AI Safety & Responsible AI
Adversarial example
An input intentionally crafted to cause a model to make a mistake.
Example
A nearly invisible image change causes a classifier to label a stop sign incorrectly.
Why it matters
It gives testers concrete cases for investigating how a system can be misled.
You'll hear it
“That small change fooled this detector but not the other one.”
Can you control it?
Sometimes
Sometimes. Some products or workflows let you influence an adversarial example, while others keep it fixed or manage it automatically. Check the settings and documentation for the tool you are using.
Common questions
- Must the alteration be invisible to a person?
- No. Some misleading examples are obvious while others are subtle.
- Will one example fool every AI?
- No. Success depends on the system, task and conditions.
- Does resisting one example prove broad protection?
- No. Other changes may exploit a different weakness.