best defined by the tasks it addresses rather than any single technique. Researchers distinguish narrow AI — systems built for a specific task, such as translating text or recognising images — from the hypothetical artificial general intelligence (AGI) that would match human flexibility across arbitrary tasks. All systems in real-world use today are narrow. There is no single agreed test for intelligence; Alan Turing's 1950 "imitation game" proposed conversational indistinguishability as one operational criterion, and debate over suitable measures continues.

Symbolic AI and machine learning

Two broad traditions run through the field. Symbolic AI represents knowledge as explicit rules, logic, and symbols that a system manipulates to reason — the dominant paradigm from the 1950s through the 1980s, including expert systems.

The second tradition, machine learning, derives behaviour from data rather than hand-written rules. Since roughly 2010, machine learning — and in particular deep learning — has driven most measurable progress, as larger datasets and faster hardware made data-driven methods outperform hand-built rules on perception and language tasks.

A brief history

The term "artificial intelligence" was coined for the 1956 Dartmouth Summer Research Project, widely treated as the founding event of the field. Early optimism gave way to periods of reduced funding known as "AI winters" when results lagged expectations. Renewed progress came with statistical methods in the 1990s and the deep-learning resurgence after 2012, when neural networks achieved step changes in image recognition. The 2017 transformer architecture later enabled the large language models that brought AI to wide public attention.

AI in research

In research, AI is both an object of study and a tool. Methodologically, sound AI research demands clearly stated tasks, representative datasets, held-out evaluation, and reported baselines so that claimed gains are reproducible. Reporting standards increasingly call for documenting training data, model limitations, and failure modes. Used as a tool, AI methods support fields from genomics to climate modelling, but their outputs are treated as hypotheses to be validated rather than ground truth.