Alignment is the field of research and engineering concerned with ensuring that AI systems pursue goals that are consistent with human values, intentions, and safety requirements. The core problem is that optimising a system toward a measurable objective — a loss function or reward signal — does not automatically produce behaviour that is beneficial in the broader sense; a model can score well on a metric while behaving in ways its designers never intended. Alignment work therefore spans technical approaches (reward modelling, interpretability, formal verification) and governance approaches (policies, red-teaming, deployment constraints).

Think of it like training a new employee. You can hand them a precise performance target — close 50 tickets a week — but without conveying the underlying values of the organisation, they might achieve that target by closing tickets fraudulently, ignoring difficult cases, or burning out their colleagues.

A concrete alignment technique is Reinforcement Learning from Human Feedback (RLHF), in which human raters score model outputs and those scores train a reward model that steers future generations toward preferred behaviour. Constitutional AI, developed at Anthropic, extends this by having the model itself critique its outputs against a written set of principles before RLHF fine-tuning is applied:

Principle: "Do not assist with actions that could harm a third party."
Step 1 – Model drafts response.
Step 2 – Model critiques draft against principle.
Step 3 – Model revises; revised draft goes to RLHF scoring.

Alignment connects tightly to RLHF, Hallucination (a misaligned model may fabricate rather than admit uncertainty), Fine-tuning (alignment is often applied as a post-training fine-tuning stage), Prompt Engineering (users attempt to override aligned behaviour via prompts), and Foundation Model development where alignment properties are established during Pretraining and instruction tuning.

You will encounter alignment most often in model cards, safety research papers, and in product decisions about what requests an AI assistant will or will not fulfil.