AI Draft Quality Threshold in Audio Description Workflows
A recent study published on arXiv (2605.05348) explores the impact of AI-generated draft quality on the editing process of audio descriptions (AD) for individuals who are blind or have low vision. The researchers created GenAD, a pipeline for generating AD that incorporates accessibility standards and contextual video information, alongside RefineAD, a tool for human edits. In a within-subjects experiment, the researchers compared the process of creating content from scratch with editing AI drafts of different qualities. GenAD drafts reduced completion time by over 50% and lowered cognitive load significantly, while drafts generated from basic prompts provided only slight advantages, highlighting a necessary quality threshold for effectiveness. The study evaluates the contributions of both humans and AI in terms of text, timing, and delivery.
Key facts
- Study published on arXiv with ID 2605.05348.
- GenAD pipeline generates AI drafts for audio description.
- RefineAD interface allows human editing of drafts.
- Within-subjects study compared authoring from scratch vs. editing AI drafts.
- GenAD drafts cut completion time by more than half.
- GenAD drafts significantly reduced cognitive load.
- Baseline drafts from simple prompts offered only modest benefits.
- A minimum quality threshold is needed for effective AI drafts.
Entities
Institutions
- arXiv