Artificial intelligence (AI) is a branch of computer science that develops systems able to perform tasks that often require human intelligence. These tasks include recognizing speech, finding patterns, writing text, and making predictions. Many modern AI systems learn from large amounts of data rather than relying only on fixed rules written by a programmer.
AI is often presented as a tool that can answer questions, create images, and assist doctors or teachers. Behind these services is an industry built on data, advanced computer chips, large data centers, and human labor. Understanding how AI is made helps explain both its abilities and the power of the companies that control it.
Analysis
Human intelligence develops through biological processes, personal experience, and social relationships. AI is designed by people and runs on computer systems. It can process information at a speed and scale that people cannot match. However, fluent language alone does not establish that a system has human understanding, feelings, or moral judgment. Its abilities must be judged by what it can reliably do, including where it fails.
AI production begins with data. Companies gather material from websites, books, public records, licensed collections, and other sources. Depending on the product and its settings, they may also use customer interactions. Much of this information was created by writers, artists, journalists, and ordinary people. Its use raises questions about permission, payment, ownership, and privacy. A Stanford study of major AI developers’ privacy policies found that user conversations could be used for training, with important differences in the choices and protections offered to users.
Different forms of AI use data in different ways. A system trained to recognize images may learn patterns that distinguish a damaged part from a working one. A chatbot uses a language model trained on patterns in text. The process that turns training data into an answer has two main stages: training the model and using the trained model to respond to a question.
During training, text is divided into small units called tokens. These may be words, parts of words, or punctuation marks. The model learns by predicting the next token in a passage and comparing its prediction with the actual text. Training software then adjusts the model’s internal numbers, called weights, to reduce prediction errors. Repeating this process across large amounts of text builds patterns that help the model handle language, facts, and relationships between ideas. Further training with examples and feedback helps it follow instructions and produce more useful responses.
When a person asks a question, the system converts the question into tokens and numerical values. The model processes them alongside the available conversation and the instructions that guide its behavior. A method called attention helps it weigh relationships between parts of this text. For example, it can connect a request to compare two policies with the descriptions of those policies earlier in the conversation.
The model then calculates the likelihood of different possible next tokens and selects one. It repeats this calculation, using the text already produced to help choose what comes next. These repeated steps build the full answer. The model is applying patterns stored in its weights; it is not searching through all its training documents each time. Some training material can also be memorized, which creates risks involving privacy and copied text.
Some AI services add web searches, uploaded documents, calculators, or other tools to this process. Information from those tools becomes part of the material the model uses to form its response. This can improve accuracy and provide current information, but it does not guarantee a correct answer. Without those tools, the model relies on its training and the information supplied in the conversation.
This process explains why an answer can sound convincing while containing false information. Predicting a likely continuation of text does not by itself verify that every claim is true. NIST identifies fabricated facts and sources, privacy risks, and harmful bias among the concerns raised by generative AI. The quality of the data, the training choices, and the checks applied before and after release all affect the reliability of a system.
The next requirement is computing power. Large models are trained using many advanced chips, and computing resources are needed again whenever users request answers. NVIDIA is a major supplier of these chips. Large cloud providers, including Amazon, Microsoft, and Google, sell access to the computers and data centers that developers need. This dependence can make it expensive for universities, public agencies, and smaller firms to build independent alternatives.
AI companies earn revenue through subscriptions, business contracts, and fees for access to their models. Cloud providers sell computing services, while chip makers sell the hardware. The Center for Democracy & Technology has also examined companies’ efforts to earn money through advertising, commerce, and government contracts. These business choices matter because they shape what products are built, whose needs are served, and how user information may be used.
The environmental costs are shared by the public through higher electricity demand, water use, pollution, and pressure to build new power infrastructure.
Take Action
- The Electronic Frontier Foundation: Advocates for privacy, free expression, and public accountability in AI policy.
- The Center for Democracy & Technology: Examines how AI business models and design choices affect people’s rights.
- NIST: Provides guidance on evaluating AI systems and managing their risks.