Summary
PDF accessibility is a long-standing challenge. Many documents are hard to navigate, expensive to fix, and difficult for assistive technology to interpret. This article explains why traditional tagging approaches fall short and how AI-powered semantic translation can turn PDFs into accessible Markdown. It also shows how this method supports autonomy, reduces barriers, and creates documents that work for more people.
This article is based on a presentation by Blake Bertuccelli Booth and Dylan Isaac from the University of Illinois Chicago (UIC) at A11yNYC. They presented Equality Reflow, an open-source pipeline that converts PDFs into accessible content through Markdown.
The broader challenge of document accessibility
PDF accessibility remains a major barrier for many organizations. PDFs were originally designed for printing, not for digital access. They often lack the semantic structure that screen readers and other assistive technologies rely on. This creates a difficult experience for people who depend on clear navigation, headings, and consistent reading order.
Many institutions hold large collections of PDFs. Some collections include hundreds of thousands of files. Manual remediation is expensive and slow. Costs can reach several dollars per page, which makes large-scale remediation difficult and expensive. This creates pressure to find a sustainable solution that supports both compliance and inclusion.
The challenge goes beyond technical as it’s also about autonomy and dignity. People shouldn’t struggle with inaccessible documents or rely on workarounds. They should be able to read, navigate, and interact with information in ways that work for them. This requires a shift from fixing PDFs to rethinking how documents are created and shared.
The approach described here focuses on translating PDFs into accessible formats rather than trying to repair them. It uses artificial intelligence (AI) to interpret visual structure and convert it into semantic code. This method supports more consistent results and reduces the burden on individuals and teams.
Why PDF accessibility is so difficult
PDFs were created to preserve the visual layout. They weren’t built with semantic structure in mind. A semantic structure is the underlying meaning of a document, such as headings, lists, tables, and relationships between elements. Assistive technology depends on this structure to create usable experiences.
Because PDFs lack a consistent structure, screen readers often struggle with reading order. Two-column layouts, decorative elements, and unlabeled images add barriers. Even when tools attempt to tag PDFs automatically, it often produces incomplete or inaccurate results.
Another challenge is ownership. PDFs are controlled by a proprietary company. This limits how much the format can evolve. It also makes it harder for organizations to build their own solutions. Many teams try to work around these limitations, but the format itself remains a barrier.
These challenges show why a new approach is needed. Instead of trying to force accessibility into a format that was not designed for it, organizations can shift to formats that support accessibility by default.
Why Markdown offers a more accessible foundation
Markdown is a plain text format. A plain text format is a file that can be opened and edited in any basic text editor.
Markdown uses simple symbols to represent structure.
- # represents a heading
- [] and () are used to represent links
- – represent list items.
Markdown is considered a democratic format because it does not require proprietary software. Anyone can open, edit, and understand it. This supports autonomy and reduces barriers for people who use assistive technology.
Markdown is also semantically rich. A semantically rich format conveys meaning through structure. Markdown maps directly to HTML, which is the foundation of accessible web content. Rendered Markdown becomes structured HTML that screen readers can interpret.
Another advantage is that Markdown works with AI systems. Many AI models get trained on Markdown examples. This makes it easier for AI to generate accurate, structured output. It also allows AI to act as a bridge between visual layouts and semantic code.
The flexibility of Markdown allows it to be converted into many other formats. It can become HTML, a Canvas page, or a WordPress post. This flexibility supports different workflows and reduces the need for multiple tools.
How AI acts as a semantic translator
AI can interpret visual information and convert it into a semantic structure. A semantic translator is a system that understands the meaning behind visual elements and expresses that meaning in code. This goes beyond extracting text. It requires understanding relationships, hierarchy, and intention.
Multimodal AI models process both images and text. This allows them to interpret visual cues such as font size, alignment, and proximity. For example, large, centered text at the top of a page represents a title. Underlined blue text represents a link. Small text near an image represents a caption.
AI groups related concepts together. It identifies patterns across languages and formats. This helps it understand the meaning behind visual elements even when the layout is complex. It can also identify when something is decorative or informative.
However, AI can’t do it alone. AI systems haven’t been trained with accessibility as a core requirement. They can make mistakes or overlook important details. This shows why a structured process is needed to guide the AI and support consistent results.
This approach uses an AI harness. An AI harness is a system that guides the AI through specific steps. It provides context, instructions, and checks to reduce errors. This creates a more reliable process for document accessibility.
How the Reflow process works
The Equalify Reflow process uses five steps to convert PDFs into accessible Markdown. Each step supports accuracy, structure, and consistency.
