Three years ago, a blind person who received an inaccessible PDF had limited options: ask someone sighted for help, run it through patchy OCR, or skip it entirely. Now they paste it into Claude or ChatGPT and get a readable version in seconds.
That shift is quietly becoming one of AI's most concrete, least-hyped use cases. Blind and visually impaired users are adopting AI assistants not as novelties but as daily-driver accessibility tools, filling gaps that traditional assistive technology never closed.
What Actually Works
The use cases fall into three buckets that matter most:
Image descriptions. Screen readers can read alt text, but most images on the web have none, or the descriptions are useless ("image_2024.jpg"). Multimodal AI models - meaning models that can process both text and images - now produce detailed, accurate descriptions of photos, charts, screenshots, and documents. For someone who is fully blind, this turns previously invisible content into usable information.
Inaccessible documents. Scanned PDFs, image-based forms, and poorly structured documents have always been a nightmare for screen readers. AI models handle these reliably now, extracting and reformatting text that OCR (optical character recognition, the older technology for reading text from images) would mangle.
Building custom tools. This is the one that gets overlooked. Coding assistants like Claude Code and GitHub Copilot let blind developers build their own accessible software. When an app's interface does not work with a screen reader, a blind developer can now write a custom version that works exactly how they need it. That kind of autonomy did not exist before.
The Local Model Gap
One tension in this space: the best models for accessibility work - Claude, GPT-4o, Gemini - are all cloud-based and subscription-priced. Local models (AI that runs on your own hardware instead of a company's servers) are not yet matching that quality for image understanding and document parsing. For users who process sensitive documents or want to avoid monthly fees, that gap matters. Open-source multimodal models are improving, but they are not there yet for the accuracy that accessibility use cases demand.
The broader point is worth sitting with. AI companies spend most of their marketing budget talking about productivity and creativity. But for the roughly 2.2 billion people globally with some form of vision impairment (WHO estimate), these tools are not productivity boosters. They are access to information that was previously locked behind a visual interface. That is a fundamentally different value proposition, and it deserves more attention from the companies building these models.