Fine-tuning and human feedback
A pre-trained model can be adapted to particular tasks, languages or ways of answering.
What is it?
A pre-trained model can be further adapted (fine-tuned) for particular tasks, a language or a style. Human feedback is often used too: people rate answers and the model learns which ones they prefer. That makes it more useful and more polite.
An everyday example
When a chatbot politely refuses a dangerous request or answers in neat bullet points, that is largely the result of tuning with human ratings.
Why does it matter?
Tuning decides how a model behaves, what it will not do and what tone it uses. People and their criteria make those decisions.
A common misconception
That a friendly, confident tone means a good answer. A model can learn to please without necessarily being right.
In more detail
What is fine-tuning?
Further training of a model’s parameters on specially selected data.
What is learning from human ratings?
People compare answers or rate examples; that data helps teach the model the desired behaviour.
Is that the same as giving instructions?
No. A prompt changes the current input. Fine-tuning changes the model through extra training.
Why does a model sometimes agree too readily?
Because the ratings and training goals can point in the wrong direction. A model may favour a pleasing answer over a correct one.
Remember: A friendly tone is not proof of a good answer.
Reliable sources
Last reviewed 1 November 2026. An explainer, not professional, legal or medical advice. For rules and figures that change, check the official version.