1. Extracting document text
The first step uses a tool that pulls text from the PDF. It also removes two-column layouts that screen readers struggle to read. This step captures most of the content but does not add structure.
2. Analyzing each page
The system identifies the type of document and the elements it contains. It notes headings, images, tables, and other features. This information guides the AI in later steps.
3. Building the heading structure
A clear heading structure is essential for navigation. The system reviews each heading and adjusts it based on context. It logs every decision so a human reviewer can verify the results.
4. Translating visual intention
Each page is processed by its own AI agent. This keeps the task focused and reduces errors. The AI interprets visual cues and converts them into semantic Markdown or HTML. It also corrects issues such as OCR errors, missing italics, and broken tables.
5. Assembling the final document
The system combines all pages into a single continuous document. It moves footnotes to the bottom and removes unnecessary layout elements. The result is a flexible document that supports zooming, reflowing, and customization.
Real-world examples of the process
The system handles a wide range of document types. Academic papers with two-column layouts become single-column documents that assistive technology can read. Complex tables are reconstructed using semantic HTML. Posters with unconventional layouts are converted into structured content with clear headings and event details.
The system also generates image descriptions. It uses a dedicated subagent to determine whether an image is decorative or informative. It considers context, surrounding text, and the purpose of the image. This supports a more complete and inclusive experience.
These examples show how AI can interpret visual language and convert it into an accessible structure. They also show how a guided process reduces errors and improves reliability.
What organizations and individuals can do next
Organizations can begin by identifying the scale of their PDF collections. This helps determine where automated solutions can have the greatest impact. They can also explore tools that convert PDFs into accessible formats rather than trying to repair the PDFs themselves. Organizations that are interested in working with Equalify Reflow can sign up for updates on the project.
Teams can integrate Markdown into their workflows. Markdown supports accessibility by default and reduces reliance on proprietary tools. It also makes it easier to convert content into multiple formats.
Organizations can test AI-powered tools on a variety of documents. This helps identify patterns, edge cases, and opportunities for improvement. It also supports a more inclusive approach to document creation and sharing.
Finally, teams can prepare for upcoming accessibility requirements by adopting tools that scale. Automated semantic translation reduces cost and increases consistency. It also supports autonomy for people who rely on accessible documents.
Moving toward a more inclusive document ecosystem
PDF accessibility has been a long-standing challenge. The format was not designed for the needs of today’s digital environment. By shifting to semantic translation and accessible formats, organizations can create documents that work for more people.
AI-powered tools offer new possibilities. They can interpret visual structure, correct errors, and generate accessible output at scale. When guided by a structured process, they support accuracy and reduce barriers.
This approach moves beyond fixing individual files. It creates a foundation for inclusive, flexible, and sustainable document accessibility.
Video highlights
- The problem with PDFs
- What if we translate PDFs?
- Why Markdown?
- Five-step Reflow process
- Real-world examples
- Q and A
Watch the presentation
Resources
Bio
Blake Bertuccelli-Booth is UIC’s Assistant Director of Web Accessibility Engineering. In addition to passionately working to advance the rights of people with disabilities, Blake is the creator of Equalify – an open-source web accessibility platform – and leads accessibility testing for WordPress. He has spoken at numerous conferences, including HighEdWeb, WordCamp, and WPCampus. Blake is passionate about raising the bar for digital inclusion through open tools, community-driven standards, and real-world impact.
Dylan Isaac is an AI accessibility consultant and founder of Enablement Engineering, and a former Lead AI Engineer at Deque Systems, where he invented and built axe Assistant. He’s currently building agentic harnesses to solve accessibility and education problems that were previously impossible — including Equalify Reflow, a multi-agent system that converts PDFs into accessible semantic markup. When he’s not wrangling AI agents, he’s eating wings and shoveling snow in his home in Buffalo 🦬.
FAQ
What makes PDFs difficult to make accessible?
PDFs lack a consistent semantic structure. They were designed for printing, not digital navigation. This makes it hard for assistive technology to interpret them.
Why is Markdown useful for accessibility?
Markdown is a plain text format with simple syntax. It maps directly to HTML, which supports accessible structure by default.
How does AI improve document accessibility?
AI can interpret visual cues and convert them into semantic code. It can identify headings, tables, images, and relationships between elements.
What is semantic translation?
Semantic translation is the process of converting visual layout into meaningful structure. It focuses on the intention behind the design rather than the appearance.
Can this approach handle complex documents?
Yes. The system can process academic papers, posters, tables, and two-column layouts. It reconstructs structure using Markdown and HTML.
