Embracing Real-World Pluralism: Moving beyond "gold standard" truth in AI evaluation
summary
Argues that single gold standard labels fail for safety and alignment and proposes treating disagreement as signal via Data Cascades analysis, concrete frameworks (crowdroot, GRaSP) and metrics (CrowdTruth, DIVE) that become pluralistic classifiers. Shows a real annotation split — 18 safe, 15 unsafe, 4 unsure with nine vs six different reasons — and reports building datasets including an adversarial nibbler corpus with ~20,000 prompt–image pairs and demographic trisections that reproduce demographic sensitivity in benchmarks.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 0:51Data quantity to cultural awareness
- 3:51Binary assumptions in evaluation
- 6:18Context in data quality and safety
- 9:06Ignoring diverse perspectives
- 10:33Pluralistic AI research milestones
- 13:07AI for a pluralistic world
- 14:58Three pillars for pluralism
- 16:46Diverse data collection at Google DeepMind
- 18:33Disagreement across demographics
- 19:37Metrics for pluralistic evaluation
- 21:00Cultural awareness benchmarking
- 22:29Future directions in pluralistic AI
- 23:11Q&A
related talks
39mThe Search for Transparency and Accountability in the Age of...
41mWhen AI Failures Reflect Our Own Social Biases
23mDeveloping an intersectional framework to analyze biases in artificial intelligence and deep neural
01h 16mLifting the Hood on Big AI - The Future of Transparency and Accountability in AI
31mOperationalizing Responsible AI
45mThe Myth of Neutrality: How AI is widening social divides
42mRethinking AI Alignment and Evaluation under Heterogeneous Preferences
44mAI, Corporate Responsibility & Democratic Legitimacy