What Is Artificial Intelligence?

The Definition Problem

Defining AI precisely is harder than it sounds, because the definition keeps shifting. In the 1960s, a calculator was considered a form of artificial intelligence, because arithmetic was a task that previously required human thought. Today nobody calls a calculator AI. This pattern repeats: once a machine can do something reliably, people stop calling it intelligence and start calling it automation.

The working definition most researchers use today is practical: AI is a system that takes input, processes it through a learned or programmed model, and produces output that would require intelligence if a human did it. This includes classifying images, translating languages, generating text, playing games, diagnosing diseases, and navigating physical environments.

That definition is deliberately broad. It covers systems that follow explicit rules (if temperature exceeds 100 degrees, trigger an alarm) and systems that learn from data (given 10 million photos, figure out which ones contain cats). The second category, machine learning, is what most people mean when they talk about AI in 2026.

Narrow AI vs General AI

The most important distinction in AI is between narrow (or weak) AI and general (or strong) AI.

Narrow AI is designed for one specific task. A chess engine plays chess. A spam filter classifies emails. A translation model converts English to French. Each of these systems is extremely good at its designated task and completely useless at everything else. The chess engine cannot write a poem. The spam filter cannot drive a car. Every AI system in production today, including large language models like GPT-4 and Claude, is narrow AI.

This claim about language models deserves explanation, because they seem general-purpose. GPT-4 can write code, answer questions, summarize documents, and generate creative fiction. But its capabilities are bounded by the statistical patterns in its training data. It cannot do anything that requires physical interaction with the world, genuine understanding of causation, or reasoning about situations completely outside its training distribution. It is a very capable narrow AI that happens to be narrow across a wide range of text-based tasks.

General AI (often called AGI, artificial general intelligence) would be a system with the same cognitive flexibility as a human. It could learn any intellectual task a person can, transfer knowledge between domains seamlessly, reason about novel situations from first principles, and understand the world, not just generate text about it. AGI does not exist. There is no scientific consensus on when it will exist, or whether current approaches can lead to it at all.

Some researchers believe AGI is decades away. Others believe current scaling trends (bigger models, more data) will eventually produce it. A smaller group believes it is fundamentally impossible with current architectures. The honest answer is that nobody knows, because nobody has a complete theory of what intelligence actually is.

How Modern AI Actually Works

Modern AI is almost entirely based on machine learning, specifically on neural networks. The process works in three stages: training, evaluation, and deployment.

During training, the system processes large amounts of data and adjusts millions or billions of internal parameters to minimize errors on a specific task. A language model processes trillions of words and learns to predict the next word in a sequence. An image classifier processes millions of labeled photos and learns to distinguish between categories. The math behind this process is calculus and linear algebra, specifically gradient descent applied to differentiable functions.

During evaluation, the trained model is tested on data it has never seen before. This is critical. A model that performs well on its training data but poorly on new data has memorized instead of learned, a problem called overfitting. Researchers use held-out test sets, cross-validation, and benchmark suites to measure how well the model generalizes.

During deployment, the trained model is put into production to handle real inputs. A deployed language model receives user prompts and generates responses. A deployed image classifier receives photos from a security camera and flags potential threats. The model's parameters are frozen at this point, it is not learning from new data unless explicitly retrained.

A Brief History

AI has gone through multiple waves of excitement and disappointment since the field was formally established at the Dartmouth Conference in 1956.

The 1950s and 1960s were optimistic. Early researchers believed human-level AI was twenty years away. Programs could prove mathematical theorems and play checkers. But the optimism was premature. The systems were brittle, worked only on toy problems, and could not scale.

The 1970s and 1980s brought the first "AI winter," a period of reduced funding and interest after the early promises failed to materialize. Expert systems, which encoded human knowledge as if-then rules, had a brief resurgence in the 1980s but proved too rigid and expensive to maintain.

The 1990s and 2000s shifted the field toward statistical methods. Machine learning algorithms that learned from data replaced hand-coded rules. Support vector machines, random forests, and Bayesian methods became the standard toolkit. Progress was steady but not dramatic.

The 2010s changed everything. In 2012, a deep neural network called AlexNet won the ImageNet competition by a massive margin, proving that deep learning with large datasets and GPU computing could achieve results impossible with previous methods. This triggered an explosion of research and investment. By 2017, the transformer architecture was introduced, and by 2018, pre-trained language models (BERT, GPT) demonstrated that massive models could transfer knowledge to almost any text task.

The 2020s are the era of foundation models. GPT-3 (2020), GPT-4 (2023), Claude, Gemini, and their successors showed that scaling up model size and training data produced systems with surprisingly broad capabilities. AI moved from a specialized technical tool to something hundreds of millions of people interact with daily.

What AI Is Not

AI is not conscious. It does not have desires, goals, experiences, or understanding. When a language model writes "I think the answer is 42," the word "I" does not refer to a thinking entity. It is a statistical prediction of what word is likely to follow "I think the answer is" given the model's training data.

AI is not infallible. Every AI system has error rates, blind spots, and failure modes. Language models confidently generate false information (called hallucinations). Image classifiers can be fooled by tiny, imperceptible modifications to pixels. Self-driving systems misinterpret unusual road conditions. Using AI responsibly means understanding where it fails, not just where it succeeds.

AI is not magic. Every capability an AI system demonstrates can be traced to patterns in its training data and the mathematical operations in its architecture. When GPT writes a sonnet, it is because it trained on millions of sonnets and learned the statistical patterns of meter, rhyme, and word choice. The result can be beautiful, but the mechanism is pattern matching at scale, not creativity in the human sense.