Alignment

Alignment is the effort to make a model’s behavior match what humans intend and value — not just what’s statistically likely.

It covers:

• Harmlessness — not producing dangerous or abusive content.
• Helpfulness — actually doing what was asked.
• Honesty — admitting uncertainty rather than fabricating.

RLHF and other techniques are the main tools. It’s as much a research field as an engineering one.