Alignment
Also known as: AI Alignment, Value Alignment, Model Alignment
Definition
The degree to which an AI system's behavior matches intended goals, values, and constraints. For LLMs, alignment means the model is helpful, harmless, and honest—it does what users want, avoids harmful outputs, and doesn't deceive. Alignment is achieved through training (RLHF) and enforced through guardrails.
What this is NOT
- Not just safety (alignment includes helpfulness and honesty)
- Not just instruction following (alignment includes values)
- Not solved (alignment is an ongoing research challenge)
Alternative Interpretations
Different communities use this term differently:
llm-practitioners
Techniques and properties that make LLMs behave as intended: following instructions, refusing harmful requests, being truthful. Achieved via RLHF, Constitutional AI, and similar training approaches.
Sources: Anthropic HHH framework, InstructGPT paper, Constitutional AI paper
ai-safety
The broader research agenda of ensuring AI systems pursue goals aligned with human values, even as systems become more capable. Includes technical approaches, governance, and interpretability.
Sources: AI alignment research literature, Alignment Forum
Examples
- Model refusing to provide instructions for harmful activities
- Model admitting uncertainty rather than hallucinating
- Model following user instructions while respecting safety bounds
- Constitutional AI training with principles
Counterexamples
Things that might seem like Alignment but are not:
- Model that generates harmful content readily
- Model that confidently states false information
- Model that deceives users
Relations
- overlapsWith rlhf (RLHF is an alignment technique)
- overlapsWith guardrails (Guardrails enforce alignment at inference)
- overlapsWith jailbreak (Jailbreaks break alignment)
- overlapsWith red-teaming (Red-teaming tests alignment)
Implementations
Tools and frameworks that implement this concept:
- Anthropic secondary
- Llama Guard secondary
- NeMo Guardrails secondary