Design systems for every user including people and LLMs | ODSP916

Guust Ysebie argues that accessibility is not just a usability or compliance checkbox in AI-driven systems: it directly affects how reliably LLMs can interpret and reuse content.

Overview

Shared accessibility challenges for humans and LLMs

• Accessibility improves understanding for people and also improves machine interpretability.
• In AI-driven systems, structure, semantics, and clarity become prerequisites for safer and more predictable outcomes.

PDF rendering vs understanding

• The session explains the PDF drawing language and how PDFs are rendered.
• A key point is the difference between:
  • Documents optimized for visual rendering
  • Documents optimized for semantic understanding

Structured PDFs and LLM extraction

• LLMs can extract data effectively when PDFs are structured.
• Tagged PDFs are highlighted as providing HTML-like text structure while preserving visual fidelity.
• Structured PDFs enable better tool development and improve accessibility.

Example document walkthrough

• The presenter reviews a sample PDF containing:
  • Tables
  • Lists
  • A watermark

Why OCR can fail

• Traditional OCR can introduce semantic loss.
• One example discussed is missing sublists due to OCR losing hierarchical structure.

Key takeaway

• Design documents for understanding, not just rendering.

Benefits of accessible, structured PDFs

• Better “smart” data extraction
• Easier search
• More AI-ready content for downstream tooling and LLM-based workflows