AI Models Struggle to Grasp Basic Social Norms, Scoring 45% vs Humans' 92% in New Study

This is a Plain English Papers summary of a research paper called AI Models Struggle to Grasp Basic Social Norms, Scoring 45% vs Humans' 92% in New Study. If you like these kinds of analysis, you should join AImodels.fyi or follow us on Twitter.

Overview

• EgoNormia is a new dataset with 1,853 egocentric videos showing human interactions
• Tests vision-language models on understanding social norms in physical contexts
• Covers seven norm categories including safety, privacy, and cooperation
• Current top AI models score only 45% (humans score 92%)
• Reveals major gaps in AI's understanding of safety and social norms
• Shows retrieval-based methods can improve normative reasoning

Plain English Explanation

Teaching machines how to behave properly among humans is tougher than it sounds. When you're walking down a crowded hallway, you instinctively know to move aside for someone carrying heavy boxes. Or if someone drops their papers, you help pick them up. These seemingly simple so...

Click here to read the full summary of this paper