What the Term Means
Artificial intelligence, or AI, is a broad label for computer systems that perform tasks that would normally require human intelligence, such as recognizing speech, translating text, spotting patterns in images or choosing a good move in a game. The phrase dates from the mid-1950s, when researchers gathered at a summer workshop at Dartmouth College to explore whether machines could be made to simulate aspects of learning and reasoning.
There is no single sharp definition. People often call a task “AI” when it seems difficult for a computer, and stop calling it that once it becomes routine. For that reason the term covers a wide range of methods, from hand-written rules to systems that learn from examples.
Two Broad Approaches
Early AI work leaned heavily on explicit rules. Programmers wrote down knowledge as if-then statements, and the program applied them. “Expert systems” of this kind could do useful work in narrow fields, but they were hard to build and tended to break down when situations did not match the rules.
The approach that dominates today is learning from data. Instead of writing every rule, developers give a program many examples and let it adjust itself to perform well on them. This is the territory of machine learning, which is a subset of AI.
How Machine Learning Works in Outline
Though the details are technical, the basic cycle is straightforward.
- Collect examples. For a system that sorts photos, this might be many images labeled by people as “cat” or “not cat.”
- Choose a model. A model is a mathematical function with many adjustable numbers, often called parameters. At the start the numbers are set essentially at random, so the output is poor.
- Measure error. The system makes predictions on the examples and compares them to the correct answers. The gap is summarized as a number, often called the loss.
- Adjust. A training procedure nudges the parameters in a direction that reduces the error. This is repeated many times across the data.
- Test. The finished model is tried on new examples it has not seen. Performance there, not on the training data, indicates how useful it is.
A common kind of model is the neural network, which is loosely inspired by the brain but is much simpler than real neurons. It is built from layers of simple calculations. With enough layers and data, such networks can pick out complicated patterns, such as the shapes that make up a face or the structure of sentences.
Learning comes in a few main flavors. In supervised learning the examples come with correct answers. In unsupervised learning the system looks for structure, such as groups of similar items, without labels. In reinforcement learning a program learns by trial and error, receiving rewards for outcomes, which has been used for game-playing and some control tasks.
Where AI Is Used
AI systems are now built into many everyday tools, often without being noticed.
- Language and speech. Voice assistants, automatic captions, translation and text prediction on phones.
- Images. Photo organizing, face unlock, and reading text from pictures.
- Recommendations. Suggestions for what to watch, read or buy, based on patterns in past behavior.
- Navigation. Route planning and traffic estimates that work alongside GPS navigation.
- Fraud and security detection. Flagging unusual payments or suspicious activity.
- Science and medicine. Helping researchers analyze images and large datasets. In health care, some systems assist clinicians in reading scans, though they are generally intended to support, not replace, professional judgment. The wider context is covered in discussions of modern medicine.
- Generating content. Newer systems can produce text, images and code in response to prompts, usually after training on very large collections of data.
Many of these systems run on large remote computers through cloud computing, which provides the processing power that training and serving big models requires.
Narrow Versus General
Almost every working AI system today is narrow: it is built for a particular kind of task and does not transfer its skill freely. A program that plays chess very well cannot, by itself, book a flight. The idea of a system with broad, humanlike flexibility, sometimes called artificial general intelligence, is a research goal and a topic of speculation, but there is no agreement on whether or when it could be achieved. Related questions, such as whether a program could ever understand anything, are explored in discussions of machine thinking.
Limits and Risks
AI is useful, but it is not magic, and several limits are worth knowing.
Data dependence. A system reflects the data it learned from. If the examples are unrepresentative or contain human prejudice, the system can reproduce those patterns. This is sometimes described as bias in AI, and it parallels the way people are affected by cognitive bias.
Errors. Models can be confidently wrong, especially on inputs unlike their training data. Generative systems in particular can produce plausible text that is inaccurate, so important claims should be checked against reliable sources.
Opacity. Large neural networks are difficult to inspect, so it can be hard to explain why one produced a particular answer.
Privacy and security. Training often involves large amounts of personal or copyrighted material, which raises legal and ethical questions that societies are still working through. Protecting data also depends on tools like encryption basics.
Social effects. AI may change how some jobs are done and how information is produced and trusted. Experts differ about how large and how fast those changes will be, and any specific forecast should be treated cautiously.
A Practical Way to Think About It
A useful mental model is that AI systems are powerful pattern-finders and tools, not people. They can save time and assist with demanding tasks, but their results are best treated as inputs to human judgment, particularly where the stakes are high. Understanding the basics of how they learn makes it easier to use them sensibly and to judge claims made about them.