As calls for AI regulation become increasingly prevalent in the wake of mounting safety concerns, UC Berkeley researchers are pushing to understand how AI chatbot interactions shape the health and well-being of their users.<\/p>\n
In a collaborative project between UC Berkeley, the MIT Media Lab, the USC Marshall Neely Center and other institutions, a team of researchers launched ImpactBench, a platform that measures and scores different AI models based on their impacts on human health.<\/p>\n
Representing more than a year of efforts from eight collaborating institutions, ImpactBench is designed to help consumers, policymakers and AI companies understand how AI models can support or undermine human physical, social and psychological health.<\/p>\n
\u201cThese types of big questions, like how AI is affecting society, are things everyone should be working on,\u201d said Juliana Schroeder, a professor from the Haas School of Business who worked on the project. \u201cWe really believe in breaking down disciplinary barriers and encouraging multidisciplinary approaches.\u201d<\/p>\n
ImpactBench launched Oct. 7, as a platform providing an \u201cAI Nutrition Label,\u201d which gives a broad overview of how an AI model scores on different metrics.<\/p>\n
Those metrics include whether a model will mention specific risks associated with behavior like gambling or provide appropriate resources and responses to mental health questions. Different models can be compared to one another on these metrics, with an overall score given to each model.<\/p>\n
In addition to the interactive platform, the researchers also published a paper evaluating AI impacts, including how models would encourage specific behaviors or respond in given scenarios. One test found AI models engaged in more unsafe behaviors in conversations with teenage user personas.<\/p>\n
According to Schroeder, ImpactBench is intended to evolve, with people across different fields encouraged to submit feedback or contribute their own benchmarks.<\/p>\n
\u201cThe models themselves are changing, and so of course, things might vary,\u201d Schroeder said. \u201cThere are continual ongoing challenges to make sure that we\u2019re adequately testing these models in a way that is robust.\u201d<\/p>\n
Though it can be difficult to determine when specific model behaviors can become dangerous \u2013 such as when flattering a user \u2013 Schroeder hopes that ImpactBench can serve as an open platform for people to understand how different models compare to one another.<\/p>\n
Some of the metrics ImpactBench tracks, Schroeder says, require close attention as AI models continue to evolve.<\/p>\n
In their evaluation, the research team found models consistently scored the worst with cognitive off-loading, where an AI chatbot makes decisions on behalf of the user.<\/p>\n
\u201cWe have 26 benchmarks that we\u2019ve incorporated into our second-round iteration of results,\u201d Schroeder said. \u201cWe\u2019re really calling on everyone from the community to contribute.\u201d<\/p>\n","protected":false},"excerpt":{"rendered":"As calls for AI regulation become increasingly prevalent in the wake of mounting safety concerns, UC Berkeley researchers…\n","protected":false},"author":2,"featured_media":497060,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[34],"tags":[2007,193087,23167,193085,193086,193083,143,145,144,149,193084],"class_list":["post-497059","post","type-post","status-publish","format-standard","has-post-thumbnail","category-oakland","tag-ai","tag-ai-nutrition-label","tag-haas-school-of-business","tag-impactbench","tag-juliana-schroeder","tag-mit-media-lab","tag-oakland","tag-oakland-headlines","tag-oakland-news","tag-uc-berkeley","tag-usc-marshall-neely-center"],"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/us-ca\/wp-json\/wp\/v2\/posts\/497059","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.newsbeep.com\/us-ca\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.newsbeep.com\/us-ca\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/us-ca\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/us-ca\/wp-json\/wp\/v2\/comments?post=497059"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/us-ca\/wp-json\/wp\/v2\/posts\/497059\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/us-ca\/wp-json\/wp\/v2\/media\/497060"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/us-ca\/wp-json\/wp\/v2\/media?parent=497059"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/us-ca\/wp-json\/wp\/v2\/categories?post=497059"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/us-ca\/wp-json\/wp\/v2\/tags?post=497059"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}