AI Alignment
In one line
AI alignment is the field — and the practical work — of making AI systems behave in line with human intent, values and safety constraints.
Going deeper
Alignment is the work of making an AI do what we actually want — not just follow instructions literally. RLHF, system prompts, guardrails and safety evaluations are all tools in service of alignment. It is not a single technique but a portfolio of approaches.
Marketers see alignment most clearly in how an AI handles brand-related queries. Better-aligned models tend to be more cautious with facts and less prone to bad guesses; weakly aligned open models hallucinate more often.
A common misread treats alignment as 'censorship', but it is broader than that. Helpfulness, honesty and harmlessness are the three axes, and a model that ignores any one of them is not really aligned.
Related terms
RLHF
RLHF (Reinforcement Learning from Human Feedback) trains an LLM using human preference signals so it produces more helpful, safer responses — the recipe behind the leap in ChatGPT-style quality.
LLMGuardrails
Guardrails are the layer of input/output checks added around an LLM to block unsafe responses, policy violations and leakage of sensitive information.
LLMJailbreak
A jailbreak is a prompt-level trick that bypasses an LLM's safety restrictions to force it into producing content the model is supposed to refuse.
LLMFine-tuning
Fine-tuning takes an already pretrained LLM and trains it further on a narrower dataset to specialise it for a domain, task or voice — the most common path for adapting an LLM to your own data.
GEO·AEOHallucination
Hallucination is when an LLM produces confidently-stated information that is simply wrong — and it is one of the biggest threats to citation accuracy in GEO.
AI AgentConstitutional AI
Constitutional AI (CAI) is Anthropic's alignment technique where the model critiques and revises its own answers against a written set of principles — a 'constitution' — instead of relying entirely on human-labeled feedback.