AI là gì?

Artificial intelligence is the effort to make machines do things that would require judgment if a person did them: recognize a face, translate a sentence, recommend a film, drive a car or answer a question. Because the phrase covers so much, it helps to use a precise definition. The European Union's AI Act, whose definitions have applied since 2 February 2025, describes an AI system as a machine-based system that operates with some autonomy, may adapt after deployment, and infers from its input how to generate outputs such as predictions, content, recommendations or decisions that can affect physical or virtual environments. The key word is infers. A calculator follows fixed steps; an AI system works out an output that nobody spelled out in advance for that exact input.

The field has two broad families of methods, and most real products mix them. The first is symbolic AI, sometimes called rule-based or classical AI. People write down knowledge as rules and the program searches through possibilities. IBM's Deep Blue is the textbook case: it evaluated about 200 million chess positions per second using 32 processors and a hand-tuned scoring function. The second family is machine learning, in which the program is not told the rules but learns them from examples. A spam filter trained on millions of labeled emails, or a model that predicts delivery times from past trips, works this way. Deep learning is the branch of machine learning that uses neural networks with many layers, and since the early 2010s it has driven most of the visible progress in speech, vision and language, as the 2015 Nature review by Yann LeCun, Yoshua Bengio and Geoffrey Hinton describes.

A modern machine learning system is built in a fairly fixed sequence. Someone defines the task and collects data, such as photos with labels or text from the web. They choose a model, a mathematical function with adjustable settings called parameters. Training repeatedly adjusts those parameters so the model's outputs match the examples better, as judged by a score called the objective or loss. The trained model is then tested on data it has never seen, because doing well on the training examples alone proves little. Finally it is deployed inside a product, where it receives real inputs, and its behavior is monitored because the world drifts away from the training data over time.

The history runs back more than seventy years. In 1950 Alan Turing opened his paper Computing Machinery and Intelligence with the question of whether machines can think, and proposed replacing it with a test, the imitation game, in which an interrogator chats with a hidden human and a hidden machine. He guessed that within about fifty years a machine could fool an average interrogator often enough that they would have no better than a 70 percent chance of a correct identification after five minutes. On August 31, 1955, John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon dated their proposal for a summer research project at Dartmouth College on artificial intelligence, and the 1956 workshop that followed gave the field its name. Later milestones include Deep Blue's 1997 chess win, the 2012 to 2015 deep learning breakthroughs, AlphaGo's 2016 Go victories, and ChatGPT's public release on November 30, 2022.

People often sort AI by how broad it is. Narrow AI does one kind of task, such as playing chess or transcribing speech, and every system in commercial use today is built and tested for particular tasks. Artificial general intelligence, or AGI, refers to a hypothetical system that could learn and perform most intellectual tasks a person can; there is no agreed test for it and no agreed date for it. Generative AI is a newer label for systems that produce text, images, audio or code. It is not a separate kind of intelligence but an application of deep learning. Robotics adds sensors and motors so the outputs act on the physical world.

It also helps to compare AI with things it is often confused with. Ordinary software follows explicit instructions and gives the same answer every time; AI systems can handle messy inputs but can also be wrong in ways that are hard to predict. Automation means any machine doing a task without a person, and much automation, such as a timer on a sprinkler, contains no AI at all. Human intelligence is general, grounded in physical experience and able to learn from a handful of examples; most AI systems need large amounts of data and can fail when inputs differ from what they saw in training. AlphaGo could beat a professional at Go but could not play chess without being rebuilt and retrained.

The limits are practical, not abstract. OpenAI's own launch note for ChatGPT warned that it sometimes writes plausible-sounding but incorrect answers, changes its answer when a question is rephrased, guesses at ambiguous requests instead of asking, and can still respond to harmful instructions or show biased behavior. Learned systems absorb the patterns, and the gaps, in their training data, so they can treat groups of people unequally. They can be fooled by inputs designed to trick them. And an AI system has no built-in sense of when it is outside its competence, so the people deploying it must decide where it is safe to rely on.

Using AI is cheap for individuals and expensive to build at the frontier. The Stanford AI Index 2025 reports that the cost of running a model at GPT-3.5's level fell more than 280-fold between November 2022 and October 2024, which is why capable assistants are now free or low cost. At the same time, US private investment in AI reached 109.1 billion dollars in 2024, and nearly 90% of notable new models came from companies rather than universities, because training the largest systems takes specialized chips, large data centers and large datasets. For most organizations the realistic cost is not building a model but integrating an existing one, checking its outputs, and training people to use it well.