What is ChatGPT actually doing when it writes a response?

ChatGPT is predicting the most likely next word in a sequence, one word at a time, based on patterns it learned from massive amounts of text during training. That's the honest answer. It isn't thinking, understanding, or searching a database — it's doing sophisticated pattern matching. When you type 'The capital of France is,' the model calculates that 'Paris' has the highest probability of coming next, based on millions of examples it saw during training. Then it predicts the next word after that, and so on. This is why it sometimes produces confident-sounding nonsense: it's optimizing for what sounds plausible, not what's factually true. I've found this distinction matters a lot in practice. When I ask ChatGPT to write a product description, it works beautifully because plausible-sounding text is exactly what I want. When I ask it to calculate my tax liability, it's risky because plausible-sounding math isn't the same as correct math. The underlying architecture is called a transformer, introduced in a 2017 Google research paper called 'Attention Is All You Need.' The 'attention' mechanism lets the model weigh how much each word in your prompt should influence each word it generates. That's why context matters so much — the model is literally attending to your specific words, not just pulling from general knowledge. Here's a tip: if you want better responses, give the model more context to attend to. Vague prompts produce vague predictions. Specific prompts narrow the probability space and produce sharper output. For a deeper dive, see our guide on [how to write AI prompts](/blog/how-to-write-ai-prompts). **Related**: How does ChatGPT differ from Google Search? | Why does ChatGPT sometimes give wrong answers?