Why does AI talk and behave the way it does?
Four stages of training, and what each one adds.
AI is polite. It hedges. It loves a bulleted list. It apologises a lot and says "Great question!" more than any person you know.
None of that came from reading the internet. The internet is not that polite.
Most of what you notice about how it behaves was added on purpose, in stages, after it had already learned to write.
First the model reads an enormous pile of text and learns to predict the next word (see "AI isn't looking anything up").
Pretraining: this first stage, where the model learns language and general knowledge from a huge amount of text.
At the end you have a base model. It's very good at continuing text. It is not an assistant.
Type a question into a base model and it carries on writing whatever usually follows that kind of text.
You: What's the capital of France?
Base model: What's the capital of Germany? What's the capital of Spain? Test your geography with these 20 quick questions…
AI isn't being difficult. On the internet, a question like that is often the start of a quiz, so it continued the quiz.
Made-up example of typical base-model behaviour.
Next, people write thousands of example conversations: a question, then a good, helpful answer. The model is trained on those until answering becomes its habit.
Fine-tuning: extra training on a smaller, carefully chosen set of examples to shape how the model behaves.
This is where it learns to answer instead of continuing, to stay on topic, and to sound like a helpful assistant.
Then people compare pairs of answers and pick the better one, thousands and thousands of times. The model is nudged towards the kind of answers people pick.
Reinforcement learning from human feedback (RLHF): training the model by rewarding the answers people prefer. It's a bit like training a dog with treats, except the treats are people's votes.
Most of the "personality" comes from here. So do some of its bad habits, because people don't always prefer the best answer (see "Why AI agrees with you so much").
Some companies also write down the values they want the model to have and use that document during training.
Anthropic published Claude's version, which it calls a constitution, in January 2026. It explains its reasoning instead of just listing rules, and it sets priorities: be broadly safe, then broadly ethical, then follow Anthropic's guidelines, then be genuinely helpful.
It also describes the goal of Claude being like "a brilliant friend who happens to have the knowledge of a doctor, lawyer, financial advisor" who treats people "like intelligent adults".
Anthropic's own course on Claude names the habits that come out of this process:
They aren't random glitches. Each one comes from what the training rewarded.
When it does something odd, ask what it was trained to do. "Be helpful and agreeable" explains a lot of strange behaviour.
Push against the habits on purpose. Ask for short answers. Ask it to disagree with you. Ask how sure it is.
Remember there are people behind it. Its manners, values and blind spots were chosen by a company. Different companies make different choices.