Meeting KI-Postdocs: 'Proxyeconomics and Goodhart's Law' "The downside of optimization" (Dr. Oliver Braganza) Competitive societal systems by necessity rely on imperfect proxy measures. For instance, profit is used to measure economic value, the Journal Impact Factor to measure scientific value, and clicks to measure online engagement or entertainment value. However, any such proxy-measure becomes a target for the competing agents (e.g. companies, scientists or content providers). This suggests, that any competitive societal system is prone to Goodhart’s Law, most pithily formulated as: ‘When a measure becomes a target, it ceases to be a good measure’. Purported adverse consequences include environmental degradation, scientific irreproducibility and problematic social media content. The talk will explore the notion that a systematic research program into Goodhart’s Law, nascent in current AI-safety research, is urgently needed, and will have profound implications far beyond AI-safety. https://www.uni-bonn.de/de/forschung-lehre/forschungsprofil/transdisziplinaere-forschungsbereiche/tra-4-individuals/veranstaltungen-1/proxyeconomics-and-goodharts-law-the-downside-of-optimization https://www.uni-bonn.de/++resource++plone-logo.svg

"The downside of optimization"
(Dr. Oliver Braganza)

Competitive societal systems by necessity rely on imperfect proxy measures. For instance, profit is used to measure economic value, the Journal Impact Factor to measure scientific value, and clicks to measure online engagement or entertainment value. However, any such proxy-measure becomes a target for the competing agents (e.g. companies, scientists or content providers). This suggests, that any competitive societal system is prone to Goodhart’s Law, most pithily formulated as: ‘When a measure becomes a target, it ceases to be a good measure’. Purported adverse consequences include environmental degradation, scientific irreproducibility and problematic social
media content.
The talk will explore the notion that a systematic research program into Goodhart’s Law, nascent in current AI-safety research, is urgently needed, and will have profound implications far beyond AI-safety.