BEGIN:VCALENDAR VERSION:2.0 PRODID:researchseminars.org CALSCALE:GREGORIAN X-WR-CALNAME:researchseminars.org BEGIN:VEVENT SUMMARY:Raphael Koster (Deepmind) DTSTART:20220209T170000Z DTEND:20220209T180000Z DTSTAMP:20260417T105856Z UID:Metagov/86 DESCRIPTION:Title: Human-centered mechanism design with Democratic AI\nby Raphael Koster (Deepmind) as part of Metagovernance Seminar\n\n\nAbstract\nBuilding arti ficial intelligence (AI) that aligns with human values is an unsolved prob lem. Here\, we developed a human-in-the-loop research pipeline called Demo cratic AI\, in which reinforcement learning is used to design a social mec hanism that humans prefer by majority. A large group of humans played an o nline investment game that involved deciding whether to keep a monetary en dowment or to share it with others for collective benefit. Shared revenue was returned to players under two different redistribution mechanisms\, on e designed by the AI and the other by humans. The AI discovered a mechanis m that redressed initial wealth imbalance\, sanctioned free riders\, and s uccessfully won the majority vote. By optimizing for human preferences\, D emocratic AI may be a promising method for value-aligned policy innovation .\n LOCATION:https://researchseminars.org/talk/Metagov/86/ END:VEVENT END:VCALENDAR