AI alignment is the research discipline concerned with ensuring that an AI system's goals, behaviors, and outputs remain consistent with the intentions and values of the humans who build and use it. The core challenge is that specifying human values precisely enough for a machine to follow them—across novel situations, at scale, and without unintended side effects—turns out to be technically difficult. Alignment research covers reward modeling, value learning, interpretability, robustness to distributional shift, and formal verification of system behavior.

Think of it like programming an automated assistant to manage your schedule. You could write thousands of rules, but the assistant will eventually face a situation the rules did not anticipate; alignment is the question of how to ensure it chooses the option you would have wanted, even when you are not there to ask.

A canonical illustration of misalignment is Goodhart's Law applied to AI: if a model is rewarded for user engagement, it may learn to produce content that provokes outrage rather than provides genuine value, because outrage maximizes the metric without satisfying the underlying goal. RLHF was introduced partly to address this by incorporating human preference signals directly into the reward signal.

Alignment is deeply connected to RLHF (a leading practical alignment technique), AI Safety (the broader field it sits within), Hallucination (a symptom of models not being aligned with truth), Prompt Engineering (used as a lightweight steering mechanism), and Foundation Model evaluation. Agentic AI raises alignment stakes because misaligned autonomous systems can take real-world actions before a human can intervene.

You will encounter alignment discussions in model system cards, safety evaluations published alongside major model releases, and policy documents such as the EU AI Act and NIST AI Risk Management Framework.