# Interpretability (AI)

> Looking inside an AI to see how it gets its answers.

**Field:** Artificial Intelligence  
**Category:** Safety & Security  
**Also known as:** AI Interpretability, Model Interpretability

## Think of it like

It is like opening a clock. You look at the gears to see how the hands move.

## Definition

Interpretability is the study of how an AI model gets its answers, by looking inside it. A big model is billions of numbers, and no one wrote down what each one does. Tools trace which parts of the model were most active for an answer, and what those parts stand for. This can show if a model uses a shortcut, hides a goal, or leans on a bias. It is still a young field, and most of a big model is not yet understood.

## Example

Emma's model sorts loan forms. A tool that looks inside shows it leans hard on where the person lives. So her team checks if that hides a bias.

## Why it matters

You cannot fully trust what you cannot see into. Looking inside can catch problems that tests of the answers miss.

## Common mistake

Thinking the model's own reasons are the same thing. A model can give a neat reason that is not what drove its answer.

## Three principles

- **Look inside.** Not just at answers.
- **Trace the parts.** That were active.
- **Young field.** Much still unknown.

## Related terms

- [Mechanistic Interpretability](https://learn.benibix.com/term/ai/mechanistic-interpretability)
- [Neural Network](https://learn.benibix.com/term/ai/neural-network)
- [AI Safety](https://learn.benibix.com/term/ai/ai-safety)
- [Explainable AI (XAI)](https://learn.benibix.com/term/ai/explainable-ai)

## Told apart from

- [Explainable AI vs Interpretability](https://learn.benibix.com/compare/explainable-ai-vs-interpretability): People often think explainable AI and interpretability are the same thing. The first gives a person a reason for one choice, such as what mattered most for a loan. The second looks inside the model to see how it works, part by part. A reason can help and still be a short summary. It may hide how the model truly chose.

![A model drawn as a closed box with a question mark, then the same box opened to show the parts that lit up for one answer.](https://learn.benibix.com/diagrams/sf-interpretability.svg)

---

Source: https://learn.benibix.com/term/ai/interpretability
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.
