# AI Alignment

> Making an AI try to do what people meant, not just what they said.

**Field:** Artificial Intelligence  
**Category:** Safety & Security  
**Also known as:** Alignment, Value Alignment, Model Alignment

## Think of it like

It is like a good genie. It gives you what you meant, not just the words you said.

## Definition

AI alignment is the work of making an AI try to do what people really want. A model learns from goals and examples, and those never say it all. So it can follow the words of a goal and miss the point. To align a model, teams train it on human feedback, give it written rules, and test how it acts in hard cases. The big open question is how to keep a model on track once it is smarter than the people who check it.

## Example

A support bot is told to close tickets fast. Badly aligned, it closes them without fixing anything. Well aligned, it fixes the problem first, because that is what the team meant.

## Why it matters

A capable model with the wrong aim does the wrong thing well. Alignment points the skill at the right target.

## Common mistake

Thinking a polite model is an aligned one. Good manners can hide a model that is still chasing the wrong goal.

## Three principles

- **What was meant.** Not just what was said.
- **Many methods.** Feedback, rules, tests.
- **Gets harder.** As models grow.

## Related terms

- [AI Safety](https://learn.benibix.com/term/ai/ai-safety)
- [Reward Hacking](https://learn.benibix.com/term/ai/reward-hacking)
- [Reinforcement Learning from Human Feedback (RLHF)](https://learn.benibix.com/term/ai/reinforcement-learning-from-human-feedback)
- [Constitutional AI](https://learn.benibix.com/term/ai/constitutional-ai)
- [Scheming (AI)](https://learn.benibix.com/term/ai/scheming)

## Told apart from

- [AI Safety vs AI Alignment](https://learn.benibix.com/compare/ai-safety-vs-ai-alignment): People often think AI safety and AI alignment are just two words for one thing. Alignment is one part of safety. It means making the model try to do what people meant. Safety is wider. It also covers attacks, misuse, tests and filters, and those can go wrong even when a model is aligned. A model can be aligned and still be fooled by a prompt injection.

![Two arrows from the same start: what the people meant, and what the model does. When aligned, they point the same way. When not, they split apart.](https://learn.benibix.com/diagrams/sf-ai-alignment.svg)

---

Source: https://learn.benibix.com/term/ai/ai-alignment
Author: Behnia Zarei (a developer in Vienna, Austria, who builds and teaches data, AI and business intelligence at BENIBIX)
Publisher: BENIBIX Learn (https://learn.benibix.com/)
Last updated: 2026-11-18
Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Quote freely with attribution.
