BEGIN:VCALENDAR
VERSION:2.0
PRODID:Spatie calendar-links
BEGIN:VEVENT
UID:5af5fe068dbf5d8af78db730f0de9cb0
SUMMARY:Telfer Research Seminar Series - Warut Khern-am-nuai
DTSTAMP;TZID=America/Toronto:20261127T100000
DTSTART;TZID=America/Toronto:20261127T100000
DTEND;TZID=America/Toronto:20261127T113000
X-ALT-DESC;FMTTYPE=text/html:_***M.SC. STUDENTS\, THESE SEMINARS CAN COUNT TOWARDS THE SIX MANDATORY\nTELFER RESEARCH SEMINARS SERIES\n[https://telfer.uottawa.ca/en/research/events/] REQUIRED FOR YOUR\nPROGRAM (MGT 6191/ MGT 6991 / MHS 6991) (4 SEMINARS FOR MSC\nPROJECT-BASED STUDENTS).***_\n\nWARUT KHERN-AM-NUAI\, PHD\n\nLarge language models are increasingly used in groups that discuss a\nproblem and then vote or produce a joint recommendation. Group\ndeliberation is often expected to make AI decisions more reliable. We\nfind that it can instead make them less predictable. In a fully\ndeterministic\, self-hosted benchmark\, exact reruns are identical\, but\nsmall wording changes intended to preserve a scenario’s meaning can\ngrow over repeated discussion and often change the final\nrecommendation. In deployed black-box API systems\, nominally identical\ncommittee runs also vary at temperature 0\, a setting many users expect\nto be nearly deterministic. Across 12 policy scenarios\, the results\nshow that instability is not only a consequence of deployment-side\nvariation: repeated interaction can make a committee sensitive to\nsmall differences in how a case is presented. Additional deployed\nexperiments show that role structure\, model composition\, and the\namount of prior discussion agents can revisit all change the degree of\ndivergence. Collective AI therefore presents a stability problem as\nwell as an accuracy problem. Deterministic execution alone does not\nensure that a committee’s recommendation is predictable or readily\nauditable.\n\nABOUT THE SPEAKER[WARUT KHERN-AM-NUAI]\n\nWARUT KHERN-AM-NUAI is the FRQ–IVADO Chair in Economics and Ethics\nof Artificial Intelligence and Professor of Information Systems at the\nDesautels Faculty of Management\, McGill University. His research\ninterests include the economics and ethics of artificial intelligence\,\nplatform for online marketplaces\, and management of information\nsecurity. He earned his Ph.D. in Management Information Systems from\nthe Krannert School of Management\, Purdue University\, in 2016.
LOCATION:DMS 4170
END:VEVENT
END:VCALENDAR