What Responsible AI Actually Means in 2026 – Microsoft's Sarah Bird

**Ksenia Se** (0:00)
Can I ask a question that really bothers me?

**Sarah Bird** (0:02)
Sure.

**Ksenia Se** (0:03)
Is it even possible to make AI responsible?

**Sarah Bird** (0:10)
The framing of that question matters a lot, because it's actually not about making AI responsible. Like, ultimately, humans should be responsible, right? It's about making AI that is trustworthy, and the worthy part is really important there. There's lots of examples where people are trusting AI, where they shouldn't, and so you don't just make an AI system generically trustworthy. Like, I might trust it to generate a paragraph that then I'm going to edit, right? But I might not trust it to make clinical healthcare decisions for me yet, right? And so trust is very contextual. It both is how you've built the system and where you're putting it in practice. And so I think we've absolutely seen places today where you can build AI that is trustworthy for specific task in specific context. And we're constantly extending that surface so that there's more and more where we feel that AI is fit for purpose, is like actually, you know, doing the job. But you're not going to take any AI tool and just trust it anywhere. And you shouldn't.

**Ksenia Se** (1:18)
And everything happens super fast right now. What has changed the most in responsible AI?

**Sarah Bird** (1:23)
You know, as you said, one is absolutely the pace. We're in a moment right now where, you know, every month or a couple of months, we're seeing kind of a major leap in capability. And that's fantastic because it opens up all of these new types of applications. But responsible AI is about making sure that it works really well in every application, in every context. And so there's, you know, new tools and capabilities we have to develop for each of this kind of new surface area. That is absolutely one. And then something that's very much on my mind right now is the recent breakthroughs we've had in agentic coding, because so much of how we've built responsible AI is based on the current software development lifecycle. And if the software development lifecycle is totally changing, which it looks like it is, how do we build into that new process? And so if we have a world where some organizations are experimenting with and some teams with, okay, agents write all the code and agents review all the code. And so it feels very out of place to then have a human review step inserted in there and taking like three days to review something that took two hours to make. And so I think that that change in how we do human oversight, because we still need it and it's absolutely important and you want that validation, but we need tools to do it. And so I see those as like two of the things that are really different about the moment that we're in, but a lot hasn't changed to the point of we still need human oversight. We're still working on the same problems of testing and having human oversight and all of these things. It's just we're layering more and more sort of interesting things on top of that.

**Ksenia Se** (3:10)
At Build, you announced two open source contributions from your responsible AI team. I think not that many people understand the specifics of responsible AI. Can you explain in the most approachable way, what does these contributions mean?

**Sarah Bird** (3:24)
Yeah.
We find when you are putting responsible AI into practice, and what a developer needs to do, the first thing you need to do is understand what are the risks you're looking for. And you can talk to experts, you can understand your domain. But the next thing you want to do is actually test for those risks. And testing for those risks can be challenging to build high-quality tests. And so the first thing that we're releasing is called Assert. And it's a tool that helps developers go from that high-level idea of what you care about into very detailed tests that really help you understand the behavior of your AI system.
Now, after you run the test, you're going to find that, okay, there's probably some places where your AI isn't exactly behaving the way you want it to yet. And so the next step is going to be putting in controls or mitigations or some people say guardrails, but adjusting the system so that it behaves the way you want it to every time. And so the second piece that we open-sourced is called Agent Control Specification. And what that is, is a kind of a description of the behaviors or the policies you want the agent to follow. And the idea is that every runtime can implement this specification, so no matter where an agent is, you understand what rules it's following and you know how to adjust them and enforce them. And so those are kind of key steps in the Responsible AI process. And we very much believe by open-sourcing tools, we can enable the community to really adopt them and make them their own and extend them, because we very much see Responsible AI as a place where the world is better if everybody does this well. We don't want to be competing with each other on this. We want everybody to succeed here. And open-source is such a great vehicle for that.