Reference · Glossary

Alignment

Shaping AI systems to **follow human intent and values** — helpful, honest, harmless — not just predict the next token.

When to use

Explaining RLHF, safety training, policy layers, and why models refuse some requests.

When not to

Alignment is not perfection — aligned models can still hallucinate or be jailbroken without strong guardrails.

Example

Post-training with human preference rankings teaches the model to prefer helpful, safe answers over toxic completions.