How Multimodal Large Language Models Support Access to Visual Information: A Diary Study With Blind and Low Vision People
New research reveals the shocking failure rate of AI-powered visual assistants for the blind.
Deep Dive
A new study shows multimodal AI models are transforming how blind and low-vision people access visual information through conversational assistance. However, in a real-world two-week diary study with 20 participants, the AI failed to deliver reliable help a staggering 33% of the time, either giving incorrect answers (22.2%) or refusing to answer follow-ups (10.8%). While rated as 'somewhat satisfying,' the tech's unreliability highlights a critical gap for everyday use.
Why It Matters
As AI becomes a critical accessibility tool, its high error rate poses real risks and limits for millions of users.