As mentioned in the previous section, large language models are the part of generative AI we interact with most, so they are what we'll explain here. Large language models work by predicting text one piece at a time. Trained on enormous amounts of writing, they learn which words tend to follow which others in a given context, and they use those statistical patterns to choose the most likely next word, then the next, and so on.

Context is what steers this. The word "lion," for example, is likely to be associated with "king" in a question about wildlife, but with "captivity" in a question about zoos. This means that the same model can produce different continuations depending on what surrounds the prompt.

The crucial consequence is this: a language model predicts what is likely, not what is true. It has no built-in understanding of meaning, and no way of knowing whether what it produces is correct. As models have access to more information, what is likely and what is true are more often the same, which is why these tools are useful. But not always, which is why you have to stay alert.