"This is My Fault", Really? Understanding Blind and Low-Vision People’s Perception of Hallucinati...
summary
Analysis of 3,467 real-world VQA cases (WizWiz plus Ask Vision) uses GPT-4 answers and human coding to show 17% of cases contain hallucinations, categorized into five types — distortion and fabrication plus irrelevance, fallacy, and misfocus — arising across eight usage scenarios. Interviews with 16 BLV users reveal detection under one third, frequent self-blame, varied risk tolerance (healthcare especially low), and expectations for mitigation: framing alerts, auditory/vibration warnings, transparency, verification strategies, and dataset/model fixes for tools like Be My AI, Seeing AI.
AI-written summary · about
chapters (5)
- 0:00Introduction and motivation
- 1:49Study one findings and hallucination types
- 5:45BLV people’s perception of hallucination
- 8:57Design and technical solutions
- 10:00Q&A session
related talks
34mAI and Accessibility: the Good, the Bad, and the Bollocks
33mI Can't See: Low Vision A11y & Users
15mAccessibility: A line of code can be many things
28mVigNET: An intelligent camera app that assists you...
43mThe User Experience of a Blind Person
11mOne Blind Weekend
35mThinking Inclusion with Molly
47mAn Illegally Sighted Look at VR Accessibility