Person
1 episodes · 17 quoted claims
Sarah Bird says Microsoft's first responsible AI standard took effect around 2018, while deep learning began taking off around 2013.
“I think our first responsible AI standard, uh, sort of came into effect about 2018, but if you're, you know, following the timelines of AI, deep learning really started taking off before that in, like, 2013.”
The Gender Shades study showed that facial recognition systems performed worse for dark-skinned women than for other groups, and Bird says such studies helped catalyze responsible AI programs.
“the Gender Shades study that showed facial recognition systems did not work as well for, you know, dark-skinned women as everybody else.”
Microsoft classified emotion detection as an unsuitable use of AI and retired its facial recognition features that predicted emotion.
“we moved these kind of things like emotion detection into a sort of unsuitable use of AI, um, at Microsoft, and then we retired kind of the features we had doing that.”
Microsoft's responsible AI standard divides fairness into three types describing how an AI system behaves.
“Microsoft's responsible AI standard kind of breaks fairness into three types, and the types of fairness are about how the system behaves.”
Bird gives uneven speech-to-text accuracy across accents as an example of a quality-of-service fairness issue.
“your speech-to-text system could transcribe, you know, better for someone with one type of accent than another.”
Bird says the term prompt injection is named after the SQL injection attack.
“It's named after a SQL injection attack if you're familiar kind of with that.”
Microsoft gives its AI red team the first access to new AI technology so responsible AI work begins alongside other exploration.
“we prioritize the very first people getting access are actually our AI red team because we want the responsible AI work to start, you know, before or at the same time as all of the other exploration”
Bird says Microsoft ships thousands of AI features a year and does not embed responsible AI experts in every one of them.
“Microsoft ships thousands of different AI features every year, and so we don't have experts working hand in hand with every single product because not all of them need that level of expertise.”
Bird says Microsoft has hundreds of people working in responsible AI, most in roles that did not exist a few years earlier.
“we have hundreds of people working in responsible AI. Most of those were not jobs even a few years ago”
Microsoft uses a tool, transcribed as Pirate, that expands a red teamer's attack idea into many variants for testing at scale.
“we have a tool called Pirate that helps augment that idea with many different variants or other attack techniques so that you can test many more things at scale”
Traditional security red teams looked only for security vulnerabilities, whereas AI red teams examine a much wider range of misbehavior.
“And so a traditional security red team was just looking for security vulnerabilities.”
Microsoft's AI red team is preparing for risks involving AI producing chemical or biological weapon information.
“We're starting to prepare for new types of risk around ability to produce, you know, chemical or biological weapon information.”
Microsoft already ran responsible AI systems at scale for GitHub Copilot before ChatGPT's release and adapted them to new risks.
“We had already had, you know, at scale, generative AI, responsible AI systems running in production, for example, for GitHub Copilot, and we had to adapt them to new risk.”
Microsoft's AI platform Foundry offers many different models so customers can choose among trade-offs in quality, cost and safety.
“what we try to do with Foundry, which is our AI platform, is offer as many different models as we can.”
Microsoft Research released a project called Magentic UI for experimenting with interfaces that help users understand and control agents.
“that's part of, um, a Microsoft Research project that we've released called Magentic UI.”
Bird says AI technology recently crossed a threshold of being good enough for agentic use cases.
“we've really in the last couple months kind of crossed a critical threshold where this technology is starting to be good enough for these, you know, agentic use cases.”
Bird says agents can go off task, so Microsoft is building new guardrails such as task adherence.
“an agent can get confused and go off task. And so we have to build new guardrails around things like task adherence.”