A way of building, evaluating and deploying AI so that it is safe, ethical and trustworthy, carrying on after release through monitoring in production.
Also called RAI.
Read more: Microsoft Learn
In the Ultra Transcenders books
Each book explains Responsible AI in context, with comparison tables and the common traps.
Related terms
- Accountability principle
Under this responsible AI principle, identifiable people must answer for how an AI system behaves; human oversight, override capability and governance committees support it.
- Fairness principle
The Responsible AI principle that AI should treat people fairly, so groups in similar situations aren't affected differently. It means neither high overall accuracy nor giving everyone the same output.
- Inclusiveness principle
The responsible AI principle that AI should engage and empower all people, breaking down barriers with accessibility features, language support and compatibility with screen readers and other assistive technology.
- Model interpretability
Part of the Responsible AI tooling, built on InterpretML, that reports feature importance both overall and for individual predictions, revealing any reliance on sensitive or proxy features.
- Privacy and security principle
Responsible AI principle requiring that business and personal data be protected through notice, consent and limited access. Handling unusual inputs belongs instead to reliability and safety.
- Reliability and safety principle
Responsible AI principle requiring AI to behave as designed under varied conditions, cope safely with missing or unusual input and withstand manipulation.
- Transparency principle
The responsible AI principle that people should be told when AI is involved, understand its capabilities and limits, and see how it arrived at a result, for example why a loan was refused.