Hanna Wallach says she learned about GPT-4 in the summer of 2022.
“I learned about GPT-4 in the summer of 2022, right as I was about to go out of the office for a couple of weeks.”
Person
Hanna Wallach appeared on the Microsoft Research Podcast to discuss how to measure and reduce harms from AI systems, drawing on her work on Bing Chat and GPT-4. Wallach says she learned about GPT-4 in the summer of 2022, and she was later asked to be research lead for the responsible AI harmful-content work stream on Bing Chat.
She says her team spent eight months developing a framework for measuring fairness-related harms in AI systems. Wallach says many static benchmark datasets lacked clear definitions of what they measured, raising doubts about their reliability and validity. The conversation also covers Microsoft's three-stage approach to responsible AI: identify, measure and mitigate.
1 episodes · 18 quoted claims
Hanna Wallach says she learned about GPT-4 in the summer of 2022.
“I learned about GPT-4 in the summer of 2022, right as I was about to go out of the office for a couple of weeks.”
Microsoft's Ether committee's name stands for AI Ethics and Effects in Engineering and Research.
“Ether stands for AI Ethics and Effects in Engineering and Research.”
Microsoft's Ether committee started a rapid responsible AI evaluation of an early GPT-4 version while Wallach was away.
“Ether kicked off a rapid, responsible AI evaluation of this early version of GPT-4 that was available to us at that point in time while I was out of the office.”
Early testers found GPT-4 showed promise for identifying harmful content.
“some folks had even figured out that it actually showed promise as a tool for even identifying harmful content.”
GPT-4 was integrated into Bing to power what became known as Bing Chat.
“I learned that GPT-4, or the latest version of GPT-4 was being integrated into Bing, specifically to power what would end up becoming known as Bing chat.”
Hanna Wallach was asked to be research lead for the responsible AI harmful-content work stream on Bing Chat.
“I was asked to serve as the research lead for a responsible AI work stream on harmful content.”
Wallach's applied science team spent eight months before Bing Chat work developing a framework for measuring fairness-related harms in AI systems.
“this team had spent the eight months prior to the start of our development work on Bing Chat, developing a new framework for measuring fairness-related harms caused by AI systems.”
Microsoft's Responsible AI Standard specifies a three-stage approach: identify, measure and mitigate.
“Microsoft Responsible AI Standard, which a bunch of us in Microsoft Research were involved in developing, specifies a three-stage approach. So identify, measure, and mitigate.”
Wallach defines stereotyping as generalizations about groups of people that uphold unjust social hierarchies.
“at a high level, stereotyping refers to generalizations about groups of people that uphold unjust social hierarchies.”
Wallach's framework uses harm taxonomies grounded in social science and linguistics to generate templates for test inputs.
“our framework uses taxonomies of harms that are grounded in the social sciences and linguistics to generate templates.”
Wallach's team converted human annotation guidelines into automated guidelines for GPT-4 to evaluate outputs.
“we ended up digging into this further by converting our annotation guidelines for humans into automated annotation guidelines for GPT-4.”
Wallach's team was preparing the Bing Chat harm-measurement work for academic publication.
“we're writing all of this up at the moment for academic publication and we're hoping that some of this stuff will come out over the next few months.”
Wallach says many static benchmark datasets lacked clear definitions of what they measured, raising doubts about their reliability and validity.
“many of them lacked sufficiently clear definitions of the phenomena that were actually being measured, which then in turn led us to question their reliability and their validity as measurement instruments”
Wallach says filtering was long the main way of mitigating harms from AI language systems.
“for a really long time, the main way of mitigating harms caused by AI systems, and this is especially true for harmful content generated by language generation systems, was filtering.”
Early Bing Chat users could evade its metaprompt by asking it to ignore previous instructions.
“early on, users could actually evade Bing chat's metaprompt by simply asking it to ignore previous instructions.”
RLHF stands for reinforcement learning from human feedback, a mitigation in which a model is fine-tuned using human preferences between pairs of outputs.
“RLHF, which stands for reinforcement learning from human feedback.”
Wallach's team collaborates with Natasha Crampton's team in Microsoft's Office of Responsible AI, which provides policy and governance expertise.
“my team collaborates really heavily with Natasha Crampton's team in Microsoft's Office of Responsible AI, who bring policy and governance expertise to our AI ecosystem.”
Sarah Bird's team in Microsoft's AI Platform runs many responsible AI engineering efforts, including integrating OpenAI models into Microsoft products.
“I also collaborate heavily with Sarah Bird's team in AI Platform, who run many of our responsible AI engineering efforts”