# AI Safety

> The work of making sure AI does not cause harm. It runs from design to daily use.

**Field:** Artificial Intelligence  
**Category:** Safety & Security  
**Also known as:** Safe AI, AI Risk

## Think of it like

It is like the brakes and seat belts on a car. They let you go fast without getting hurt.

## Definition

AI safety is the work of making sure AI systems do not cause harm. It covers small harms, such as a chatbot giving bad advice, and big ones, such as a powerful system acting against people. Part of it is building models that want the right things. Part of it is testing them for harm before release. And part of it is guarding them once they are live, so a mistake or an attack is caught early.

## Example

Before a new model ships, Lucy's team tests it with hundreds of risky requests. They fix what it gets wrong, then add filters that watch it once it is live.

## Why it matters

AI is used for health, money and code. A safe model is one people can trust with real work.

## Common mistake

Treating safety as one filter added at the end. It has to be part of how the model is built and tested.

## Three principles

- **Build it right.** Want the right things.
- **Test for harm.** Before release.
- **Guard it live.** Catch problems early.

## Related terms

- [AI Alignment](https://learn.benibix.com/term/ai/ai-alignment)
- [Guardrail (AI)](https://learn.benibix.com/term/ai/guardrail)
- [AI Red Teaming](https://learn.benibix.com/term/ai/ai-red-teaming)
- [Interpretability (AI)](https://learn.benibix.com/term/ai/interpretability)
- [Responsible AI](https://learn.benibix.com/term/ai/responsible-ai)
- [AI Security](https://learn.benibix.com/term/ai/ai-security)
- [Existential Risk from AI](https://learn.benibix.com/term/ai/existential-risk)

## 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.
- [AI Safety vs AI Security](https://learn.benibix.com/compare/ai-safety-vs-ai-security): People think AI safety and AI security are two words for the same job. They are not. Safety asks if the AI itself could cause harm, through a wrong goal or a mistake. Security asks if an attacker could make it cause harm, with a trick or stolen access. A model can be well meaning and still easy to attack, so a team needs both.

![Three layers around a model: build it to want the right thing, test it for harm, and guard what goes in and out once it is live.](https://learn.benibix.com/diagrams/sf-ai-safety.svg)

---

Source: https://learn.benibix.com/term/ai/ai-safety
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.
