Responsible AI
Unit 6 of 8 · Module 1: Plan and prepare to develop AI solutions on Azure
Read this lesson
Why responsibility matters
Every engineer should think about how their software affects users, and society. With AI, that matters even more. AI systems inform decisions using probabilistic models, and those models depend on the data they were trained on. Their human-like feel makes apps easy to use, but it can also lead people to trust them too much. Incorrect predictions, or misuse, can harm individuals or whole groups. So you need to mitigate those risks, and make sure your system is fair, reliable, and protects people from harm and discrimination. Microsoft uses six core principles for responsible AI.
Fairness
First, fairness: AI systems should treat all people fairly. Picture a model that helps a bank approve loans. Its predictions shouldn't be biased by gender, ethnicity, or anything else that gives some applicants an unfair advantage or disadvantage. There are tools to evaluate, quantify, and mitigate unfairness, but tools alone aren't enough. Build fairness in from the start: review your training data so it represents everyone who might be affected, and evaluate performance for different groups of users throughout development.
Reliability and safety
Second, reliability and safety. Think of an autonomous vehicle, or a model that diagnoses symptoms and recommends prescriptions. If those systems are unreliable, lives are at risk. So AI software needs rigorous testing and deployment management before release. And because models are probabilistic, apply sensible thresholds when you evaluate the confidence scores of their predictions.
Privacy and security
Third, privacy and security. Models are built on large volumes of data, and that data can include personal details that must stay private. Even in production, the system keeps using new data to make predictions and take action. So put safeguards in place to protect data and customer content.
Inclusiveness
Fourth, inclusiveness. AI should empower and engage everyone, and benefit every part of society, regardless of physical ability, gender, sexual orientation, ethnicity, or anything else. One of the best ways to get there: involve as diverse a group of people as possible in design, development, and testing.
Transparency
Fifth, transparency: AI systems should be understandable. Users should know what the system is for, how it works, and what its limits are. Share the factors that affect accuracy, like how many cases the model was trained on, and which features influence its predictions most. Share confidence scores, too. And when an app uses personal data, like facial recognition, make it clear how that data is used, how long it's kept, and who can access it.
Accountability
Finally, accountability: people are accountable for AI systems. Even when a system seems to run on its own, it's the developers who trained and validated the models, and defined the logic that acts on their predictions, who are responsible for it. So work within a framework of governance and organizational principles that makes sure your solution meets clearly defined responsible and legal standards.
Facts from Microsoft's documentation: Responsible AI for Microsoft Foundry; Overview of Responsible AI practices for Azure OpenAI in Foundry Models (CC BY 4.0), rewritten in Loop's own words and examples. Narration, animation, and quiz by Loop. Microsoft, Azure, and Microsoft Foundry are trademarks of Microsoft. Loop is independent and not affiliated with or endorsed by Microsoft.