AI Glossary / Alignment

🏗️ Models and training

Alignment is the collective term for efforts to ensure AI goals and behaviors match human intentions and values - so the model is helpful, honest and harmless, even in situations its creators didn't anticipate. In practice, it's a combination of training techniques (like RLHF), safeguards and stress tests. In the debate about AI's future, alignment also means the broader question: how to ensure increasingly powerful systems do what we actually want - not what a poorly formulated objective literally tells them.