ChatGPT made artificial intelligence part of everyday life almost overnight. Here’s what’s going on behind the chat box, and why it can be brilliant and wrong at the same time.
ChatGPT is an AI chatbot made by OpenAI, a U.S. company. When it launched in November 2022, it became one of the fastest-growing consumer apps ever, and it set off a race among tech companies to build rival chatbots and AI tools.
At its heart is a large language model, or LLM. This is an AI system trained on an enormous amount of text, from books and articles to websites and computer code.
During training, the model is repeatedly shown text with the next word hidden and asked to predict it. Over billions of examples, it adjusts a vast number of internal settings until it becomes very good at predicting what comes next. Along the way, it picks up grammar, facts, writing styles and patterns of reasoning.
That alone produces a system that can continue text, but not necessarily one that is helpful. So the model goes through a second stage of training, where people rate its answers and it learns which responses are useful, accurate and safe. This is what turns a text predictor into an assistant that answers questions and follows instructions.
Training these models takes huge numbers of specialised chips, such as Nvidia’s, running in large data centres, which is why AI has become so expensive to build and so important to chipmakers.
Why chatbots get things wrong, and how they make money
Because a language model generates the most likely response rather than looking facts up in a database, it can produce answers that sound convincing but are wrong. These errors are often called “hallucinations”.
Newer chatbots reduce the problem by searching the web or reading documents before answering, and by showing their sources. But it hasn’t gone away, so it’s wise to check anything important, such as medical, legal or financial information.
Chatbots also only know what was in their training data up to a certain date, unless they’re connected to search, and they can reflect biases in the text they learned from.
ChatGPT now has many competitors, including Google’s Gemini, Anthropic’s Claude and Meta’s AI models. The companies behind them make money mainly through paid subscriptions for individuals and businesses, and by charging developers to build the technology into their own apps and services.
For businesses, the big question is how to use these tools to save time and cut costs, from drafting documents to answering customer questions. For workers, it’s how jobs will change as AI takes on more everyday tasks.
