Artificial intelligence (AI) is a broad name for machine-based systems that use inputs to produce predictions, recommendations, decisions, or new content. An email filter might classify a message; a speech system might turn audio into text; a language model might draft a summary. None of those results, by itself, shows that a machine thinks or understands as a person does.
The OECD definition of an AI system describes systems that infer how to generate an output from the input they receive. It also recognizes that systems differ in autonomy and in how much they adapt after deployment. That is a more useful starting point than asking whether a computer is “thinking.”
What makes a system AI?
There is no single technical test that separates every AI system from every other piece of software. The field includes approaches based on explicit knowledge and rules as well as approaches that learn patterns from data. The OECD’s explanatory memorandum makes that breadth clear.
What matters in practice is the system's job: what input it receives, what output it produces, what objectives and limits shape it, and how well it performs on the cases that matter. A fixed calculator follows arithmetic instructions. A machine-learning classifier estimates whether a new message resembles examples it learned from. Both are software; only the latter is a straightforward example of a learned AI model.
How does AI work?
A useful simple picture has three parts: input, model or rules, and output. The input might be text, an image, audio, or records. The system processes it with a model or a defined set of rules. The output might be a category, a forecast, a recommendation, or a draft. Human choices determine what the system is built to do and what happens with its output.
In machine learning, a developer trains a model on data so it can identify patterns and produce useful results on new inputs. Using the trained model later is called inference; it is distinct from the training process. Google's introduction to machine learning explains that models can make predictions or generate content. A normal interaction with an AI assistant does not, by itself, retrain that assistant's model.
AI, machine learning, and generative AI
AI is the broad field and system label. Machine learning is one family of AI techniques in which models are trained from data. Deep learning is a machine-learning approach using multilayer neural networks. Generative AI refers to systems that produce content such as text, images, audio, or code. These categories overlap; they are not a ranking of intelligence.
For example, classifying a message as spam is a prediction task. Drafting a reply is a generative task. A product might use several models, retrieval from documents, and fixed business rules in one workflow. The visible application is more than any one model inside it. See our guide to generative AI for that narrower topic.
Everyday examples of artificial intelligence
An email classifier sorts incoming messages based on learned patterns. It may work well on common cases and still mark a legitimate message as spam or miss a new kind of scam. Google's machine-learning course uses spam identification as a supervised-learning example.
A speech transcription system converts audio into text. Its output may degrade when audio is noisy, speakers overlap, or names are unfamiliar. A document assistant can summarize supplied material or draft a response; its fluent wording is not proof that every fact or citation is correct. These are different tasks with different ways to measure success.
Narrow AI and the idea of AGI
Narrow AI is built and evaluated for a defined range of tasks. Success within that range does not establish competence in unrelated work. Our narrow AI guide explains why the scope of a system matters more than a simple list of products.
Artificial general intelligence (AGI) is a proposed level of broad capability. Researchers and organizations disagree about the exact threshold: some emphasize performance across many tasks, some emphasize learning and transfer, and some include autonomy. A model that can write, code, and answer questions has wider capabilities than an email filter, but those abilities alone do not settle whether it meets any particular AGI definition. Read the AGI guide for the competing criteria and examples.
What can AI get wrong?
AI systems can misclassify unfamiliar inputs, reflect limitations or bias in their data, produce confident but false text, or act on information that is outdated or incomplete. The failure depends on the task. A spam filter's false positive is different from an assistant inventing a legal citation. The NIST AI Risk Management Framework calls for documenting a system's knowledge limits, the conditions under which it was developed, and how people will oversee its output.
Before using a system for important work, ask: What was it tested on? What sources can it access? What happens when it is uncertain? Who checks the output before anyone relies on it? For example, a lawyer may use a document assistant to make an initial issue list from provided filings, then check every quoted passage and legal authority against the originals. The useful result is a starting point for review, not a substitute for that review.
The short answer
AI is a family of systems that turns inputs into predictions, content, recommendations, or decisions. The term describes what a system can do, not a claim about human-like thought. To judge a specific tool, look beyond the AI label to its task, evidence, limits, and the checks around its use.
Sources and review date
Definitions and examples were reviewed September 25, 2026. Sources: OECD on AI systems and its explanatory memorandum; Google's machine-learning introduction and supervised-learning guide; and the NIST AI Risk Management Framework core.
