Particle.news

Download on the App Store

AI Vision Models Struggle with Basic Visual Tasks

New research reveals significant limitations in AI's ability to perform simple visual reasoning, despite advanced claims.

Overview

  • Researchers from Auburn University and the University of Alberta tested top AI models on basic visual tasks.
  • AI models like GPT-4o and Gemini 1.5 Pro failed at simple tasks such as identifying overlapping shapes and counting objects.
  • The study highlights that AI's visual 'understanding' is more about pattern recognition than true visual comprehension.
  • Models performed well on familiar patterns like the Olympic Rings but struggled with slight variations.
  • The findings suggest that current AI models lack generalizable visual reasoning capabilities.