LECTURES A GRATIS GLOBAL SERVICE
⌕ SEARCH GRATIS GLOBAL ↗
LECTURES
Human-AI Interaction Under Societal Disagreement
SOURCE: YOUTUBE · NO TRACKING UNTIL YOU PRESS PLAY · TROUBLE PLAYING? WATCH AT THE SOURCE ↗

Human-AI Interaction Under Societal Disagreement

53 MIN · EN · STATUS: [ STREAMING ]
RATE THIS
STANFORD

Mitchell Gordon, a computer science PhD student in Stanford's Human-Computer Interaction group, argues that machine learning pipelines handle disagreement badly, typically resolving conflicting annotator opinions through majority vote and erasing minority viewpoints into an aggregate pseudo-human. He presents Jury Learning, an interactive architecture that lets developers explicitly declare which people or groups, and in what proportion, should determine a classifier's prediction, using the metaphor of a jury. He also introduces the Disagreement Deconvolution, a metric transformation showing how standard evaluation measures overstate performance on user-facing tasks like comment toxicity detection by averaging over individuals whose judgments genuinely diverge. Drawing on examples from toxicity detection, poster design, and medical treatment decisions, the talk builds toward a new pipeline for encoding values into human-AI systems that bridges human-computer interaction principles with machine learning practice. Delivered as a Stanford seminar talk, it runs just under an hour.

At a glance

Lecture facts

Runtime compared with the other 387 Computer Science lectures
Runtime53 m
Compared with Computer ScienceShorter than 76%
Source channelStanford Online (YouTube)