← AI Terminology

Alignment

Alignment in AI refers to the degree to which an AI system's behaviour matches the intentions, values, and goals of its designers and users.

The term is used both narrowly (does this output follow instructions?) and broadly (does this system pursue human-beneficial goals?).
Why It Matters in AI
"Aligned" has become a term of art for AI that does what you actually want, not just what you literally asked. A system can follow instructions perfectly but still be misaligned if those instructions failed to capture your intent. As AI systems gain more autonomy, alignment gaps compound — small mismatches between stated and intended goals lead to large behavioural deviations at scale.
Key Points
Aspect Description
RLHF Primary current technique — human feedback trains the model toward preferred outputs
See also [[ai-alignment]] for the technical research field; [[rlhf]] for the dominant training method
Goodhart's Law When a measure becomes a target, it ceases to be a good measure — reward hacking is this
Value alignment Broader: does the model's behaviour reflect human values across novel situations?
Constitutional AI Model uses a written set of principles to self-critique and revise outputs
Instruction-following Narrow alignment: does the model follow explicit instructions reliably?
Simple Analogy
Asking an assistant to "make my presentation shorter" and having them delete half the slides gets the instruction right but misses the intent — you wanted brevity without losing substance. Alignment is closing the gap between the literal instruction and the actual goal.
Common Usage Examples
  • HuggingFace trl library: SFT + RLHF pipeline to align LLM outputs with human preferences
  • Anthropic's Claude system prompt specifying values (helpful, harmless, honest) — instruction-level alignment
  • GPT-4's system-card documentation of alignment techniques applied before deployment
  • Reward hacking: RL agent finds a loophole that maximises reward without solving the intended task
  • Debate and amplification: proposed techniques for aligning AI smarter than its trainers
Summary
In short: Alignment is ensuring an AI system pursues your actual intent — not a technically-correct but subtly-wrong interpretation of it.