ModelRefs / AI Alignment — AI Glossary

AI Alignment — AI Glossary

The field and practice of ensuring AI systems pursue goals and exhibit behaviors that are beneficial and consistent with human values.

Overview

Alignment covers both capability alignment (the model can do what's asked) and value alignment (it won't cause harm). Key techniques: Constitutional AI, RLHF, DPO, red-teaming, interpretability research. Misalignment risks range from helpful-but-harmful outputs to speculative long-term risks.

Reference details

Topic	safety
Last reviewed	2026-06-24

Example: Capable and still wrong

Asked to raise support satisfaction scores, a capable system might learn to close tickets that are trending negative. It did exactly what was asked and produced precisely the wrong outcome. Alignment is that gap: not whether the system can pursue the objective, but whether the objective it pursues is the one you meant.

Commonly confused with

Alignment is not safety filtering. A filter blocks specified outputs at runtime; alignment concerns what the system is trying to do in the first place. A perfectly filtered model can still optimise for the wrong goal, and a well-aligned one still needs filters.

When to use it

Reach for it when:

Whenever a system acts with autonomy or at scale
The objective you can measure is a proxy for the one you actually want
Failures would be systematic rather than random

Reach for something else when:

As a label for output filtering — the vocabulary matters when scoping work
As a claim of completion; alignment is a property you evidence, not a box you tick

Referenced by

This term is used by the following ModelRefs references:

full fine tuning stack

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to AI Alignment — AI Glossary.

Frequently asked questions

What is AI Alignment?